The scenario

The model described by this commentary aims to address this specific scenario:

Suppose artificial general intelligence (AGI) arrives in Australia in the 2030s and does most of the work people are now paid to do. Suppose also that the people it displaces do not find comparable work. Can the country still pay its way? Can it pay everyone an income, keep its budget and its foreign debts under control, and leave people across the income range better off than today's tax and transfer system would, without inequality rising?

The model behind the commentary simulates the Australian economy year by year from 2030 to 2055. It is anchored to 2026 national accounts, tax and superannuation data. Dollar amounts are given in the dollars of the year they are paid, and what a person receives is also shown as purchasing power: what it buys at today's (2026) shop prices.

AGI does not arrive all at once in the model. It follows an S-shaped path, slow, then fast, then slow: what AI can do, what it adds to the economy and what it costs all move along it, and firms take it up along an S-curve of their own (chapter 3).

The result

Although AGI is assumed by this model to cause a very large reduction in employment and hence wages and income tax revenue, the economy and people can survive and thrive by transitioning to raising revenue from a share of AGI-induced productivity gains. By taxing those gains in proportion to the human labour displaced (and not just allowing those gains to accrue only to consumers and capital owners), enough revenue can be raised to guarantee every adult a daily income sufficient to support an expanded economy, higher investment and consumption and greater national wealth.

A representative set of policies, called D18, passes all four of the model's key tests (external/foreign liabilities, fiscal, purchasing power and equality - see Appendix A for details).

D18 is one response to one scenario. The model exists to explore: what AGI could do to an economy like Australia's, and what a government could do about it. Other responses pass the tests too, other scenarios can be set, and every setting in this commentary is a control on the model's page. Readers are encouraged to change them and look for better answers (chapter 4).

• Work. Employment falls from 14.7 million in 2030 to 9.1 million in 2035 and to its lowest point, 6.5 million, in 2042, and recovers only to 7.0 million by 2055, as the population grows. With no AGI, 18.2 million people would be employed by then. This is the result of the core premise of the model: AGI greatly reduces the demand for labour.
• An income for everyone. A Guaranteed Daily Income (GDI) replaces pensions and most benefits. It is phased in as AGI takes over jobs rather than by the calendar, and is fully in from 2036, when it pays an adult $147 a day, which buys what $114 buys at today's shop prices, and a child a quarter of that. It rises faster than output per person, measured in dollars, to $435 a day by 2055 for adults and $108 for children, which buy what $237 and $59 buy today. Reminder: unless specifically noted, dollar figures in this commentary are in the dollars of the year they are paid, and "buys today" means at 2026 shop prices, GST and levy included (chapter 2).
• Workers who lose their jobs. While the GDI is being phased in, it is not yet enough to live on. A worker displaced after 2030 is therefore also paid today's unemployment payment, JobSeeker and Rent Assistance, on the same terms as someone already receiving it: on top of the GDI, and withdrawn on the same schedule as today's other payments as the GDI phases in and eventually fully replaces all other transfers. This costs 4.2% of output at its peak in 2038 and is last paid in 2047. With it, nobody in the lowest income quintile under today's system lives in a household more than a tenth worse off than under today's system in any year, and no more than 9% of the next lowest quintile do.
• People. Households are compared as people, each counted once, by purchasing power per equivalent adult. By 2055 each quintile of people has between 1.65 and 1.87 times the purchasing power of the same quintile under today's system: the poorest quintile 1.87 times and the top quintile 1.84 times. The quintiles are ranked afresh under each system, so this compares the poorest quintile then with the poorest quintile under today's system, not the same people. The purchasing-power test examines four points of the distribution in the same way, and all four pass: the people at the 10th, 20th, 50th and 90th percentiles have 1.73, 2.00, 1.84 and 1.89 times the purchasing power of the people at those points under today's system. A no-AGI Australia that kept today's policies would leave those four points 1.25 to 1.27 times better off.
• People 67 and over. Members of the national investment fund that replaces the current superannuation system who are aged 67 and over draw 5% of their balance a year, everyone else 2.3%. Households 67 and over who rely mostly on the pension start 9% ahead of today's system and those on a part pension 16% ahead; those living on their own savings start 1% behind, are 3% behind in 2031, and are ahead by 2035.
• No loss for the bottom three quintiles. Taken together, the poorest three quintiles are never more than 0.5% below their 2030 purchasing power, in 2031. The transition starts slowly: in 2031 the GDI pays only $4 a day, and the poorest quintile slips 0.5%. The second and middle quintiles slip 1.7% and 1.4% in 2033, the year jobs start to go in large numbers. The poorest quintile then gains 41% by 2040, and stands still until 2048 as today's pensions and benefits are withdrawn, before gaining again.
• Who is worse off. The transition is paid for mostly by people in the upper half of the income distribution. The fourth quintile's purchasing power is 7.7% below its 2030 level in 2034. In 2034 the quintile of people who are best off under today's system is on average 9% behind it, and more than two fifths of them live in households more than a tenth worse off, through a high wage lost to AGI, higher income tax, and prices pushed up by the levy. The fourth quintile is ahead of today's system from 2036 and the top quintile from 2037.
• Inequality. The model's internal measure of the Gini coefficient of purchasing power is 0.177 under today's system. Under D18 it falls to 0.124 in 2036, when the GDI is fully phased in and displaced workers are also receiving the unemployment payment, rises to a peak of 0.175 in 2048, the year the last of today's payments is withdrawn, and ends at 0.167 in 2055. Inequality is lower than the current measure for each year.
• The budget. The model's government debt is gross debt: everything Commonwealth, state and local government owe, before netting off their financial assets (chapter 3). It never rises above its 2030 level of 64.7% of output, and ends at 37%. The budget runs its largest deficit, 4.3% of output, in 2037, while the GDI, today's pensions and the unemployment payment are all being paid. From 2039 it borrows about 3% of output a year to fund a public investment programme, and without that programme it is in surplus from 2040; with it, the budget is in surplus from 2052. Two margins are thin: debt's worst five-year average is 62.5% of output against the model's internal fiscal test limit of 71.2%, and the Gini's peak of 0.175 is against a model limit of 0.177.
• The rest of the world. Net foreign liabilities rise from 22% of output in 2030 to a peak of 34.5% in 2037. Australia then becomes a net creditor from 2048.
• If the premise is wrong. With a quarter, half or all of the displaced re-employed, D18 still passes all four model tests.

What it requires

D18's main policies:

• a GDI phased in as AGI takes over jobs, fully in once the jobs taken over reach 40% of the jobs the economy would have without AI, paid in full for every day an adult does not earn wages (including weekends) and every day for every child, and once phased in, rising faster than economic output per person;
• today's unemployment payment for workers displaced after 2030, paid on top of the GDI and withdrawn on the same schedule as today's other payments;
• today's pensions and benefits phased out over ten years on the GDI's clock, but never faster than the GDI rises, so 40% of them are still paid in 2040, 17% in 2045 and none from 2048;
• employer superannuation and payroll tax are abolished, with the saving passed on in lower prices;
• an AGI levy in place of the GST: a tax on sales of everything households buy except housing, with a rate which rises with the share of an industry's labour that AGI has taken over;
• income tax with a tax-free threshold of at least $40,000 from 2035 and no rate above 50%, Medicare included: 45% from $65,000, 48% from $135,000 and 50% from $190,000; and company tax at 35%. The thresholds are stated in 2026 dollars, and tax brackets increase as a proportion of economic output per person.
• deferred pay: in 2031–35, income above $100,000 a year (in 2026 dollars) is partly saved compulsorily and returned, indexed, from 2038 once the economy has room;
• the superannuation pool, about $6,000 billion by 2030 in 2026 dollars, run as a national fund owned by its members, who draw (tax-free) 5% of their balance a year from age 67 and 2.3% before, so that the fund as a whole pays out about 3% each year;
• a public investment programme that waits until the AI build-out is over and then uses the economy's spare capacity;
• once government debt is below half of output, the budget's surpluses go to raising the GDI.

What it relies on

These are the assumptions of the model and of D18:

1. AGI arrives, and on a large scale. D18's income and taxes are sized for a very large shock, but they scale themselves to it: because the GDI and the withdrawal of today's payments follow the jobs AGI takes over, a smaller shock brings a smaller GDI. With the technology of the published studies, D18 passes all four tests (chapters 7 and 8). What fails badly is AGI that displaces workers but adds little to what the economy can produce (chapter 7). AGI's productivity gain matters too: D18 assumes AGI eventually adds 60% to what the economy can produce; at about 18%, the size the published studies imply, it still passes, but only just.
2. The pace of AI. D18 assumes firms take AI up within about five years. At the pace of any earlier technology, from the smartphone to electricity, it still passes all four tests; with the world gaining only two thirds as much as Australia, the middle paces fail the fiscal test's 2055 limit. Paid on a calendar instead of following the jobs, the GDI would fail at every pace (chapter 7).
3. The rest of the world goes through the same transition. D18 assumes the rest of the world gains three quarters as much from AGI as Australia does, both in its demand for Australian exports and in its companies' profits. It passes at two thirds, and down to 63%. At 60% the fiscal test fails narrowly, and at 50% it fails clearly. At higher world gains, the policy package outperforms.
4. AGI lifts growth. The no-AGI baseline grows at 2% a year, the projection in Treasury's 2026 Intergenerational Report. The AGI world's trend growth is one point higher (3%) from 2031, as tools made by AGI spread into work that AGI does not do itself. Without that the package still passes, with government debt ending at 45% of output rather than 37%.
5. Interest rates that fall once the build-out is over, but not below inflation. The model's interest rate stays above 4% until 2037, falls to its 1% floor in 2040 and 2041 as capacity falls idle, and from 2050 is about 2.5%, level with inflation, so the government's debt costs it about nothing in real terms from then on. D18 does not need lenders to accept less than inflation. With the rate never below 3% or 4%, with the rate held at 5% throughout, or with ordinary inflation of 1.5% or 1%, D18 passes in the base case and fails the fiscal test only if the rest of the world also gains just two thirds as much as Australia. With no ordinary inflation the fiscal test fails.
6. An AGI productivity levy that can exceed 100%. The most any firm pays is 100% in 2030, rising to 115% in 2035 and 120% from 2040, on its price before tax, and the levy is the largest single tax. As a share of the price the buyer pays, those are 50%, 53.5% and 54.5%. Capped at 100% throughout the package still passes, with debt ending at 47% of output rather than 37% and fewer people in work. If a tenth goes uncollected the package still passes, but with a quarter uncollected it fails.
7. A state that does not grow with the economy. The economy grows dramatically, but real government services per person rise by only 5% by 2055. That is a parameter choice, not a policy necessity: services per person could grow by 33%, about as fast as with no AGI, and D18 would still pass. At 46% the fiscal test fails.
8. Tax and transfers on an unprecedented scale for a market economy. Government revenue reaches 65% of output and the GDI 40% by 2055.

Of 118 single parameter changes to D18's assumptions, each to a value a reasonable person could choose, 21 break at least one test: 14 the fiscal test, 9 the external test and 8 the inequality test. Chapter 7 lists them.

What it is not

This is not a forecast, and D18 is not a policy bill ready to legislate. The model shows that one internally consistent set of policies passes four demanding tests under stated assumptions, and which of those assumptions the result relies on. D18 is neither unique nor optimal: a variant that withdraws the GDI more slowly also passes (chapter 7), and others surely do too. It is one of many possible responses to one of many possible scenarios, offered so that the questions can be argued over with numbers.

The model cannot say whether any of this could be legislated, how people would behave under it, or what a society would be like with most adults outside paid work. Those questions may matter more than anything in the model (chapter 17).

Creative Commons Attribution 4.0

This model and commentary are licensed under CC BY 4.0. You are free to share and adapt for any purpose, even commercially, with attribution. The model is instantiated by its web page.

1.1 Why think about this now

Very large amounts continue to be invested in AI research and its commercial use. The heads of two leading laboratories have put dates on artificial general intelligence: Anthropic CEO Dario Amodei by about 2028, and Google DeepMind co-founder Demis Hassabis by about 2030. Both have a stake in the claim, and AGI is not certain. But progress is broad, and AGI could come from any of several research groups. Dismissing the possibility of widespread economic disruption seems as reckless as ignoring the warnings of climate scientists in the 1990s. Amodei has put the prospect bluntly: "Cancer is cured, the economy grows at 10% a year, the budget is balanced - and 20% of people don't have jobs."

Treasury now treats AI as one of the forces that will shape the next forty years. The 2026 Intergenerational Report names it first among five major transitions: "AI is developing rapidly, and will be a defining influence on the economy over the next 40 years" (page vii). Its projections are cautious all the same. They assume labour productivity grows 1.2% a year, as in the 2023 report, and they expect AI to help Australia reach that figure rather than exceed it: "The rise and adoption of AI is likely to support the achievement of Treasury's long-term labour productivity growth assumption over time" (page xvi). Chapter 11 returns to the report.

This commentary starts from the other direction. It asks what happens if AI does far more than that, if the unprecedented amounts being invested in AI are justified. Milton Friedman observed that when a crisis comes, the actions taken depend on the ideas that are lying around.^{1 2} If AGI arrives on anything like the timelines its builders suggest and its financiers assume, then analysed and credible policy options need to exist before the disruption, not after it.

1.2 Why Australia

Australia is a useful case, and not only because it is ours.

• Its tax base relies on labour. Income tax raises about a third of all government revenue, and payroll tax is the states' largest tax of their own. An economy in which wages shrink loses the base that funds most of the state.
• It has a very large retirement pool. Superannuation held about $4,300 billion in June 2025, and will hold about $6,000 billion by 2030, in 2026 dollars: nearly two years of national output.³ How that pool is invested, and who owns what it earns, becomes a central question when super earnt through employment no longer funds retirement.
• It has to finance a current account. Australia runs a current-account deficit of about 2% of output and owes the rest of the world, net, about a fifth of a year's output. An AGI transition that needs imported equipment has to be paid for.
• It will mostly buy AGI rather than build it. Without substantial local investment (consistent with Australia's poor record on R&D funding as a share of GDP), the "frontier" systems will probably be built elsewhere. What Australia can control is how the gains are taxed and shared, and how much of the spending on AI stays here.⁴
• Its public data are good. ABS national accounts and tax statistics, ATO data, APRA's superannuation figures and Jobs and Skills Australia's occupation-level AI exposure scores are enough to anchor a model of this kind.

1.3 How the model came about

The model began in early 2026 with a narrower question: whether Australia could pay a universal income, the Guaranteed Daily Income, by funding it with a variable GST that rose as firms' wage bills fell as a share of their revenue.

Work on it showed that approach could not work on its own. The variable GST is a consumption tax. As its rate rises it raises prices and shrinks the spending it taxes, so it erodes the base it needs. The question was no longer only how to fund a transition. It became whether a post-AGI economy can consume its own output and still pay its way with the rest of the world: whether there are policy settings that carry Australia through without ending in large budget and trade deficits.

Answering that needed a larger model: a national fund built from the superannuation pool; an external account that tracks stocks as well as flows, with the foreign share of business profit derived from who owns what; an interest-rate rule and a price level; 95 household types built from census data; a base year calibrated to the 2024–25 national accounts; and a technology clock on which AGI arrives slowly, then quickly, then slowly. Every dollar the model moves has a payer and a receiver, and automated checks hold it to that (Appendix C). The hard years of the transition turn out to be caused by an AI build-out that has to be financed, from abroad or by restraint at home (chapter 6).

1.4 How to read this commentary

A few conventions run through every chapter.

• Dollars. Amounts that have to add up, such as the budget, debt, output and the fund, are in the dollars of the year they belong to. What a person receives or pays, such as the GDI, pay or a tax threshold, is given in the dollars of the year and also as what it buys at today's (2026) shop prices, GST and levy included. Settings, such as the GDI's $105 a day and the tax thresholds, are stated in 2026 dollars, as they are on the model's page. The model assumes ordinary inflation continues at 2.5% a year, the middle of the Reserve Bank's target band; AGI's price falls and the levy's price rises come on top of it. By 2055 a dollar buys 54.4% of what it buys at the shops today, against 47.9% in a no-AGI Australia. Section 2.5 explains.
• Output. "Output" means the model's measure of national output in the same year, before consumption taxes. Shares of output therefore run a little higher than the familiar shares of GDP at market prices. The test thresholds are adjusted for this (appendix A).
• Government. The budget, revenue, spending and debt are those of all three levels of government together: the Commonwealth, the states and local government.
• Two comparisons. "Today's system" means the same households under the tax and transfer rules of 2026. "A no-AGI Australia" means the model's baseline, which keeps today's policies and grows at the 2026 Intergenerational Report's 2% a year.
• The base case. The package, D18, was set on the model and frozen before this commentary was written. Every figure here comes from one run of that model.
• Notes. Numbered notes are collected at the end of each chapter.
• The model itself. The model is published as an interactive page where every figure here can be reproduced and which lets you run the model with assumptions from other studies (see Part III) or change settings yourself.

2.1 The premise: no comparable work for the displaced

The model assumes that workers displaced by AGI are not re-employed in comparable work. In the model's terms, the reallocation rate is zero. This is the premise of the exercise, not one of its findings. It is also the reason the model exists.

Most published modelling of AI's economic effects assumes the opposite. Lennox and Dixon's study for the Centre of Policy Studies holds total labour supply fixed and lets wages in each occupation adjust until everyone is employed (chapter 8). EY's modelling finds a small net gain in jobs (chapter 9). In these models a worker whose tasks are automated moves to other work, and total employment ends where it would have been anyway. This model can reproduce that. Run with those studies' assumptions, it puts employment in 2055 at about 17 to 18 million, close to the no-AGI path of 18.2 million. The disagreement is about the labour market, not the technology.

There are three reasons to start from zero instead.

• It is the case that needs a plan. If displaced workers find comparable work, today's tax and transfer system mostly copes, and a policy model has little to add. If they do not, the wage base that funds most of the state shrinks within a decade. A country that has not thought about that case in advance will be improvising in a crisis.
• Re-employment is least certain for a general-purpose technology. Past waves of automation took over particular tasks, and people moved to tasks machines could not do. AGI is, by definition, a technology for which that list of tasks is short and getting shorter. Whether new kinds of work appear fast enough is an open question.
• It is a bound. Zero re-employment is one end of the range. Any real outcome lies between it and the published studies' full re-employment, and the model can be run anywhere in between. Section 2.4 does that.

2.2 The layoff trap

Brett Hemenway Falk and Gerry Tsoukalas open their paper The AI Layoff Trap with the sentence that best states this model's concern:

If AI displaces human workers faster than the economy can reabsorb them, it risks eroding the very consumer demand firms depend on.

Their argument is that even knowing this does not stop it happening: "An automating firm captures the full cost saving but, under competitive pricing, bears only a fraction of the resulting aggregate demand destruction; the rest falls on rivals."

Each firm automates more than would be best for firms as a whole, and the loss falls on owners as well as workers. They find that more competition and better AI make the excess worse, and that wage adjustments, free entry, capital income taxes, worker equity, a universal basic income, retraining and bargaining do not remove it. "A Pigouvian automation tax can."

This model is not a model of their prediction. Firms here adopt AI when it is cheaper than the people it replaces, and the model does not track each firm's share of the demand it destroys. But the two point the same way. The model's AGI levy rises with the share of an industry's work that AGI has taken over, so it charges more to industries that have automated more. And Falk and Tsoukalas note that once AI can do all human work, "the remaining failure is distributional, not allocative". They describe a universal basic income that "returns part of the owners' surplus to households as spending". That is the job the GDI does here.

2.3 What happens with no policy response

The model can show what the layoff trap would do to a national economy, and what today's safety net does about it. Take D18's technology assumptions but leave Australia's policies as they are: no GDI, no levy, with superannuation, payroll tax, pensions and benefits all kept, and today's tax scale. A worker who loses a job to AGI is paid today's unemployment payment, as they would be now.

Output rises. By 2040 the economy produces 45% more than it would without AGI. But nearly half of all jobs have gone, and wages have fallen from 43% of output to 26%. Households as a whole hold up. Prices at the shops are 31% lower than they would have been, so today's pensions and benefits buy more, and the unemployment payment reaches the people displaced. Spending does not collapse, and the demand that firms depend on is still there. But the payment is a safety net, not a wage. In 2035 a fifth of people live in households more than 20% worse off than under today's system.

The budget pays for all of it. The unemployment payment costs 4% of output by 2035, 6% by 2040 and 9% by 2045, and there is no new revenue to meet it. The deficit passes 10% of output in 2036. Debt passes 100% of output in 2038 and reaches 156% by 2045, rising by about eight points a year from 2040. It is hard to believe any government could keep borrowing on that path.

Even that outcome is optimistic. The payment is borrowed and spent while the economy is working at capacity, so the extra spending is imported, and from 2035 the country reaches the model's limit on borrowing from abroad (chapter 3). The model responds by slowing the adoption of AI, by up to a third. That is why 7.7 million people are still in work in 2040. Their jobs are not saved: employment falls to 4 million by 2050 as the brake comes off. And the interest rate the government pays does not rise with its debt.

Take the payment away and the layoff trap appears. If the same run pays displaced workers nothing, spending falls short of what the economy can make, and by 2040 14% of its capacity is idle. In 2035 the person at the 10th percentile is 56% worse off than the person at that point under today's system, a third of all people are more than 20% worse off, and by 2040 the Gini coefficient has more than doubled. The budget is spared the payment, and still has debt above 100% of output from 2040.

Either way, today's system does not survive AGI on this scale. With its safety net it fails its budget. Without it, it would fail its people.

The table stops at 2045 on purpose. After the mid-2040s this run is far outside what the model was built for, with debt beyond 150% of output and still rising, and its later figures are arithmetic rather than findings. The point is what happens in the first decade and a half.

2.4 If the premise is wrong

A government would have to choose its settings before it knew how many displaced workers will find work. So D18 should not fail if the easier case turns up. The model can re-employ a share of displaced workers at 65% of their former productivity:

D18 passes these scenarios, and households are better off at every level the more people find work. Debt ends lower with a quarter or all of the displaced re-employed, and higher with half, though still well inside its limit. More wage income also means more spending on imports, so the external position ends weaker: with everyone re-employed, Australia ends owning, net, 15% of a year's output more than it owes, rather than 61%, still well inside the external test. And because the GDI follows the jobs AGI takes over, it phases in more slowly when fewer jobs go.

2.5 Why 2026, and which dollars

The model's first year of AGI is 2030, but its values are anchored to 2026, built from the latest national accounts (2024–25), tax statistics and superannuation data. Readers know what things cost in 2026, and what a 2026 wage or pension feels like.

A commentary could go further, and state every amount in 2026 dollars with ordinary inflation taken out. Correcting a 2045 dollar amount for twenty years of inflation in one's head is notoriously hard, and often done badly, so a single unit might seem the kindest thing to offer a reader. It confuses more than it helps. Inflation cannot be shown in a unit that has removed it, a debt fixed in dollars appears to shrink for no visible reason, and the fund's assets have to be valued on two bases at once. So this commentary uses two measures, and never mixes them in one table.

• Dollars of the year, for anything that has to add up: the budget, government debt, output, the fund, and the accounts with the rest of the world. The deficit of $138 billion in 2035 is $138 billion of 2035's money, and it adds to that year's debt dollar for dollar.
• What it buys today, for anything a person receives or pays: the GDI, pay, a tax threshold. The amount in the dollars of the year is divided by the rise in shop prices since 2026, GST and levy included. The GDI of $102 a day in 2035 buys what $80 buys at the shops today.

Shares of output are the same in any dollar. The package's settings, such as the GDI's $105 a day and the $190,000 top tax bracket threshold, are stated in 2026 dollars, as they are on the model's page, and are indexed from there.

The model assumes the ordinary machinery of prices and wages keeps working as it does now, with inflation of 2.5% a year on average, the middle of the Reserve Bank's 2–3% target band. Two things move prices away from that trend. AGI cuts what it costs to make things, and competition passes the saving into prices: by 2040 prices before tax are 34.7% below where ordinary inflation alone would have put them. Part of that comes at once, in 2030, when the package abolishes employer super and payroll tax and those savings pass into prices too. And the AGI levy, a tax on what households buy, takes most of that saving back at the shops:

Shoppers see something close to ordinary inflation over the period as a whole: shop prices rise 78.6% between 2030 and 2055, against 85.4% from ordinary inflation alone. The levy's rise pushes shop prices up faster than usual while it is phased in, by 16.8% in the three years to 2034. AGI's continuing cost falls then hold them nearly flat from 2036 to 2039. As the technology matures in the 2040s, prices return to the trend. It is prices before tax that fall, from 2033 to 2038, by as much as 3.6% and 4.0% in 2034 and 2035; at the shops the fall is never more than 0.2% in a year.

Wages in the model do not respond to shop prices. They follow productivity and the 2.5% trend, and the average pay of those still working rises 14.1% in the same three years, a little less than shop prices do. Today's pensions and benefits are indexed to wages, and the GDI and the tax thresholds are indexed to output per person in dollars. Nothing is indexed to shop prices, so the levy's lift to the price level sets off no second round of wage and price rises. That is an assumption, not a result (chapter 17).

Debts are fixed in dollars. The model charges interest at the rates people would see, and ordinary inflation erodes every fixed-dollar amount, such as government debt, as it does today. How much that helps a borrower depends on the interest rate. A rate held at 1% for decades while inflation ran at 2.5% would have the government paying about −1.5% a year on its debt in real terms. In the model the rate is never below underlying inflation (chapter 3). It stays between 4.2% and 5.1% until 2037, falls to 1% in 2040 and 2041 while AGI is holding prices down, then follows inflation up and sits level with it from 2050, so the government then pays about nothing in real terms on its debt. The rate is below that year's shop inflation in 2033 and 2034, by 1.1 and 2.2 points, while the levy lifts shop prices and the central bank "looks through" (ie, ignores) it, and again from 2040 to 2045, by 0.2 to 0.6 points, when prices return to their trend faster than the two-year measure the central bank reads.¹

The model is a year-by-year simulation of the Australian economy from 2030 to 2055, 26 years in all. This chapter describes its components and Appendix C describes how it was checked.

3.1 Three layers

The model has three layers, each feeding the next.

The economy and its accounts. Output, incomes, spending, saving and investment for the whole economy, with an itemised government budget and an external account. Output in the base year is set at $3,200 billion in 2026 dollars.

Eighteen industries. Each industry uses a mix of occupations, and each occupation carries Jobs and Skills Australia's scores for how exposed its tasks are to automation and to augmentation by AI. AGI arrives industry by industry, on the technology clock described below. AI doing a displaced worker's tasks is assumed to cost about 8% of that worker's wage in 2030 (an adjustable model parameter), falling to a tenth of that as the technology matures, so it is almost always cheaper, and adoption follows exposure: it runs fastest in the industries whose work is most exposed. The AGI levy can hold adoption back: where the levy would push an automating firm's prices above those of a fully human competitor, the firm stops short (chapter 13). Adoption displaces workers, raises output per worker and lowers the cost of producing goods and services. On top of the gains from adoption, the AGI economy's trend growth is one point faster (by default, 3%) than the no-AGI world's (by default, 2%), as tools made by AGI spread into work AGI does not do itself.

Ninety-five household types. The household layer is built from census data: couples and single people, with and without children, renters, owners with and without a mortgage, pensioners, and so on, 95 types in all. Each type's earners are split into scenarios for who loses their job and who keeps it. Each year the model calculates every type's income, tax, transfers, housing costs and purchasing power, and compares it with the same household under today's system.

3.2 The technology clock

The change imparted by a new technology tends to be slow at first (amongst early adopters), then gathers speed with mass adoption before slowing. The model gives AGI that shape. One S-shaped curve, the clock, describes how far the AI frontier has come: near zero in 2030, moving fastest in a chosen year, and approaching its limit after that. In D18 it moves fastest in 2035 and takes ten years to go from 10% to 90% of the way. Three things follow it.

• What AI can do. The capability scale sets how much of each occupation's measured exposure AI can actually take over. At 0 it is Jobs and Skills Australia's scores as published, under which AI could take over 54% of today's jobs at full adoption. D18 starts there in 2030 and rises with the clock to 100 at maturity, where AI could take over 78%.
• What AI adds to productive capacity. At maturity the economy can produce a set share more than it could without AI: 60% in D18. The gain arrives with the frontier and with firms' take-up of AI, and runs a set time behind take-up, a year in D18, for firms to learn to use what they have adopted.
• What AI costs. The cost of AI falls along the same curve to a set share of its 2030 price, a tenth in D18, running behind capability by a set time, two years in D18. New capability arrives dear and becomes cheap later.

Firms take AI up along an S-curve of their own: in D18, 10% of the way in 2031 and 90% five years later. Work that needs robots and machinery, rather than software alone, runs behind by the embodiment lag. The jobs taken over in any year are what AI could do at that year's capability, in the industries that have taken it up, less what the levy holds back.

The clock's settings are the scenario, not the policy. They are guesses about a technology that does not yet exist. Chapter 4 explains D18's choices, and chapter 7 runs D18 at the pace of earlier technologies, from the smartphone to electricity, with the size of AGI's effect held fixed: history can say something about how fast a technology spreads, but nothing about how large AGI's effect will be.

3.3 Spending, capacity and imports

The model's most important structural choice is that output is limited by spending as well as by capacity.

The economy's capacity grows with the workforce, with ordinary productivity growth and, in the AGI world, with AGI. How much of that capacity is used depends on what households, businesses, government and foreign buyers want to buy. If spending falls short, output falls short and capacity sits idle. If spending exceeds capacity, the excess is met from abroad: imports rise and the current account goes further into deficit. That is how the model asks its central question, whether an AGI economy can consume its own output and pay its way.

Business investment responds to spare capacity and interest rates. Building the AI plant itself is investment, much of it imported. A public investment programme, when it runs, fills capacity that would otherwise sit idle (chapter 4).

The model also enforces a financing limit on borrowing from abroad. Each year the current-account deficit may be no larger than one sixth of the room left below a ceiling on net foreign liabilities of 80% of output, plus the deficit that growth alone would absorb. In 2036, D18's tightest year, that allows a deficit of 8.4% of output; D18's deficit that year is 6.4%. If the deficit would be larger, AI adoption is slowed until it fits, though that AGI brake can only be changed gradually from year to year. A year in which even that is not enough is a breach of the limit.¹

That limit is a policy the government chooses, not one the world imposes. Lenders would probably allow more, just as a bank raises the credit limit of a borrower who cannot really afford it but has assets the bank could sell. And slowing AI adoption, the limit's instrument, often leaves the external position worse rather than better, because an economy held back must then import what it cannot make (chapter 7). So breaches are reported beside the external test, not scored in it; the test itself is about net foreign liabilities. In D18 the limit is never breached, and the deficit never comes closer than 2.1 points of output to what it allows.

3.4 The budget, the fund and the rest of the world

The budget is the combined budget of all three levels of government, Commonwealth, state and local, and it is itemised: income tax, company tax, the GST and the levy, state taxes, the GDI and the transfers it replaces, government services and investment, and interest. Company tax is charged on profit after interest, including the interest companies pay abroad, as it is today.

Government debt accumulates from deficits. It is gross debt: everything the three levels of government owe, before netting off the financial assets they hold, such as the Future Fund, state investment funds, loans and deposits. The model starts it at $2,050 billion in 2030, in 2026 dollars, from the Parliamentary Budget Office's projection of national gross debt.² Interest is charged on it at a rate calibrated to the interest governments actually paid in 2024–25, and the income from their financial assets is counted in revenue. Net debt, which subtracts those assets, is the figure more often quoted in Australia, and it is more than 20 points of GDP lower: the Parliamentary Budget Office projects national gross debt at 58.0% of GDP in 2029–30, and net debt at 36.5%. Chapter 11 compares the model's debt with the Intergenerational Report's.

The national fund holds the superannuation pool, about $6,000 billion in 2030, in 2026 dollars. In D18 it holds 5% in government bonds, 55% in Australian shares and 40% overseas. Members aged 67 and over draw 5% of their balance a year and everyone else 2.3% (both groups' drawings are tax-free), so the fund pays out about 3% a year in all, and it reinvests its earnings. Its overseas assets are assumed to earn 6% a year in real terms, 3% as income and 3% as growth.

The rest of the world. The external account tracks the current account and the stocks behind it: external debt, foreign-owned Australian shares, and the fund's and other Australians' overseas assets, with the income paid and received on each. Australian business is owned by three groups: the fund, other Australian residents and foreigners. The foreign share of profit is worked out each year from who has bought and sold shares. It starts at 29% in 2030.

3.5 Interest rates and prices

The model has one interest rate. It is closer to what the government pays to borrow for ten years than to the Reserve Bank's cash rate. Mortgages cost 1.5 points more and businesses borrow at 2 points more, and the rate drives business investment, the exchange rate and households' cash flow. Government and external debt reprice to it over about seven years.

A rule standing in for the Reserve Bank sets the rate. It aims at full use of the economy's capacity. It cuts when capacity sits idle, by half a point for each point of capacity not in use, and raises the rate when spending runs ahead of what the economy can produce, moving 40% of the way each year. With the economy at capacity the rate is 5%, which is 2.5% in real terms plus 2.5% inflation, a normal ten-year bond yield. The rest of the world is assumed to go through the same transition, so world rates mirror Australia's.

The rule does not consider prices, in either direction. When AGI lowers what things cost to make, the central bank does not cut: a fall in costs is not weak demand. When the levy raises prices at the shops, it does not raise the rate either, as the Reserve Bank "looked through" the introduction of the GST in 2000.³

Two floors hold the rate up. It cannot fall below 1%, about where ten-year yields bottomed in 2020. And it cannot fall below underlying inflation, so that lenders are never asked to accept less than inflation for years on end. Underlying inflation is measured as ordinary inflation of 2.5% plus the yearly change in prices before tax, against that trend, over the two years to the year before; it leaves the levy out. While AGI is pushing prices before tax down, underlying inflation is low or negative and the 1% floor is the one that matters. Once prices return to their trend, in the 2040s, the floor rises to about 2.5%.

In D18 the rate is 5.0% in 2030 and 4.6% in 2033, and stays above 4% until 2037. It falls as capacity comes free, to 1.0% in 2040 and 2041, and from 2050 it is about 2.5%, level with inflation. From 2040 the rate is at one floor or the other: with about a sixth of capacity idle, the rule on its own would put it lower.

AGI lowers what it costs to produce goods and services, and competition passes that into prices. By 2040 prices before tax are 34.7% below where ordinary inflation alone would have put them, and 29.9% below where they stood on that measure in 2030. Prices at the shops include the levy, so they fall by much less. In 2030, before AGI arrives, shop prices are already 6.8% lower than in a no-AGI Australia. That is entirely policy: abolishing employer super and payroll tax lowers business costs, and competitive firms pass the saving into prices. AGI's cut in prices before tax is far larger than policy's, but the levy takes most of it back as tax: by 2055 shop prices are only 12.1% lower than in a no-AGI Australia.

Housing is treated separately because AGI does not make land cheaper. Rents drift up slowly, by 1% a year in real terms. House prices follow rents and the interest rate. When mortgage rates fall, the same repayment supports a larger loan, and buyers bid prices up until a mortgage on a given house costs about what it did; the model spreads that over about ten years. A mortgage grows with house prices relative to incomes, and mortgage interest follows the interest rate. A lower rate also lets households with a mortgage spend more, but only while they keep the lower repayment: as the cut is bid into house prices, the extra spending fades, even if the rate stays low.

The exchange rate follows the gap between Australia's interest rate and the world's. In D18 the world's rates follow Australia's and there is no gap, so the dollar does not move. A risk premium on government debt, if one were charged, would raise what Australians pay to borrow but would not raise the dollar: a premium for risk does not attract capital.

3.6 Comparisons with today and with a no-AGI future

Today's system. Each household in each year is compared with the same household under the tax and transfer rules of 2026. This is the comparison behind the purchasing-power test: is a household better or worse off than it would be under the system Australians know?

A no-AGI Australia. The model's baseline keeps today's policies and has no AGI. It grows at 2% a year, the Intergenerational Report's projection. It starts with a deficit of 3% of output, as Australia's budgets do, and no government acts to close it: there is no consolidation, meaning no spending cuts or tax rises to bring the budget back towards balance. So the deficit widens as interest on the growing debt compounds. By 2055 it is 4.9% of output, and gross debt is 87%. Chapter 11 compares it with Treasury's own projections.

3.7 How households are compared

What a household can buy is its income after tax and transfers, with the GDI, the fund's payout and any other income, less housing costs, at the prices it pays at the shops, including the GST and the levy on everything it buys, imported or made here.

Households differ in size, so the model compares them per equivalent adult: the first adult counts as 1, each other adult as 0.5 and each child as 0.3. The household's purchasing power is shared among its members on that basis, and the comparisons are between people. A fifth is a fifth of people, ranked by purchasing power per equivalent adult, and the Gini coefficient counts each person once. People are ranked afresh each year, so the poorest fifth in 2040 is whoever is poorest in 2040, not the people who were poorest in 2030.

3.8 Keeping up with the economy

The GDI and the income tax thresholds are set in 2026 dollars and rise each year with output per person: the GDI by 125% of its growth, the income tax bracket thresholds by 55%. The measure is output per person in dollars, at the prices it sells for, the same measure the tax base follows. When AGI lowers prices, output per person rises faster in volume than in dollars, and the GDI follows the dollars. A GDI indexed to the volume of output would gain AGI's productivity twice: once through the index, and again as the prices of what it buys fall. Australia already indexes this way: the Age Pension is benchmarked to average weekly earnings, a measure in dollars.

At 100% the GDI would keep its share of output. D18 indexes it at 125%, so its share grows: from 30% of output in 2036, its first full year, to 35% in 2045 and, with the surplus waterfall's help from 2053, 40% in 2055. That is a choice, and chapter 4 gives the reason for it.

If the GDI and the thresholds were indexed to volume, D18's GDI would reach $547 a day by 2055, which buys at the shops what $297 buys today, and cost 50% of output, and the package would fail the fiscal test, with debt ending at 59% of output. Indexed to output in dollars, it reaches $435, which buys what $237 buys today, and costs 40% of output.

3.9 The four tests, and two reported measures

A package is viable when it passes all of these four tests:

1. External. Net foreign liabilities never exceed 75% of GDP, are under 30% by 2055, and would not cross 30% within another 25 years on their final five-year trend. Breaches of the government's own limit on foreign borrowing are reported beside this test, not scored in it (above).
2. Fiscal. Gross government debt averages no more than 65% of GDP over its worst five years, and is under 50% by 2055.
3. Purchasing power. In 2055 the people at four points in the distribution, the 10th, 20th, 50th and 90th percentiles, each have at least the purchasing power of the person at the same point under today's system. People are ranked afresh under each system, so the test compares positions in the distribution, not the same people: a household may sit at a different point under the package than it does under today's system.
4. Inequality. From 2032 the Gini coefficient of purchasing power is never above today's system's.

The limits are set on GDP at market prices and restated on the model's measure of output, which leaves out consumption taxes (appendix A). They are judgements, not measurements.

Two further measures are reported beside the tests but not scored.

• No loss for the bottom three fifths. Whether the poorest three fifths of people, taken together, ever fall below their 2030 purchasing power, and each fifth's worst one-year fall.
• The foreign share of Australian business profit. A package can pass every test by selling Australian businesses to foreigners, so the model reports who ends up owning the profits made by Australian business.

3.10 What kind of model this is

It is not a computable general equilibrium (CGE) model of the kind Lennox and Dixon use (chapter 8). A CGE model finds the prices at which markets clear, usually with a fixed supply of labour, and compares one long-run equilibrium with another. This model is a year-by-year simulation in which output is limited by spending as well as by capacity, and it keeps a full set of accounts: every flow of money and every stock of assets and debts, sector by sector. Models of this kind are called stock-flow consistent.⁴ Prices follow costs rather than clearing markets, employment follows technology and population, and output is the lower of what is demanded and what can be produced.

Each year is solved by repetition until three things agree: spending equals what is produced, with anything beyond capacity imported; the income side of the national accounts equals the production side; and the brake on AI adoption that keeps the current account within its limit has settled. Every year of D18 solves this way.

3.11 How the model is checked

The model is stock-flow consistent. Every dollar one sector receives is paid by another, each sector's income less its spending equals what it lends or borrows, and every stock (debt, shares, the fund, foreign assets) changes only by the flows into and out of it. About twenty-five automated checks enforce this and the model's other identities after every change, and many are themselves tested by injecting the fault they claim to catch and confirming that they catch it. The accounting checks run at the model's presets and at a dozen other settings as well as at the base case.

The base year is checked against the national accounts. The ABS measures the government's deficit in two ways, in the national accounts and in its government finance statistics, and the model sits between them:

The model starts where Australia is: at full capacity, with about 14.5 million people employed.

D18 is one set of policies that passes all four tests, for one path the technology might take. It was chosen under limits set in advance: no personal tax rate above 50%, company tax no higher than 35%, a Gini coefficient aimed at about 0.17 rather than driven as low as possible, and a package that still passes if the rest of the world takes up AGI more slowly than assumed. This chapter sets out the technology D18 is built for, then each lever and the reason for it. Appendix B lists every setting.

Settings are stated in 2026 dollars, as they are on the model's page. Amounts paid in a later year are given in the dollars of that year, with what they buy at today's shop prices beside them (chapter 2).

4.1 One response to one scenario

The model exists to explore. Its purpose is to show what AGI could do to an economy like Australia's, and what a government could do about it, so that the questions can be argued over with numbers rather than guessed at. D18 is not a forecast, and it is not a recommendation that this is the policy Australia should adopt. It is one response to one scenario, and the evidence that responses of this kind exist.

Other responses pass the tests too. The closest is a 90-cent variant, which withdraws the GDI at 90 cents for each dollar of other income rather than a dollar. It passes all four tests, including when the world shares only 66% of the AGI economy, at a cost of slightly more debt: a worst five-year average of 63.8% of output rather than 62.5%, and 40% rather than 37% in 2055. Chapter 7 and chapter 12 come back to it.

Every setting in this chapter is a control on the model's page, and every scenario in chapter 7 can be run there. Readers are encouraged to change the technology, the world and the policy, and to look for packages that do better. The page shows at once which tests pass and why.

4.2 The technology D18 is built for

AGI does not arrive all at once, and it does not stop arriving. In D18 the AI frontier follows one S-shaped path: slow, then fast, then slow. It moves fastest in 2035 and takes ten years to go from 10% to 90% of the way. Three things follow that path:

• What AI can do. In 2030 AI could take over 54% of today's jobs at full adoption, on the occupation scores Jobs and Skills Australia published. By maturity it could take over 78%.
• What AI adds to what the economy can produce. At maturity, 60% more than the same economy without it, a year behind firms' take-up of AI.
• What AI costs. It falls to a tenth of its 2030 price, two years behind the capability.

Firms take AI up along an S-curve of their own: 10% of the way in 2031 and 90% five years later. That is faster than any earlier technology. Television took 12 years from 10% to 90% of American households and the smartphone about eight (chapter 7). But AGI is a tool that can reduce the friction of its own take-up, its cost savings make take-up an economic necessity, and global trade and competition both spur adoption and supply the means to adopt. Chapter 7 runs D18 at the pace of each earlier technology, from the smartphone to electricity. The table shows D18's path; chapter 3 explains how the model uses each line.

4.3 An income for everyone

The Guaranteed Daily Income (GDI) is paid in full to every adult for every day they do not work (including weekends), with a smaller amount for each child. Its full rate is $105 a day for an adult and $26 for a child, in 2026 dollars.

It is phased in as jobs are lost, not by the calendar. Each year the GDI pays a share of its full rate, and the share follows the jobs AGI took over the year before. It reaches its full rate once those jobs are 40% of the jobs the economy would have without AI. The comparison is with that larger workforce, which grows with the population, so that population growth is not mistaken for a recovery. In D18 the threshold is passed in 2035, and the GDI is fully in from 2036:

At D18's GDI withdrawal rate of 100%, a person who receives before-tax wage income on a day that is less than the GDI will receive a "top up". For example, imagine a part-time worker receives wages of $100 on a day in 2040: they would receive a GDI of $88 on that day, taking them to a total before-tax income of $188 for the day. If the next day they received wages of $200, they would receive no GDI.

That is, on a day a person works, the GDI for that day is withdrawn at one dollar for each dollar of the day's wage, down to nothing. Someone working full time at an ordinary wage receives no GDI on their working days, and the full GDI on every other day. At a dollar for a dollar, a day's work adds nothing to a person's income until the day's wage passes the day's GDI. The 90-cent variant would let a person keep ten cents of each dollar earned; chapter 12 discusses the case for it, and chapter 13 the withdrawal on a working day. The GDI is taxable income, like wages.

The full GDI rate rises with output per person, measured in dollars, at 125% of that growth, and does so from 2030, while it is still being phased in. A child's GDI is a quarter of an adult's: $25 a day in 2035 and $108 in 2055, which buy what $20 and $59 buy today. Chapter 3 explains why the GDI follows output per person in dollars rather than in volume.

Why follow the jobs, not the calendar. The GDI's job in the transition is to replace the wages that are lost. Paid on a calendar, it commits the budget to a date, whatever AGI has done by then. If AGI arrives later than the calendar assumed, the budget pays out years before AGI produces the output and the levy that pay for it. Tied to the jobs taken over, the GDI arrives when the wages go, and so does the levy, which is charged in proportion to how much of each industry's work AGI has taken over. The difference is large. Phased in over 2030 to 2035 on the calendar, with everything else as D18, the package fails both the external and the fiscal tests: the worst five-year average debt is 87% of output, against a limit of 71.2. Following the jobs, D18 passes at every historical pace of take-up chapter 7 tries, from the smartphone's to electricity's. The phase-in never goes backwards: if the share of jobs taken over falls, the GDI keeps the share of its full rate it had reached.

Why 40%, and why front-loaded. The threshold and the shape of the phase-in trade the budget against households in the build-out years. A GDI that is fully in only once half the jobs have gone, and phased in later, eases the budget: the worst five-year debt is 60.7% of output rather than 62.5%. But it pays less while the jobs are going, and households' total purchasing power falls 3% below what today's system would give them in 2034; in D18 it never falls below. A lower threshold, 30%, costs more and fails the fiscal test if the world shares only 66% of the AGI economy. D18's phase-in is also front-loaded, paying more early in the ramp than a straight line would (a phase-in shape of 90, where 100 is a straight line).

Why 125%. At 100%, the GDI would keep its share of output. D18 lets it grow faster because, once wages stop carrying most of households' income, the incomes that remain are concentrated higher up the distribution. Wages are earned by fewer households. The fund's payouts and other income from investment go to households in proportion to what they own. Housing does not get cheaper the way goods and services do, and weighs most on poorer households. A GDI that only kept its share of output could not offset that. At 125% its share grows from 30% of output in 2036, its first full year, to 35% in 2045 and, with the surplus waterfall's help (below), to 40% by 2055.

The GDI takes over the jobs that superannuation, the Age Pension, JobSeeker, the Disability Support Pension and other working-age payments do today: an income for people who are not working, whatever the reason. Workers' compensation stays as it is. By 2036 the GDI costs 30% of output, and by 2055 40%. It is the largest component of the budget.

The GDI is a floor, not a ceiling. It pays every adult the same amount, so most of the package's long-run reduction in inequality comes from it (chapter 5).

4.4 Workers who lose their jobs after 2030

While the GDI is being phased in it is not enough to live on. In 2033 it pays an adult $9,604 for a full year, which buys what $8,379 buys today; a single person renting who lost their job under today's system would receive JobSeeker and Rent Assistance of $27,270 a year in 2026 dollars. Most of the jobs go in exactly those years: 2.3 million workers have been displaced by 2033 and 6.3 million by 2035.

So D18 pays a worker displaced after 2030 today's unemployment payment as well as the GDI, on the same terms as a person already receiving it: the payment in full, less the share of today's payments that the phase-out has withdrawn by that year (next section). A person who lost their job in 2029 and one who loses it in 2032 are treated alike.

Today's payment here is JobSeeker at its September 2026 rates and, for renters, Rent Assistance, both indexed to wages as the model indexes today's other payments.¹ In a couple, a partner with no earnings is paid too, as today. The model leaves out Parenting Payment and the rise in Family Tax Benefit when a family's income falls; both would pay more.

One person shows how it works: a single renter with no other income.

With the GDI alone this person would have had, in 2033, less than a third of what JobSeeker and Rent Assistance buy today. With the payment they are ahead of today's system from the first year, and the payment fades as the GDI grows, until from 2048 the GDI is all there is. Between 2040 and 2047 what they receive buys about the same each year: the payment is being withdrawn about as fast as the GDI grows.

The payment costs the budget $8 billion in 2031, $48 billion in 2033 and $135 billion in 2035, in the dollars of those years, and between $135 billion and $184 billion a year from 2035 to 2041. At its peak, in 2038, that is 4.2% of output. It then falls as today's payments are withdrawn, to 3.4% of output in 2040 and 1.5% in 2045, and is last paid in 2047. A displaced worker receives $20,762 on average in 2033, in that year's dollars. That is less than the single renter's payment above, because part of the payment has already been withdrawn, and because many displaced workers do not rent or have a partner who is still earning.

Why the same terms, and not a top-up. The cheaper design is a floor: pay a displaced worker the higher of the GDI and today's payment, not both. The model can run all three:

The GDI alone fails the inequality test: in 2033, while the GDI is still small, the Gini reaches 0.179. The top-up does what its name says, and it costs little. It helps the poorest fifth most. It does much less for the second and middle fifths, because the households left behind there have lost a wage well above any benefit, and the top-up all but stops once the GDI passes JobSeeker, in 2035. And it does not treat like alike: a person on JobSeeker since 2029 keeps most of it and receives the GDI as well, so they receive more than a worker displaced in 2031 and topped up. Paying on the same terms costs about 2 points of output on debt's worst five-year average, and it is the only one of the three in which a displaced worker is no worse treated than someone already unemployed.

What it does not fix. No payment at JobSeeker's level makes whole a household that has lost a good wage. In 2033, 18% of the people in the middle fifth still live in households more than a tenth worse off than under today's system, and 20% of the fourth fifth and 24% of the top fifth do. They are households such as a couple with children, a mortgage and two wages totalling $155,000, one of whose earners has been displaced. Today's system would not make them whole either. Chapter 5 follows them.

4.5 Today's pensions and benefits

Pensions and benefits are not switched off when the GDI starts. They are phased out over ten years on a schedule that keeps more of them in place early and withdraws more at the end. The schedule runs on the same clock as the GDI: it follows the jobs taken over until the GDI is fully in, and the calendar after that. On that schedule alone, they would all be gone by 2042. The unemployment payment to displaced workers follows the same schedule.

A second rule overrides the schedule: today's transfers may never be withdrawn faster than the GDI rises. Each year the share withdrawn can grow only by as much as the rise in the GDI can replace for the average recipient, in dollars. When the GDI grows slowly, as it does in the late 2030s once its phase-in is over, the phase-out slows too:

Under the rule, pensions and benefits are not gone by 2042 but wound down over the next six years: 40% of them are still paid in 2040, 17% in 2045, and none from 2048. The reason for the rule is political as much as economic. A fall in the poorest households' purchasing power as pensions end would be hard for any government to defend, even if those households stayed ahead of where today's system would leave them. The rule protects the average recipient, not every one (chapter 5).

Why this schedule. How fast today's payments go is where two of the tests pull against each other. The fiscal test wants them withdrawn sooner, the more so because the unemployment payment is withdrawn with them. The inequality test wants them withdrawn later, because these payments go to the poorest households, and the Gini coefficient peaks around the year the last of them goes, before the budget's surplus has begun to raise the GDI. A schedule that withdrew them in a straight line fails the inequality test: the Gini's peak is 0.180, in 2044, against a limit of 0.177. One that withdrew them later fails the fiscal test if the world shares only 66% of the AGI economy. D18's schedule lies between: it passes both tests in both worlds, with a Gini peak of 0.175 in 2048.

The phase-out starts in 2030 with a tunable setting of 3% of today's transfers withdrawn.

4.6 Lower costs for employers

Employer superannuation contributions and payroll tax are both abolished from 2030. Each is a cost to business, and in a competitive market the saving is passed on in lower prices rather than kept as profit or paid out as higher wages. Workers are unlikely to bargain the abolished super into their pay once the GDI takes over super's role, and firms cannot keep the saving as profit when their competitors cut prices.

The effect is immediate. In 2030 prices at the shops are 6.8% lower than they would otherwise have been. That matters because the GDI is paid in dollars: every price fall raises what it buys.

4.7 The AGI levy in place of the GST

The GST stays at 10% in 2030 and is phased out by 2035. While it lasts it is broadened to everything households buy, including fresh food, health, education and rent, which are exempt today. In its place comes the AGI levy, a tax on sales whose rate follows automation. Each industry's rate is a maximum rate multiplied by the share of the industry's human work that AGI has taken over, and it is charged on the whole price before tax. An industry that still employs most of its people pays little. One that has automated most of its work pays close to the maximum. The levy reaches everything households buy except housing: rent, mortgage interest and owners' own homes are never levied.

An example shows how it works, at today's prices. Before AGI, preparing a tax return costs $600: $500 of an accountant's time and $100 of everything else. AGI takes over 90% of the work, $450 of the accountant's time, and the AI that does it costs $36, 8% of the pay it replaces. Competition passes most of the saving into the price, which falls to $227, and the firm keeps $41 of extra profit. The industry's levy rate is the maximum, 120%, times the 90% of the work AGI does: 108% of $227, or $246. The customer pays $473, against $660 today with GST. The accountant's $450 is gone as pay; the GDI replaces it as income, and most of the saving reaches the government as levy, which pays for the GDI.

The maximum rate rises steadily, from 100% of pre-tax prices in 2030 to 115% in 2035 and 120% from 2040. Stated as a share of the price the buyer pays, as tax-inclusive rates usually are, those maximums are 50%, 53.5% and 54.5%. Across everything households buy, consumption taxes rise from 7.5% of what they spend before tax in 2030 to 43% in 2035 and about 53% from 2040. That is why prices at the shops rise by 16.8% in the three years to 2034, although AGI is lowering what almost everything costs to make (chapter 6).

Together with a levy on imports, it raises about 21% to 23% of output from 2036 onwards; the levy on its own raises between 17% and 19% from 2040, which makes it the largest single tax. Three things recommend it over heavier income or company tax. It taxes what AGI produces wherever it is produced, including imports. It is less visible, and likely to be less resented, than income tax at the same weight. And unlike company tax, it gives firms no reason to move their profits offshore. In the model it is also the only thing that makes the cost of labour matter to a firm's decision to automate, because a firm that automates past the point where the levy makes its prices higher than a fully human competitor's loses customers (chapter 13).

A maximum above 100% is unusual: the tax is more than the firm's own price before tax. D18 does not need it to pass the tests, but it uses what it buys. Capped at 110%, D18 still passes all four tests, with government debt ending at 41% of output rather than 37% and 6.6 million people in work in 2055 rather than 7.0 million. Capped at 100% throughout, it still passes, with a worst five-year average debt of 68.6% of output against the limit of 71.2, debt ending at 47%, and 6.2 million people in work. Collection is a larger risk. Nothing in the model lets firms avoid the levy, and there would be a strong incentive to avoid a tax this large: if a tenth of it goes uncollected, the package still passes, but a quarter fails the fiscal test (chapter 13).

4.8 Income tax and company tax

Income tax follows a self-imposed limit: no rate above 50%, and the top rate not reached by pulling the thresholds down so that many ordinary earners pay it. D18 charges 50% above $190,000 and 48% from $135,000, the same thresholds as today's top two rates. Below that it is heavier than today's scale: 45% from $65,000, and 33% from the tax-free threshold, rising to 38% by 2055; the threshold rises from $25,000 in 2030 to $40,000 in 2035 and $50,000 by 2055. These are the whole rates. There is no separate Medicare levy, so 50% is the most anyone pays, against 47% at the top of today's scale with the Medicare levy.²

The thresholds are stated in 2026 dollars and rise each year with 55% of the growth in output per person, measured in dollars. So the thresholds in force are higher:

Thresholds that rise more slowly than incomes mean that, over time, more of each income falls into the higher brackets: the top rate starts at about twice the average worker's pay in 2030, and at about 1.5 times it by 2055.

Heavier top rates also pass the tests. Rates of 54% from $121,000 and 49% from $100,000, on D18's other settings, end with debt at 30% of output rather than 37%, and a Gini peak of 0.167 rather than 0.175. But by 2040 wages are only a quarter of output, so rates on the highest wages do more to redistribute than to fund the budget, and the package passes without them. Driving inequality much lower would blunt the incentive to take risks.

Company tax is 35% from 2030, against 30% today. The choice is finely balanced. Under dividend imputation Australian shareholders pay their own personal rate on company profits, so the company rate matters to them less than it appears. But a lower rate reduces the incentive to shift profits offshore, and a rate of 40% would invite companies to move, a response the model cannot capture. All things being equal, 35% seems likely to serve the economy better in the long run. Company tax raises about 15% of output from 2040, largely because profits are a much larger share of the AGI economy. At a 30% company rate throughout, D18 only just passes the fiscal test, with a worst five-year average debt of 71.0% of output against the limit of 71.2. D18 keeps dividend imputation and, as today, refunds unused franking credits in cash. Interest a company pays abroad is deducted from its profit before company tax, as it is today.

4.9 Deferred pay

In the five build-out years, 2031 to 2035, each adult's income above $100,000, in 2026 dollars, is partly saved compulsorily, raising (or "reserving") 2% of output a year. This restrains spending by less than a tax of the same size on everyone would, because the households who pay it spend a smaller share of their income than the average household does (chapter 3). Balances are indexed to prices, so they keep their value but earn no real interest.

The money comes back over ten years, but only once the economy has room for it: repayment begins in the year after one in which less than 97% of capacity is used. In D18 repayment runs from 2038 to 2047.

This is Keynes's proposal from How to Pay for the War (1940): when the economy is fully stretched, have those who can bear it defer part of their spending, and pay it back when the capacity is free. Chapter 6 explains why the build-out needs it. ³

4.10 The national fund

The superannuation pool becomes a national fund. APRA put the pool at about $4,300 billion in June 2025; by 2030 it will hold about $6,000 billion in 2026 dollars. Every member's balance carries over at its value. What changes is the assets behind it, which become 5% government bonds, 55% Australian shares and 40% overseas assets. For most members that changes little: a balance is still a claim on a large, diversified portfolio. Self-managed funds, which hold about a quarter of super today and often a family farm or business premises, would need a transition of their own, which the model does not attempt (chapter 14).

The fund keeps its earnings and reinvests them, as super funds do today, and it buys back Australian shares held by foreigners. Its income is taxed each year at members' own marginal rates, not at super's concessional 15%: about 41% to 44% of it from 2040. Its growth, the retained earnings of the companies it owns and the growth of its overseas assets, is not taxed, and nor is what members draw.

Members draw a fixed share of their balance each year: 5% at 67 and over, 2.3% below. The draw is compulsory and untaxed, and the pool pays out about 3% a year in all. The fund is not a national endowment. It is members' money, built from their superannuation, and a low mandated payout would rightly be called expropriation. Today's rules allow nothing to be drawn before retirement and then require minimum drawdowns rising with age, from 5% of the balance at 65 to 74 to 14% from 95. The two rates keep something of that shape. People 67 and over draw what today's rules ask of them in their late sixties and early seventies, which keeps them from going backwards in the early years of the transition (chapter 5). Younger members get an income from their balance that they would not get today.

When a member dies, the balance passes to their beneficiaries as a balance in the fund, so the fund is handed on between generations. That is a departure from super today, whose death benefits are paid out in cash, and the package depends on it, though it could bear some cash death benefits too (chapter 14).

In the dollars of each year the fund is worth $6,267 billion at the end of 2030 and $25,167 billion at the end of 2055: 190% of output and then 216%. Its value barely grows in the build-out years, $6,607 billion in 2033 and $6,508 billion in 2035, although it is earning more each year. The model values the fund's shares at the price of the capital behind them, and AGI is making that capital cheaper (chapter 5).

4.11 The public investment programme

The programme is the package's use for idle capacity. It is countercyclical: nothing is spent while the AI build-out fills the economy, and afterwards it expands by up to $150 billion a year toward a cap of $1,000 billion a year, aiming to keep the economy near 95% of capacity. It pauses whenever the budget deficit would exceed 3% of output. It first spends in 2039, reaches $229 billion in 2040 and its cap in 2047, all at 2026 prices for capital goods.

It is a physical programme: the budget pays what the works cost at the time, and AGI makes them cheaper. The programme costs the budget 4.3% of output in 2040 and 11.0% in 2045, and at its cap 12.7% in 2050 and 10.9% in 2055.

Eighty per cent of it is productive investment, and 60% of that sells its output abroad: tourism, education, processing, energy. That assumes the rest of the world grows richer with AGI too, and that a larger global middle class will want what Australia can offer: space and coastline, a stable and safe society, education in English, and minerals and energy processed with abundant sun and wind. The world's demand for Australian exports, set below, governs how fast that demand grows. The model also limits it: the price the programme's exports fetch halves once they reach 12% of output, because no market takes unlimited quantities of anything. The rest of the programme builds what is productive but not traded, such as transport, which should aim first at replacing imports. Its import content is 20%. Its returns fall as it grows: the return on the marginal project halves once the productive public capital stock reaches a year's output. Chapter 16 asks how much more of the idle capacity it could use.

4.12 The surplus waterfall

When the budget runs a surplus and government debt is below 50% of output, the surplus is spent in a fixed order: first on lifting the GDI, by up to 5% a year, and the rest on repaying debt. The GDI lift is permanent, so it cannot be taken back in a bad year; and because it spends only a surplus that has actually appeared, it pays out when things go well and not when they go badly. In D18 the budget is in surplus from 2052 and the waterfall first lifts the GDI in 2053; by 2055 it has raised it by 6.6%.

4.13 How big the state becomes

In D18, government services per person grow only 5% in real terms between 2030 and 2055, while output per person more than doubles. Services fall from 23% of output to 9%. In a no-AGI Australia services per person grow 31%, keeping pace with its much slower growth.

This is a setting, not a result. The model lets the state grow with output per person, not at all, or anywhere in between, and D18 sets it near the bottom of the range. The package cannot pass the tests if the state keeps its share of output: a larger state spends the AGI dividend, and imports with it, and the external and fiscal tests both fail. But it has some room. Services per person could grow by 33% by 2055, about as fast as they do with no AGI, and D18 would still pass all four tests; at 46%, the fiscal test fails. Chapter 15 discusses what that means.

4.14 What the package assumes about the world

Three settings describe the world rather than policy.

• The world shares the AGI economy. D18 assumes the rest of the world gains three quarters as much from AGI as Australia does, in two ways set separately: its demand for Australian exports grows with three quarters of Australia's AGI output gain, and profit rates on its companies, including those the fund owns, rise by three quarters as much as Australian companies' do. Incentives to adopt AGI are much the same everywhere, especially among Australia's trading partners and developing economies, so 75% seems reasonable. The two need not move together: a world can share AGI's profits and still not buy Australian goods. D18 still passes if the world shares only 63%. Chapter 7 shows the result with both lower together, and with each alone down to half.
• Most AI spending stays onshore. 85% of what Australia spends on AI services is captured in Australia, in data centres, energy and local firms. AI is assumed to be cheap, 8% of the pay of the work it replaces in 2030 and a tenth of that by the 2040s, so spending on it is a small share of output. With none of it onshore the package still passes, with a worst five-year average debt of 67.9% of output rather than 62.5%. It would matter much more if AI were expensive (chapter 7).
• Growth. Trend growth is 2% a year in both worlds, the Intergenerational Report's projection. The AGI world adds one percentage point a year from 2031 for AGI diffusion: tools made by AGI lift productivity in work AGI does not itself do, as a better battery, designed with AGI's help, makes a tradesperson faster without AGI doing any of the trade. Three per cent is also Australia's average over the past forty years, so for an economy transformed by AGI it is a conservative figure.

4.15 A worker's super in 2030

Consider what happens to someone who is still working in 2030:

• They lose their employer's super contribution.
• They pay more income tax, 33% from the tax-free threshold and 45% from $65,000 in 2026 dollars.
• And the GDI pays them $3 a day.

Against that, prices at the shops are 6.8% lower than they would otherwise have been.

On the model's measure the median person ends 2030 3% ahead of today's system, the 10th percentile 5% ahead and the 90th percentile 2% ahead. The measure compares income a household can spend now. Super contributions are saving for later, so under today's system they do not count as income either, and their loss does not show.

What replaces them is the fund and the GDI. A member keeps their balance in the national fund, which pays them 2.3% of it every year from the start rather than nothing until retirement. And super's purpose, an income in retirement, is taken over by the GDI, paid for life and rising with output per person. A 2030 worker gives up a stream of contributions into their own account, and gains a guaranteed income for life, paid for every day they are not working, that by 2036 is $147 a day and by 2045 $248 a day, which buy what $114 and $171 buy today. And if AGI takes their job before the GDI has reached its full rate, they are paid today's unemployment payment as well.

This chapter follows D18 from 2030 to 2055. The tricky build-out years, 2031 to 2035, get a more detailed examination in a chapter of their own (chapter 6).

5.1 Work and output

Employment falls over five years. From 14.7 million in 2030 it drops to 12.7 million in 2033 and 9.1 million in 2035, and reaches its lowest point, 6.5 million, in 2042. It then recovers slowly, to 7.0 million by 2055. A no-AGI Australia would employ 18.2 million by then. In the model, displaced workers do not find other jobs. They leave paid work, and the unemployment payment and then the GDI are their income.

The relatively slight recovery in employment comes from population growth. Employment grows 8.2% between 2042 and 2055, while the population grows 11.6%. In the model the human share of each industry's work is set by how far AGI has reached into it, so growth in what the economy can produce creates no new jobs for people: AGI and capital supply it. More spending does create some, because employment moves with the share of capacity in use, but few (chapter 16). That is probably too pessimistic. Richer households may want more of what we currently, at least, think people do best, such as care, hospitality and personal services, and some displaced workers would find that work. Chapter 2 shows the package still passes if many of them do.

Output grows much faster than it would without AGI. Real output is 23% above its 2030 level by 2035, 71% above it by 2040 and nearly three times its 2030 level by 2055. By 2050 it is about 69% larger than in a no-AGI Australia. Output per person more than doubles between 2030 and 2055, against a third with no AGI. In the dollars of each year, output is $3,302 billion in 2030, $4,996 billion in 2040 and $11,651 billion in 2055.

Who earns it changes completely. Wages fall from 46% of output in 2030 to 32% in 2035, 25% in 2040 and 22% by 2055. After-tax profit rises from 12% of output to about 26%. Most of what households receive now comes as the GDI, the national fund's payouts and dividends, not as pay. Those still working are well paid: average pay per worker rises from $104,000 in 2030 to $190,000 in 2040 and $373,000 in 2055, in the dollars of each year, which buys what $101,000, $145,000 and $203,000 buy today. In 2055 that is 2.35 times what the GDI pays an adult for a full year.

The economy does not use all it could produce. It runs close to full capacity through the build-out, at 98% to 99% from 2031 to 2036, and from 2038 it uses between 81.7% and 88.5% of its capacity, 84.8% on average. Use is highest in 2038, falls to about 85% while the public investment programme is still growing, and falls again once the programme reaches its cap in 2047, to 83% in 2050 and 82% in 2055. The rest is capacity that exists but finds no buyer. Chapter 16 asks how much more of it could be put to use.

5.2 People by fifth

The table follows each fifth of people, ranked by purchasing power per equivalent adult each year, against the same fifth in 2030.

The transition starts slowly. In 2031 the GDI pays only $4 a day and the levy is still small, and the four lower fifths slip by less than 1%. They gain in 2032, and the second and middle fifths slip again in 2033, by 1.7% and 1.4%. That is the year jobs start to go in large numbers, 1.3 million of them, while the GDI is still only a fifth of its full rate and shop prices rise 5.7% (chapter 6). By 2033 the poorest fifth is 3% ahead of 2030, and the second and middle fifths are level with it. The fourth fifth is the one that falls: 4.3% below 2030 in 2033 and 7.7% in 2034. These are households that have kept good wages and pay the transition's taxes. The top fifth, with more of its income from investment, stays ahead of its 2030 level throughout. From 2035, as the GDI reaches its full rate, every fifth gains strongly.

The poorest fifth then stands still. Its purchasing power is 41% above 2030 in 2040 and between 41% and 44% above it from 2041 to 2048, as today's pensions and benefits and the unemployment payment are withdrawn; it falls slightly in 2045 and 2047, by 0.1% and less. From 2049 it rises again. Taken together, the bottom three fifths gain in every year from 2034. By 2055 every fifth has between 1.63 and 1.84 times its 2030 purchasing power.

One caution about the middle fifth. The model gives today's middle fifth about half the pensions and benefits it actually receives, because its household types spread adults who do not work evenly across the top three fifths. So in the model the middle fifth has less to lose as today's transfers are withdrawn than it would in reality. The build-out years are hardly affected, because only 5% of today's transfers have gone by 2033; but by 2040, when 60% have gone, the middle fifth's gains are overstated somewhat. The bottom three fifths together are not affected, because the model gives them about the right share between them.

The top two fifths also pay deferred pay in 2031 to 2035. Their income does not change, but what they can spend does. At its lowest, in 2034, the fourth fifth can spend 11.4% less than in 2030; the top fifth's lowest is 7.3% less, in 2033. They get it back, indexed, from 2038.

Against today's system, each fifth ends between 1.65 and 1.87 times as well off as the same fifth under today's system: the poorest fifth 1.87 times, the second 1.87, the middle 1.82, the fourth 1.65 and the top 1.84. The purchasing-power test reads four points of the distribution: the people at the 10th, 20th, 50th and 90th percentiles end with 1.73, 2.00, 1.84 and 1.89 times the purchasing power of the people at those points under today's system. Both compare positions in a ranking made afresh under each system, not the same people. Followed household by household, the gains are larger at the bottom and smaller at the top, because the GDI is a flat amount and carries those who start lowest furthest up the ranking: by fifth of people as ranked under today's system, the average household ends 2.07, 2.11, 1.92, 1.67 and 1.55 times as well off. A no-AGI Australia that kept today's policies would leave the four percentile points 1.25 to 1.27 times better off. So D18 puts them between about 40% and 60% ahead of that as well.

Households as a whole are never worse off under D18 than under today's system. Their total purchasing power is at its lowest in 2031, at 1.026 times today's, and reaches 1.79 times by 2055.

5.3 Who is worse off than under today's system

Averages hide the households the transition treats worst. The model counts, in each fifth of people ranked by their purchasing power under today's system, those who live in households more than a tenth worse off than today's system would leave the same household that year:

Nobody in the poorest fifth is in that position in any year, and nobody in the second fifth after 2034. That is the unemployment payment's doing: paying displaced workers the GDI only, the figures for 2033 would be 10%, 16% and 19% for the poorest, second and middle fifths (chapter 4).

The households that are worse off are in the upper half and the build-out years, and they are there for two reasons. Some have lost a good wage, which nothing in the package replaces: a couple with children, a mortgage and two wages totalling $155,000, one of whose earners is displaced, is at 0.80 of today's system in 2033. Others have kept their jobs and are paying for the transition, through income tax of 45% from $65,000 and shop prices pushed up by the levy. On average the fourth fifth is about level with today's system through the build-out, between 0.97 and 1.02 of it from 2031 to 2035, and the top fifth is behind it, at 0.95 in 2033 and 0.91 in 2034. The fourth fifth is ahead from 2036 and the top fifth from 2037, and by 2055 the fourth fifth is 1.67 times as well off as under today's system and the top fifth 1.55 times.

These counts rest on the household layer, which pays displaced workers between 74% and 87% of what the budget pays them in total, because its household types contain fewer earners than the economy has workers. The budget's figure is the one the fiscal test uses. The counts here, and the Gini coefficient, rest on the smaller one. For each displaced worker the two pay the same.

5.4 People 67 and over

The package is set so that people 67 and over do not go backwards in the early years. Fund members of that age draw 5% of their balance a year, about what today's rules require at 65 to 74, and the draw is untaxed (chapter 4). Households 67 and over, against the same households under today's system:

Households that live on their own savings start 1% behind today's system, are 3% behind in 2031, 1% behind in 2032 and 2033, and 7% ahead by 2035. Those on a part pension or mostly on the pension start 16% and 9% ahead, mainly because the 5% they may draw from the fund is more than these households draw from their super today, and are 21% and 17% ahead by 2033, as the GDI arrives on top of a pension that is withdrawn only slowly. The model recognises these households by their circumstances, with no earner and an income from super, and its group living on their own savings is smaller than the real one, so the figures for that group are indicative.

5.5 Owners, mortgage holders and renters

Every tenure gains, and by similar amounts:

Housing does not get cheaper the way goods and services do, because AGI does not make land. And lower interest rates raise house prices. Mortgage rates stay near 6% through the build-out, fall to 2.5% in 2040 and 2041, and are about 4% from 2050. A lower rate lets the same repayment support a larger loan, and buyers bid prices up to match, over about ten years. By 2055 house prices are 3.93 times their 2030 level in dollars, and 2.20 times against prices at the shops; with no AGI they would be 2.32 times and 1.23 times. Mortgaged households lag in the mid-2030s, mainly because most of them have an earner, and households with an earner gain least in the build-out, but they catch up.

5.6 Pensioners and the end of today's transfers

The no-loss rule (chapter 4) stretches the end of today's pensions and benefits past the schedule, so that on average a recipient gains as much from the GDI's growth as they lose from the transfer's withdrawal. In D18 40% of today's payments are still paid in 2040, 17% in 2045 and none from 2048.

The rule protects the average recipient in dollars, not every recipient, and not against rising prices. From the late 2030s to 2047 the households that depend on today's payments stop gaining:

Households that live mostly on the pension gain 40% by 2038 and then gain only 4% more in the nine years to 2047. Working-age households already on today's payments do less well: they lose a tenth of their purchasing power between 2038 and 2047. In 2047 they are still 36% better off than under today's system, but a decade of standing still or going backwards, while other households gain, is the price of withdrawing today's payments in the 2040s, and a better design would look for a gentler path for these two groups.

Single pensioners have most to lose. Today's pension pays a single person about a third more than each member of a couple, while the GDI pays every adult the same. A supplement for single adults is one of the design questions the package leaves open (chapter 17).

5.7 Inequality

The Gini coefficient of purchasing power per equivalent adult, counting each person once, is 0.177 under today's system. Under D18 it starts at 0.163 in 2030 and falls to 0.124 in 2036. It then rises to a peak of 0.1748 in 2048, and falls back to 0.167 by 2055. In a no-AGI Australia it hardly moves, staying at about 0.180.

The fall to 2036 comes from the GDI and the unemployment payment. Both pay the same amount to people whatever they used to earn, so they are worth far more, relative to their other income, to poorer households than to richer ones. In 2035 the GDI is 36% of households' purchasing power, today's pensions 9.5% and the unemployment payment 5.8%.

The rise from 2035 to 2048, 0.045 in all, is the transition's supports being taken away. The unemployment payment's end adds 0.022 and the end of today's pensions 0.016. As those payments go, households that still have a wage pull further ahead of those that do not, which adds 0.027. Housing costs, which do not fall with other prices and weigh most on poorer households, add 0.011. Income tax and the growing GDI pull the other way, by 0.011 and 0.012. The Gini peaks in 2048, the year the last of today's payments is withdrawn. From 2053 the budget's surplus lifts the GDI, by 6.6% by 2055, and the Gini falls.

Inequality is one of the package's two thin margins. Its peak is 0.002 below the test's limit, and eight of the 118 stress changes in chapter 7 fail the inequality test. The GDI is a floor, not a ceiling, and nothing in the package limits what anyone can earn or own.

The model's Gini coefficients are much lower than the ABS's, which put the Gini coefficient of disposable household income at about 0.32.¹ The model measures inequality only between its 95 household types and their displacement scenarios, each represented by a typical household. It cannot see the spread of incomes within a type, which is much of the inequality in a real population, or the very top of the distribution. So its absolute level cannot be compared with the ABS's. Its changes can be read, because the inequality test compares the package with today's system on the same measure.

5.8 The budget: the transition, borrowing to invest, then surpluses

The state becomes much larger in money terms, because most of the GDI is paid out of taxes. Revenue rises from 42% of output in 2030 to 61% in 2037 and 65% by 2055. Where it comes from changes too:

The budget runs in four phases.

• 2030 to 2033: small surpluses. The budget opens with a surplus of 2.2% of output, as higher income tax arrives before the GDI does, and stays in small surplus to 2033, while few jobs have yet gone.
• 2034 to 2038: the transition's peak. The GDI completes its phase-in, today's pensions are still being paid, and the unemployment payment grows to its largest. The deficit is 1.2% of output in 2034, 3.7% in 2035, 4.2% in 2036 and 4.3% in 2037, $178 billion in that year's dollars, and 3.2% in 2038. It is a short-term deficit of the kind that may be argued as justifiable when the trend is towards balance.
• 2039 to the early 2050s: borrowing to invest. From 2039 the budget borrows about 3% of output a year, the limit the package sets, and every dollar of it, and more, goes into the public investment programme. Without the programme the budget is in surplus from 2040, by 1.3% of output in 2040 and 8.1% in 2045. The programme is designed to earn an income, to use capacity that would otherwise sit idle, and to sell abroad or replace imports so that it pays its own way with the rest of the world.
• From 2052: surpluses. Once the programme reaches its cap, in 2047, it stops growing, and the deficit narrows. The budget is in surplus from 2052, by 1.4% of output in 2055. With debt below 50% of output, from 2053 the surplus lifts the GDI by up to 5% a year, and the rest repays debt. These surpluses come while the economy has spare capacity; chapter 16 discusses what else they could do.

Government debt here is gross debt, what all three levels of government owe (chapter 3). It never rises above its 2030 level of 64.7% of output. It falls to 58.4% in 2033, as the early surpluses and a growing economy outpace borrowing, rises back to 64.1% in 2037 as the GDI and the unemployment payment are paid while the old system is still in place, and falls to 61.0% by 2040 and 36.9% by 2055. In dollars the debt keeps growing while the budget is in deficit, from $2,280 billion in 2035 to $4,711 billion in 2050, and is $4,302 billion in 2055. The ratio falls because output, in dollars, more than triples between 2035 and 2055, and because the interest bill is small: 2.4% of output in 2035 and 0.9% in 2055. From 2050 the government borrows at about the rate of inflation, so its debt costs it about nothing in real terms. The debt's worst five-year average is 62.5% of output, against a limit of 71.2%.

5.9 The rest of the world

The AI build-out is financed partly from abroad. The current account deficit widens from 1.9% of output in 2030 to 5.8% in 2034 and 6.4% in 2036, and is still 4.4% in 2037. Net foreign liabilities rise from 22% of output to a peak of 34.5% in 2037.

After that the external account improves steadily. The current account deficit narrows to 3.9% of output in 2040 and 1.3% in 2045, and turns to a surplus of 1.8% in 2050 and 4.0% by 2055. Australia becomes a net creditor from 2048, and by 2055 it owns, net, foreign assets worth 61% of a year's output. The fund's overseas income, the programme's exports and the falling foreign share of profit all contribute, and so, late in the period, does the idle capacity: an economy that does not buy what it could produce imports less.

The net figure is the difference between much larger gross positions:

The timing item is overseas assets the fund sells in one year to pay for imported AI plant: the sale reduces external debt that year, and the asset leaves the fund's books the next.

In the build-out years every holding of shares falls in value, the fund's overseas assets and foreign holdings of Australian shares alike. That is how the model values shares: at book value, meaning the capital behind them valued at the price of new capital. AGI makes capital goods cheaper, by 7.0% in dollars between 2030 and 2035, faster than companies' retained earnings add to it. Nothing is lost by it: output is measured at the same falling prices, and the companies' profits rise. At market prices, which follow profits, the foreign position would look better rather than worse, because Australians own more shares abroad than foreigners own here (chapter 17). After 2035 the fund's overseas assets grow with its reinvested earnings, while foreign holdings of Australian shares grow more slowly than output.

5.10 Who owns Australian business

The foreign share of Australian business profit falls from 29% in 2030 to 25% by 2035, 24% by 2040 and 19% by 2055. With no AGI it stays at 31.5%. It falls mainly because the fund keeps buying Australian shares back from foreign holders.

Chapter 14 discusses what this ownership costs and what it buys.

The five years from 2031 to 2035 are the hardest stretch of the transition. By 2035 more than six million people have lost their jobs, and more go in the two years after. The AI plant has to be built and paid for, and the budget moves from surplus into its largest deficits. They are where the external position is under most pressure, and where households are most exposed.

6.1 An economy near full capacity

Australia starts the transition at full capacity, as it is today. In the build-out years it has to do two things at once: build the plant and systems that AGI runs on, and keep producing everything households, government and foreign buyers already want:

Investment rises by nearly six points of real output as the AI plant goes in. At the same time 5.6 million people leave paid work by 2035, and 7.6 million by 2037, and the GDI and the unemployment payment replace their wages. Firms take up AI fastest in 2034 and 2035, when a fifth of the economy adopts it each year, so most of the jobs go in 2033 to 2036.

6.2 Consumption has to give way, or be imported

An economy already at full capacity cannot build a large new stock of plant and keep consuming as before. Something has to give. Either households consume a smaller share of what the economy produces, or the extra is bought from abroad.

In D18 it is some of each. Household consumption falls from 41.5% of output in 2030 to 38.4% in 2034, and the current account deficit widens from 1.9% of output to 5.8% in 2034 and 6.4% in 2036. Net foreign liabilities rise by 12 points of output in seven years, to 34.5% in 2037. This is where the external position is under most pressure: every year of the build-out has to be financed.

6.3 The central bank holds the interest rate

Three things restrain spending in these years: interest rates, prices and deferred pay.

The central bank's rule answers spare capacity and excess demand, not the fall in costs AGI brings (chapter 3). Prices before tax fall as AGI spreads, but the economy is still at full capacity, and the bank does not cut. The rate is 5.1% in 2031, 4.6% in 2033 and still 4.3% in 2036, and falls only as capacity comes free, from 2038. Mortgage rates stay near 6%. That holds back household spending and investment, which is why the economy runs a little under full capacity rather than at it, and why less has to be imported.

It also costs the budget, which borrows at 4% to 5% through the build-out, and still at 4.2% in 2037, the year its debt ratio is highest after 2030. A bank that also cut for falling costs, by 1.5 points for each point prices before tax ran below their trend, would take the rate to 1% by 2035 with the economy still full; government debt's worst five-year average would then be 60.0% of output rather than 62.5%. That would make the package look more comfortable than a real central bank would allow.

The bank still "looks through" the levy. Prices at the shops rise 5.7% in 2033 and 6.6% in 2034, and the rate does not respond, on the assumption that a tax's one-off lift to prices does not turn into a wage and price spiral (chapter 3). And in a real economy, spending beyond capacity on things that cannot be imported, such as meals out, haircuts and building work, would push their prices up, which the model does not show (chapter 17).

6.4 The levy's price effect

The pressure on households comes through prices. As displacement speeds up, the AGI levy rises with it, and consumption taxes go from 8% of what households spend before tax in 2031 to 32% in 2034 and 43% in 2035. AGI is pushing prices before tax down at the same time, by 4.7% between 2031 and 2034, but in these years the levy rises faster. Prices at the shops rise 16.8% over the same three years.

The tax lands evenly. In 2035 households in each fifth pay between 22.7% and 25.6% of their net income in consumption tax, against 5% to 7% in 2030.

Incomes keep up, for most. The average pay of those still working rises 14.1% between 2031 and 2034, a little less than shop prices. The GDI rises from $3 a day in 2031 to $47 in 2034 in what it buys. The poorer fifths stand still for the first three years and then gain fast. The poorest fifth slips 0.5% in 2031 and 0.2% in 2033 and gains 5% in 2034 and 11% in 2035. The second and middle fifths slip in 2031 and again in 2033, by 1.7% and 1.4%, the year jobs start to go in large numbers while the GDI is still a fifth of its full rate.

6.5 Buying ahead of the levy

The levy rises over five years, and everyone can see it coming. A household that knows a car or a fridge will carry a much higher tax in two years' time has a reason to buy it now. In 2033 and 2034 the interest rate is also below shop inflation, by 1.1 and 2.2 points, so borrowing to buy early costs little. Korinek and Lockwood show that a consumption tax falls on existing wealth as well as on income, a result that holds for a rate that stays constant; a rate that rises over time also changes when people choose to spend (chapter 12). The model does not show it, because households spend a share of their cash income each year whatever happens to prices next year.

In the build-out, buying ahead would work against the package. It would add spending in the years the economy has least room for it, mostly on goods that are imported, and take it away later, when there is capacity to spare. Deferred pay, below, leans against it for the households that pay it, since it takes spending power away in exactly those years. Taxing durable goods when they are bought on the levy's later rate, or phasing the levy in faster for them, would close the gap; the model leaves both open (chapter 17).

6.6 The unemployment payment

The build-out is when the unemployment payment matters (chapter 4). By 2033, 2.3 million workers have been displaced, by 2035 6.3 million and by 2036 7.6 million, and the GDI does not reach its full rate until 2036. A displaced worker is paid today's unemployment payment on top of the GDI, less the share of today's payments already withdrawn.

Without it, on D18's other settings, the poorest fifth would fall 9.3% in its worst year, the second fifth 2.8% and the middle fifth 2.0%, and in 2033 10% of the people in the poorest fifth, 16% in the second and 19% in the middle fifth would live in households more than a tenth worse off than under today's system. With it, the figures are none, 9% and 18%. Without it, the package fails the inequality test (chapter 4).

It adds spending at exactly the time the economy has least room for it, and some of that spending is imported. On its own it adds 3.9 points of output to the peak of net foreign liabilities and 2.0 points to government debt's worst five-year average. Phasing in the GDI with the jobs lost, rather than on a calendar, makes room for it: the GDI is small while the payment is doing most of the work.

6.7 Deferred pay: Keynes's 1940 answer

In 1940 John Maynard Keynes faced a similar problem. Britain's economy was at full stretch making war material, workers' incomes were rising, and there was nothing more to buy. Taxing the extra income away for good would be unjust to working people; letting them spend it would bid up prices. In How to Pay for the War he proposed deferred pay: part of each person's income above a threshold would be paid into an account, and returned after the war, when the economy had capacity to spare.

D18 does the same in the build-out. From 2031 to 2035 each adult's income above $100,000 a year, in 2026 dollars, is deferred at a rate between 12.4% and 14.0%, raising 2% of output a year:

On top of income tax of up to 50%, that is a heavy burden. It falls on those who can best bear it: the fourth fifth's spending power is at its lowest in 2034, at 88.6% of its 2030 level, and the top fifth's at 92.7% in 2033.

Who pays also decides how much spending it holds back. The households that pay would have saved some of that income anyway, since the better-off spend a smaller share of what they earn, so $70 billion of deferred pay restrains spending by less than $70 billion. The model takes it out of spending at the payers' own rate of spending (chapter 3).

The money is not taxed away. Balances are indexed to prices, so they keep their value, though they earn no real interest (chapter 17). Everyone who paid gets it back.

6.8 Release only into spare capacity

Repayment starts only after a year in which the economy uses less than 97% of its capacity. In D18 the economy is still above that in 2036, so repayment runs from 2038 to 2047, $40.5 billion in the first year and, with indexation, $50.5 billion in the last.

Tying repayment to spare capacity matters. On a fixed timetable, repayment would add spending in any version of events where the build-out ran late and the economy was still full. Released only into spare capacity, deferred pay does what Keynes intended: it moves spending from years that cannot absorb it to years that can.

6.9 Without deferred pay

Take deferred pay out of D18 and the package still passes all four tests. What changes is the external position: net foreign liabilities peak at 37.5% of output instead of 34.5%. The budget does a little better in the build-out without it, because the spending that deferred pay holds back would have paid levy: debt's worst five-year average is 60.6% of output rather than 62.5%, though debt ends higher, at 39% of output rather than 37%. Deferred pay is there to protect the external position in the years when the AI plant is being imported, and the budget pays something for that protection.

A package that passes four tests at one setting of a hundred-odd assumptions proves little on its own. This chapter moves the assumptions, one at a time, to values a reasonable person could have chosen instead, and reports what breaks. It also runs D18 at the pace of earlier technologies, and sets it beside the closest alternative packages.

7.1 The world

The single most important assumption is about the rest of the world. If other countries go through the same transition, they buy more of Australia's exports and the fund's overseas investments earn more. If Australia goes through it alone, neither happens. The model sets the two separately, and D18 puts both at 75%: the rest of the world makes three quarters of Australia's gain from AGI. Moving both together:

D18 passes from 63% up. At 60% the fiscal test fails on one of its two limits: debt's worst five-year average is 67.0% of output, inside its limit of 71.2%, but debt ends at 57% against 54.8%. At 50% it fails clearly, with debt ending at 68% of output, though the external test still passes. Below 50% the fiscal test fails by more, and at 25% the external test fails as well.

The two halves need not move together. A world can share AGI's profits and still not buy Australian goods, or buy them and not share the profits. Moved one at a time:

The world's profits can fall to half on their own and D18 still passes. Its demand for exports matters more: with export demand at 50% the fiscal test fails, unless the world's profits are at 100%. Export demand matters more to the budget, because it decides how much of the economy's capacity is used and so how much the levy collects. The world's profits matter more to the external position, because they are the income on Australia's foreign assets.

D18 assumes 75%. Incentives to adopt AGI are much the same everywhere, especially among Australia's trading partners and in developing economies, so it seems reasonable to expect most of the world to go through much the same transition. Two thirds is a cautious case, and D18 passes it, with debt's worst five-year average at 65.0% of output. Half would be very low. The purchasing-power test passes at every setting in both tables. How much of AGI the world shares decides whether Australia can pay its way, far more than how its households fare.

There is another way to see this assumption. Korinek and Lockwood argue that once AI systems own much of the world's capital, the right response is to tax that capital directly (chapter 12). That tax belongs to the countries that host AGI. Australia, which buys its AI, cannot levy it. Its stand-ins are the fund's overseas holdings, the levy on imports, and the world's sharing in AGI: a world that grows richer buys more from Australia and pays more on Australia's investments. If AI's owners abroad reinvested their profits rather than paying them out, the fund's overseas income would be the exposed part. D18 still passes with that income at 2% a year rather than 3%, though debt ends at 52% of output, close to its limit.

7.2 The pace of AI: lessons from earlier technologies

D18's technology moves fast: firms go from 10% to 90% of the way to using AI in five years, and the frontier moves fastest in 2035 (chapter 4). No earlier technology spread that quickly. History can say something about how fast a technology spreads, and how long its productivity takes to arrive. It cannot say how large AGI's effect will be. So each path below keeps D18's size, AI able to take over 78% of today's jobs at maturity, adding 60% to what the economy can produce and costing a tenth of its 2030 price, and takes its speed from an earlier technology.¹

In every path the frontier matures as the technology spreads: it moves fastest at the middle of the take-up, and takes as long to go from 10% to 90% of the way. The productivity lag is the time between the technology being in use and its output gain. Most of electricity's famous forty-year lag was slow take-up, which the take-up column carries; its ten years is the reorganisation of factories around the new power on top. The internet's boost came while it spread. Television and smartphones are consumer technologies with no productivity record of their own, so their lags are nominal.

On D18's policy, every pace passes:

At the pace of any earlier technology, D18 passes all four tests. With the world at two thirds, the four middle paces fail only the fiscal test's 2055 limit, with debt ending at 56% to 58% of output against 54.8%: their transitions arrive later, and the budget has fewer years of AGI's revenue before 2055. At electricity's pace the transition is still under way in 2055: 8.2 million workers have been displaced rather than 11.2 million, and the GDI is fully in only from 2054.

The reason is that D18's GDI, and the withdrawal of today's payments with it, follow the jobs AGI takes over rather than the calendar (chapter 4). The slower the technology, the later the budget pays out, and the later the levy that pays for it arrives. Put the GDI back on a calendar, phased in from 2030 whatever happens to jobs, and every pace fails the external and fiscal tests, D18's own included: at D18's pace the worst five-year average debt is 87% of output, and at the internet's 198%. Paid on a calendar, a GDI commits the budget to a date. Tied to the jobs, it commits it to an outcome.

The clock's own settings were also moved one at a time, as part of the stress table below: the year of fastest change from 2033 to 2038, the frontier's S-curve over 6 to 20 years, firms' take-up over 3 to 10 years, the productivity lag from none to three years, the cost lag from none to four years, AI's eventual cost three times as high, and capability in 2030 50 points lower or higher on the capability scale. D18 passes every one. The narrowest is a productivity lag of three years, with debt's worst five-year average at 69.9% of output.

7.3 Growth and AGI diffusion

D18's trend growth is 2% a year, plus one point from 2031 for the spread of AGI-made tools into work AGI does not do itself:

Slower growth leaves households less well off and debt higher, and still passes. Faster growth is what fails. At 3%, the past forty years' average, every household is better off and the budget is stronger, but the distribution stretches: the Gini peaks at 0.182, above today's system's 0.177. The same happens if AGI's eventual gain in productive capacity is 110% rather than 60%: the Gini peaks at 0.183. A richer Australia than D18's would probably need a more generous GDI, or a slower withdrawal of today's payments, to stay inside the inequality test. With debt ending at 22% of output rather than 37%, it would also have the budget to pay for one.

7.4 Interest rates and inflation

D18's central bank does not cut for falling costs, and its rate is never below underlying inflation (chapter 3). The rate is still 4.3% in 2035, falls to 1.0% in 2040 and 2041, and from 2050 is about 2.5%, level with inflation, so the government pays about nothing in real terms on its debt from then on. Each of those features can be moved:

D18 does not depend on cheap money. It passes with the rate never below 2%, 3% or 4%, or held at 5% for the whole 25 years. With the world at two thirds, floors of 3% and 4% fail the fiscal test; at 4% the inequality test fails as well.

Lower ordinary inflation is a different matter, because debts are fixed in dollars and inflation is part of what shrinks them against output. At 2%, 1.5% and 1% D18 passes, but at 1.5% and 1% it fails if the world is also at two thirds. With no inflation it fails. The model assumes the Reserve Bank meets its target; with a sixth of capacity idle from the 2040s, that may not be easy (chapter 17).

The table also shows what low rates do to house prices. Had the rate been allowed to stay at 1%, houses would be 3.11 times their 2030 price against prices at the shops by 2055, rather than 2.20 times; with the rate held at 5% they would be 1.33 times.

7.5 What AI costs, and where it is made

D18 assumes that 85% of what Australia spends on AI services is captured here: in Australian data centres, on Australian power, and in the revenue of Australian firms, where it is taxed. It also assumes AI is cheap: doing a displaced worker's tasks costs 8% of that worker's wage in 2030, falling to a tenth of that as the technology matures. AI spending is then 2.3% of output in 2035, about its peak, and falls to 0.6% by 2045. The large sums in the build-out are for automation equipment, the robots and plant in Australian workplaces, which have to be installed here wherever the AI runs.

With so little spent on AI itself, where it is made matters little while AI is cheap. It matters a great deal if AI is dear:

With AI as cheap as assumed, D18 passes even with none of the spending onshore, with debt's worst five-year average at 67.9% of output rather than 62.5%. If AI is dear, because a few providers can charge close to the pay of the work it replaces, much of the value AGI creates goes to them in their prices, and the package then needs most of that spending to be earned and taxed in Australia. With 85% onshore it passes at every cost tested.

So there is an economic case, as well as one of sovereignty, for Australia hosting the AI it uses, and it strengthens the dearer AI turns out to be. The reasons of sovereignty stand on their own: control over systems a society comes to depend on, their security, and whose law and values they follow.

7.6 Three settings with no outside anchor

Most of D18's settings can be set against something: a statistic, a published estimate, today's practice. Three cannot, and each was moved across its range.

• The levy's maximum. The most any firm pays is 100% of its price before tax in 2030, 115% in 2035 and 120% from 2040. Capped at 110% throughout, D18 still passes all four tests, with government debt ending at 41% of output rather than 37% and 6.6 million people in work in 2055 rather than 7.0 million. Capped at 100% it still passes, with debt's worst five-year average at 68.6% of output and debt ending at 47%. Capped at 80% it fails the fiscal test. So D18 does not need a levy above 100%, though it uses what it buys. Collection is the harder question: a twentieth or a tenth of the levy uncollected still passes, but a quarter fails the fiscal test, with the worst five-year average at 76.7% of output, and half fails the external test as well.
• The adoption brake. How far AI adoption is held back when capacity is idle. Anywhere from none to twice D18's setting, every test passes; employment in 2055 runs from 6.1 million to 7.8 million.
• AGI's eventual gain in productive capacity. D18 assumes 60%. Lennox and Dixon's reference case, 1.31% a year of extra productivity over about two decades (chapter 8), corresponds on this model's shorter window to about 2.8% a year, which on the technology clock is an eventual gain of about 18%; their fat-tailed case corresponds to about 27%. D18 passes at both, narrowly at 18%: debt's worst five-year average is 70.3% of output against the limit of 71.2. A smaller gain leaves less capacity standing idle, which partly offsets the smaller economy. So D18 does not rest on productivity well above the published estimates, though its margin then becomes thin.

7.7 The stress table

Of 118 single changes, 21 break at least one test. Fourteen fail the fiscal test, nine the external test and eight the inequality test. They fall into nine groups.

AGI arrives and does little. Only one technology tested fails: if AGI disappoints, taking over jobs on the occupation scores as published but adding only 5% to what the economy can produce, both the external and fiscal tests fail badly. With the technology of the published studies, Lennox and Dixon's or Jobs and Skills Australia's, D18 passes all four tests (chapter 8). In those worlds few jobs go, so the GDI phases in only part of the way and today's payments are only partly withdrawn: the package scales itself to the shock. And if firms pass on only 60% of their cost savings, debt ends at 62% of output, over the limit of 54.8%.

AGI does more. If AGI's eventual gain is 110% rather than 60%, or trend growth is 3% rather than 2%, the inequality test fails (above).

The clock. Every setting of the technology clock tested passes, and so do the smartphone's and television's paces (above).

The economy's closure. The package fails if the state keeps its share of output (chapter 15), or if households save much more of their income (a spending rate of 80% rather than 87%, before the effects described in chapter 17), which fails the inequality test as well as the fiscal one. It fails with the published studies' assumption that everything the economy can produce is sold, which turns the transition into a demand boom the external account cannot finance. And it fails the inequality test if employment does not grow with the population, or if today's transfers are withdrawn on their schedule without the no-loss rule: the poorest fifth then falls by 5.7% in one year.

Interest rates. D18 fails the fiscal test if lenders charge a risk premium of 0.1 of a percentage point for every point of debt above 40% of output, and passes at half that.

Shocks from abroad. The package fails the fiscal test if the Australian dollar is 15% higher. It fails the external and fiscal tests if a 50% share market fall hits in 2035; a 35% fall in 2040 passes. It also fails both if the fund sends 80% of its money overseas, because it would have to sell Australian shares to foreigners to do so.

The fund. If the fund reinvests none of its earnings, or if every member draws 5% of their balance a year, net foreign liabilities never come down far enough, and the external and fiscal tests fail. If it reinvests half its earnings, the inequality test fails by a whisker, with the Gini peaking at 0.1773 against 0.1770. D18 still passes if the fund's overseas assets yield 2% a year in income rather than 3%, or grow by 1.5% a year rather than 3%; in both cases debt ends at 52% of output, not far inside the limit.

The levy. A quarter of the levy uncollected fails the fiscal test; a tenth passes.

The package's own settings. A public investment programme of fixed size, spending through the build-out as well as after it, fails badly. What the programme does depends on when it spends: into spare capacity it pays its way, into a full economy it is mostly imports. A GDI of $90 a day rather than $105 fails the inequality test, with the Gini peaking at 0.183. And the GDI phased in on the calendar rather than with the jobs fails the external and fiscal tests (above). Company tax at 30% throughout rather than 35% passes, but only just: debt's worst five-year average is 71.0% of output.

One more of the 21 is a matter of measurement rather than of the world: if the household layer gave every owner one blended housing cost rather than treating renters, mortgage holders and outright owners separately, the Gini would peak at 0.180, over its limit.

7.8 The 90-cent variant, and others

D18 is not the only package that passes. The closest alternatives differ in how fast the GDI is withdrawn against a day's wage. Korinek and Lockwood argue that a transfer should not depend on whether its recipient works (chapter 12), and a lower withdrawal rate moves towards that:

The 90-cent variant lets a person keep ten cents of each dollar earned on a working day. It passes all four tests in both worlds, at the cost of a little more debt and a slightly smaller GDI. It is a little less robust: it fails five of the stress table's changes that D18 passes, among them the fund's overseas income at 2% and company tax at 30%, and passes one D18 fails, the fund reinvesting half its earnings. The 80-cent variant, even with the levy's maximum raised to 110% in 2030 and 120% in 2035 to pay for it, fails with the world at two thirds.

In the model, almost no group of households is better off under the 90-cent variant: what earners keep on their working days is offset by a GDI that is smaller for everyone. Its benefit would come from people working more, especially part-time in the build-out years when the economy is short of capacity, and the model has no labour supply that responds to it. So the model shows what a lower withdrawal rate costs, and not what it would buy. Chapter 12 discusses it further.

Other responses pass too. A GDI fully in only once half the jobs are gone, on a later phase-in, passes with less debt, at the cost of households falling behind today's system in the build-out (chapter 4). A smaller public investment programme leaves debt much lower: with a cap of $150 billion a year, debt ends at 14% of output, and the 10th percentile ends a little better off, at 1.80 times today's system rather than 1.73. A heavier top income tax eases both of D18's thin margins (below). Every one of these can be run on the model's page.

7.9 Where the fragility sits

D18 has two thin margins.

The first is the budget. Government debt's worst five-year average is 62.5% of output against a limit of 71.2%, and anything that weakens the budget through the 2030s takes it over: AGI doing far less than assumed, firms keeping their cost savings, a dollar that is too high, levy that is not collected, a world that buys less, households that save much more, lenders charging a premium on debt, or a state that keeps its share of output. With the world at two thirds the margin on debt in 2055 is thinner still: 50% against 54.8%.

The second is inequality. The Gini's peak, 0.1748 in 2048, is 0.002 below the limit. It comes in the year the last of today's payments is withdrawn, and before the budget's surplus has started to lift the GDI. Anything that makes the incomes of those still working or owning grow faster than the GDI in the 2040s takes it over: faster growth, more productive AGI, a smaller GDI, or employment that does not grow.

The two margins pull against each other. Withdrawing today's payments later would ease the inequality test and tighten the fiscal one; withdrawing them sooner would do the reverse (chapter 4). A smaller GDI, which is a way to buy fiscal margin, fails the inequality test. Of the changes tested, the one that eases both by much is heavier income tax at the top: with 54% from $121,000 and 49% from $100,000, debt's worst five-year average is 60.0% of output and the Gini peaks at 0.167. The self-imposed limit of 50% rules that out. A government adopting D18 would be relying on both margins holding through the 2030s and 2040s, and would want to know in advance which of its limits it would relax first if they did not.

The external position is no longer where the package is tightest. Net foreign liabilities peak at 34.5% of output against a limit of 82.1%, and the government's own limit on foreign borrowing never comes within 2.1 points of output of binding. The purchasing-power test never fails.

7.10 What the package survives

Ninety-seven of the 118 changes leave all four tests passing. Among them:

• the technology: AGI half as capable or twice as capable at maturity; AI costing 20% of the wages it replaces rather than 8%; an eventual gain of 25% rather than 60%; the technology of Lennox and Dixon or of Jobs and Skills Australia; every setting of the technology clock tested, and the smartphone's and television's paces;
• growth: trend growth of 1.5%; half the diffusion of AGI-made tools, or none;
• the premise: a quarter, half or all of the displaced re-employed (chapter 2);
• interest rates and the world: no rate rule, with the rate held at 5%; a risk premium of 0.05 of a percentage point for every point of debt above 40% of output; world rates of 3% or 7%; the world not following Australia's rates at all; a lower dollar; a 35% share market fall in 2040;
• the fund: the fund entirely in Australia or half overseas; overseas assets earning 7%, 5% or 4.5% rather than 6%; a fifth of its earnings paid out rather than reinvested;
• the programme: a cap of $150 billion, $250 billion or $600 billion rather than $1,000 billion; a capital-output ratio of 4 or 10 rather than 7; half as much of it exportable; 35% imported; returns and export prices falling twice as fast;
• the levy: a twentieth or a tenth of it uncollected; half the levy on imports uncollected; shoppers switching harder away from heavily taxed goods; the levy held flat at 110%;
• the package: no deferred pay; no buy-back of foreign-held shares; the GDI fully in at 30% or 70% of jobs displaced; company tax at 30%.

James Lennox and Janine Dixon's Occupations, Tasks and Generative AI: A Computable General Equilibrium Analysis (Centre of Policy Studies, Victoria University, working paper G-367, May 2026) is the closest Australian study to this one. It uses the same data on which occupations AI can augment and which it can automate, and it asks what AI does to the whole economy. It reaches a very different answer because it models a fixed aggregate labour supply.

8.1 What they modelled

Lennox and Dixon build a task-based computable general equilibrium (CGE) model of Australia with 45 industries and 97 occupations. Each occupation's tasks can be done by people alone, by people augmented by AI, or by AI alone, using Jobs and Skills Australia's separate scores for augmentation and automation potential. They solve for a long-run equilibrium in which generative AI "is fully integrated into production", "which we interpret as being roughly two decades".

In their reference case, "the reference simulation yields a 29.8% real GDP increase: roughly one third from task-level productivity gains, the rest from capital deepening and general equilibrium reallocation. Real consumption—our long-run welfare metric—rises by 16.2%." About 47% of tasks end up augmented and 9% automated, and the labour share of income falls from 60% to 52%. The labour market's adjustment is large but mostly within occupations: "extensive-margin task reallocation equivalent to two thirds of current work." Their sensitivity cases with a fatter-tailed distribution of task productivity roughly double the GDP gain.

8.2 The closure

The difference that matters is one sentence: "Aggregate labour supply is fixed, as is common in the CGE literature." Wages in each occupation adjust until everyone who wants work has it. The authors are clear about what that leaves out: it lets them "focus squarely on the long-run reallocation of labour within the economy, abstracting from potential responses of hours, participation and equilibrium unemployment."

That is a reasonable choice for their question, which is how AI reshapes the economy once it has settled. It is the opposite of this model's premise, and the two models are built differently as well: theirs finds the prices that clear every market in a long-run equilibrium, while this one follows the economy year by year with output limited by spending (chapter 3). In their model every displaced worker finds work by construction; in this one, none does. Neither is a forecast: they are the two ends of the range, and chapter 2 shows the package passing its tests at points in between.

8.3 Their technology in this model

The model has a preset that reproduces Lennox and Dixon's technology assumptions: their split of augmentation and automation, their 1.31% a year of AI productivity growth, which on the model's technology clock amounts to AI eventually adding about 8% to what the economy can produce, and all displaced workers re-employed. It keeps the model's own year-by-year economy, in which output is limited by spending and investment responds to capacity and interest rates. It can be run with today's policies or with D18's:

With today's policies, their technology gives the model about 10% more output than a no-AGI Australia by 2050, with employment close to the no-AGI path. That is well short of their 29.8%. The model reproduces something closer to the part they attribute to task-level productivity; the rest of their gain comes from capital deepening in a long-run equilibrium, which the model's year-by-year investment does not reach in the same time.

Today's policies fail two tests here, but the no-AGI baseline fails the same two (chapter 11). On this technology AI does not break today's system.

D18 on their technology passes all four tests, and households consume 15% more than in a no-AGI Australia by 2050, close to Lennox and Dixon's own 16.2%. The package does not overreact, because its GDI follows the jobs AGI takes over. With their technology only 7% of jobs go by 2055, so the GDI phases in only about a fifth of the way, and only 5% of today's payments are withdrawn. The levy, charged in proportion to each industry's automation, stays small too.

8.4 What the comparison shows

The two models disagree about the labour market, not about the technology. Given Lennox and Dixon's technology and labour market, this model agrees that today's tax and transfer system copes. Given this model's premise, their framework has no way to express the problem, because it holds employment fixed.

The practical lesson is that a government does not know which world will arrive, and its policy should not have to guess. Today's settings should not be relied on in a world that looks like this model's (chapter 2). A package whose income and taxes follow the jobs actually lost does little if AGI does little, and a great deal if it does a great deal. That is the case for tying the GDI to displacement rather than to the calendar (chapter 4).

Lennox and Dixon's model was also used for the dynamic projections in Jobs and Skills Australia's 2025 study Our Gen AI Transition, through Victoria University's VUEF model. The published model has a preset for those projections too; it behaves much like the Lennox and Dixon preset, and passes and fails the same tests.

EY's The economic upside of Artificial Intelligence for Australia (5 August 2026) is a short study with a clear method. Its use of Jobs and Skills Australia's occupation-level AI exposure data to drive an economy-wide model inspired this model's approach.

9.1 What they modelled

EY maps Jobs and Skills Australia's Gen AI Transition data from occupations to industries, weighted by employment, and turns it into productivity shocks in EYGEM, EY's computable general equilibrium model. AI both automates labour-intensive tasks and augments workers. Two assumptions keep the effects modest. The productivity gain in any industry is capped to a parameter, Umax: "Umax acts as a reasonable cap on the potential labour productivity uplift from AI adoption, rather than an unconstrained estimate of productivity gains. This was set to 10%." And uptake is phased: "from 20% initially to 100% by 2035–36".

The results are correspondingly modest. AI is "expected to deliver between a 2.0 per cent to 2.4 per cent increase in multifactor productivity and boost GDP by between $95 billion (2.6 per cent) to $116 billion (3.2 per cent) by 2036." On jobs, "there would be a relatively small addition to the labour market: between approximately 36,000 to 44,000 additional full-time equivalent jobs, in net terms", and "some jobs would be restructured or displaced, but new roles created elsewhere in the economy would more than offset those losses."

9.2 Their assumptions in this model

The model has a preset for EY's technology: a small augmentation effect, no automation beyond it, and all displaced workers re-employed.

With today's policies the model gives about 3% more output than a no-AGI Australia, the same size as EY's 2.6% to 3.2%, and employment follows the no-AGI path. The two models agree when they are given the same assumptions.

With D18's policies the result is the same as for Lennox and Dixon's technology. All four tests pass, and households consume 7% more than in a no-AGI Australia by 2045. With only 2% of jobs gone by 2055, the GDI phases in only a tenth of the way, and today's payments are hardly withdrawn: the package scales itself to the shock (chapter 4).

9.3 What the comparison shows

EY describes a world in which AI is one more productivity improvement, adopted steadily, with the labour market absorbing it. In that world today's tax and transfer system needs no redesign, and this commentary has nothing to add.

This model exists because that is not the only world on offer. EY's cap of 10% on the productivity gain in any industry rules out, by assumption, the case this model studies. The cap is reasonable for a ten-year study of the AI now in use. It is not a statement about AGI.

Economic Scenarios for Transformative AI, by Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter and Peter McCrory (The Anthropic Institute, working paper 2026-02, September 2026), is a study of the United States, not Australia. It is useful here for its assumptions: it sets out, in numbers, how large and how fast the effects of AI might be over the next few years.

10.1 Three scenarios to 2030

The paper "presents a framework for assessing the economic consequences of AI between 2026 and 2030". It defines three scenarios by how many cognitive tasks AI can do and how widely it is used, and it says plainly that they are "not predictions, and we attach no probabilities to them".

• Modest change. AI affects 20% of cognitive tasks by 2030 and is used in a fifth of the cases where it could be. It "adds less than half a point to GDP growth by 2030" and raises unemployment by a tenth of a point.
• Substantial change. AI affects 30% of cognitive tasks, used in 40% of cases. By 2030 GDP is 8% higher and cognitive employment 4% lower. This is the scenario closest to what the median respondent to their survey expects.
• Extreme change. AI affects half of cognitive tasks, used in 60% of cases. "GDP growth then rises to 15 percent per year", with GDP 32% higher by 2030. The labour share of income falls from 60% to 45%, and nearly one in five cognitive workers is unemployed. Cognitive wages end 11.5% below their no-AI path, while wages in all other occupations end 34% above it.

10.2 Their assumptions against this model's

Three differences stand out.

Where displaced workers go. In the paper, workers in cognitive occupations who lose their jobs can move to other occupations, and wages there rise. That route is open in all three scenarios. This model's premise is that AGI closes that route too: it reaches physical work after a lag, and displaced workers do not find comparable jobs.

How far it goes. The model's technology goes well beyond the paper's extreme scenario. D18's economy grows by 23% between 2030 and 2035 and by 71% by 2040, against the extreme scenario's 32% gain by 2030. And wages fall from 46% of output to 25% by 2040, a much larger fall than the extreme scenario's labour share, while employment falls by more than half.

What is measured. The paper models the United States and measures the labour share of all income. This model measures wages as a share of Australian output before consumption taxes. The comparison is of orders of magnitude, not precise figures.

10.3 The shorter horizon

The paper stops in 2030 because it models AI's effect on cognitive tasks only: "we do not extend the analysis beyond 2030". Robotics and the automation of physical work would change the picture after that, and the authors leave them out rather than guess.

This model starts where the paper stops. Its AGI economy begins in 2030 and runs to 2055, and it includes physical work, reached by AGI more slowly than cognitive work. If the paper's extreme scenario is where the world is heading by 2030, this model asks what comes next, and what Australia would need in place to meet it.

Treasury's Intergenerational Report 2026: Australia's future to 2066 is the official long-run view of the economy and the budget. It matters to this model in two ways: it supplies the no-AGI baseline's growth rate, and it states Treasury's view of AI.

11.1 Treasury's 2% and the no-AGI baseline

The report's central growth figure is stated in one sentence: "Average annual growth in real gross domestic product (GDP) is projected to be 2 per cent over the next 40 years, below the 3 per cent experienced over the past 40 years." It assumes labour productivity grows 1.2% a year.

The model's no-AGI baseline grows at the same 2%. It keeps today's tax and transfer policies, and by 2055 its output is 64% larger than in 2030, and a third larger per person. The AGI world shares that 2% trend and adds one point a year for AGI diffusion, the spread of AGI-made tools into work AGI does not do itself. With that point the AGI world grows at the 3% of the past forty years before any of AGI's direct gains are counted. That seems conservative for an economy that AGI transforms.

11.2 Treasury on AI

The report puts AI first among the forces shaping the next forty years: "AI is developing rapidly, and will be a defining influence on the economy over the next 40 years." It cites the Productivity Commission's estimate "that AI could boost multifactor productivity in Australia by at least 2.3 per cent over 10 years." It notes that "most Australian and international studies have found limited effects to date", and it expects AI to support its productivity assumption rather than exceed it: "The rise and adoption of AI is likely to support the achievement of Treasury's long-term labour productivity growth assumption over time."

This is the EY world of chapter 9: AI as a strong productivity improvement, with the labour market adjusting. The report does not model a case in which AI displaces most paid work. That is not a criticism. An intergenerational report projects the path that current evidence supports. This commentary asks about a different path, one current evidence does not yet show but that the people building AGI say, and their financiers assume, is coming.

11.3 The fiscal outlook

The report projects the Commonwealth budget to 2065–66. Payments rise to 27.7% of GDP. The underlying cash balance ends in a deficit of 1.8% of GDP. Gross debt falls from 33.1% of GDP in 2025–26 to a low of 22.2% in the mid-2050s, then rises again to 27.4% by 2065–66.

The model's no-AGI baseline looks worse: a deficit of 3.0% of output in 2030 widening to 4.9% by 2055, and gross debt rising from 66% of output to 87%. The two are measured differently in three ways.

• Coverage. The report covers the Commonwealth. The model covers all three levels of government, Commonwealth, state and local, and its debt covers them all. Both measure gross debt, before netting off financial assets (chapter 3).
• Denominator. The model measures output before consumption taxes, which is smaller than GDP at market prices, so its ratios run about a tenth higher (appendix A).
• Policy. The baseline starts from a deficit of about 3% of output for the whole government sector, between the national accounts' and the government finance statistics' measures of it, and assumes no consolidation: no government cuts spending or raises taxes to bring the budget back towards balance. It shows what today's settings would do if left alone, not what a government would do.

The baseline is a comparison for the AGI world, not a fiscal projection. It is worth saying all the same that it fails two of the four tests that D18 passes: its debt ends above the fiscal limit, and its Gini coefficient, 0.180 in 2030, is above today's system's 0.177. Today's settings, left alone, do not meet the standard the package is held to.

Anton Korinek and Lee Lockwood's Public Finance in the Age of AI: A Primer (Brookings Center on Regulation and Markets, working paper, January 2026) asks the question this model asks, from the other end.¹ The model starts from a set of policies and tests whether they survive a transition. Korinek and Lockwood start from the theory of optimal taxation and ask what a tax system should look like once AI does most of the work. Their answer is close to D18's, which is encouraging, and where it differs it points to things the model leaves out.

12.1 What they argue

Their starting point is that transformative AI (TAI) "may gradually erode the two main tax bases that underpin modern tax systems": first income from work, and later the consumption of people. So the tax system should move in sequence, from taxing work, to taxing consumption, and eventually to taxing the capital that AI itself owns.

In their first stage, little paid work is left but people are still the consumers. The most a tax on work can raise shrinks with labour's share of income, towards nothing. A tax on consumption becomes the main instrument. Taxing consumption is the same as taxing work and existing wealth together: "a constant consumption tax is equivalent to a mix of a constant labor tax and a one-time tax on initial capital" (p. 10). Paired with a transfer paid equally to everyone, a high consumption tax moves everyone towards the same level of consumption, without distorting how much people save. For a sense of scale, they estimate that raising all of today's American tax revenue from consumption would need a tax of about 37.5% (p. 15).

Two kinds of tax that matter little today become important. When some goods cannot be taxed, or are easy to buy without paying the tax, or can be made at home instead, a uniform consumption tax no longer falls evenly, and "TAI might reinstate the relevance of Ramsey taxation", meaning different rates on different goods (p. 17). And taxes on the rents from fixed factors, "the classic case" being "unimproved land", become more valuable as other bases shrink (p. 18).

If the consumption tax cannot be raised as far as it should be, because of administration, evasion or home production, then a tax on the normal return to capital becomes worthwhile: "taxing the normal return to capital becomes optimal as a second-best tool for redistribution" (p. 15).

Their transfers are paid as lump sums. Citing Korinek and Juelfs (2024), they argue that "conditioning transfers on work is only desirable if work gives rise to positive utility effects" (p. 9): otherwise people should be free to choose whether to work without it changing what they receive.

In their second stage, AI systems own capital and spend it on their own purposes, and neither a tax on work nor a tax on people's consumption can reach that value. The answer is to tax the AI's capital directly, at a rate equal to the human discount rate, "commonly estimated around 4 percent" (p. 4), or about 11% on less standard assumptions (p. 24).

Their theory compares steady states. It assumes a closed economy, full employment, and taxes that stay constant over time. They do set out a roadmap across the transition (their Table 2, p. 4). Before the first stage come sovereign wealth funds, windfall clauses, broad ownership of AI companies, and steering technology towards augmenting work rather than replacing it. The first stage brings the consumption taxes above, and the second taxes on AI's capital, on computing and on robots. They argue that "proactive institutional adaptation is preferable to reactive crisis response" (p. 5). But their emphasis is on where the tax system should end up, more than on the years of getting there. This model's emphasis is the reverse.

12.2 How D18 maps onto it

D18 sits almost entirely in their first stage, and its core is close to what their theory recommends.

12.3 Where the paper supports D18

The levy and the GDI. Their result on a consumption tax paid back as an equal transfer is very nearly a description of the levy and the GDI. It is why most of D18's reduction in inequality comes from the GDI (chapter 5).

It also offers a less alarming way to state the levy. Like the GST, the levy is set as a rate on the price before tax, and its maximum of 100% in 2030, 115% in 2035 and 120% from 2040 looks extreme. Stated as a share of the price the buyer pays, the way most countries state their consumption taxes, those maximums are 50%, 53.5% and 54.5%. Consumption taxes of 53% of what households spend before tax are about 35% of what they pay at the till: by Korinek and Lockwood's equivalence, much like a tax of 35% on income from work and on existing wealth, if the rate were constant. Their own figure for replacing all of America's taxes with a consumption tax, 37.5% before tax, is about 27% of what the buyer pays. D18's is larger because its GDI replaces most of what wages do today. Chapter 13 sets out the levy's design.

Capital taxes. In their first stage, a tax on capital is worse than a tax on consumption. D18 taxes company profits at 35% and the fund's income at members' own rates, so it leans on capital more than their theory would like. But their own exception applies: a capital tax is worthwhile when the consumption tax is constrained, and D18's levy is. Capped at 100% throughout, D18 still passes, but debt ends at 47% of output rather than 37%. If a quarter of the levy goes uncollected, the fiscal test fails. And with company tax at 30%, D18 only just passes: the worst five-year average debt is 71.0% of output against the limit of 71.2.

Two further reasons fall outside their model. Company tax is the only tax that reaches the share of Australian profit that foreigners own: 29% of it in 2030, and still 19% in 2055. And from 2038 D18 uses only 82% to 89% of its capacity. The harm a capital tax does in their model comes from discouraging the saving that builds capacity, and in D18 capacity is not what is short. Chapters 13 and 14 return to these choices.

One smaller tension. Korinek and Lockwood note that because retirement saving is exempt, today's income tax already works partly like a consumption tax. D18 moves the other way: the fund's income is taxed at members' own rates, about 41% to 44% of it from 2040, rather than at super's 15%.

12.4 Where the paper challenges D18

Housing and land. This is their strongest challenge. D18 levies everything households buy except housing, so housing becomes by far the largest untaxed purchase, and Ramsey's logic says the gap should be closed. Land is also the classic fixed factor, whose rent they would tax more heavily as other bases shrink. The commentary treats rising housing costs as something the GDI has to offset (chapter 4); their framework treats land rent as a base to tax.

The model has a broad land tax, so the trade-off can be sized. At 0.5% of land value a year, D18's debt ends at 26% of output rather than 37%, with the worst five-year average 60.3% rather than 62.5%. But the inequality test fails: the Gini peaks at 0.178 against today's system's 0.177. At 1%, debt ends at 14% and the Gini peaks at 0.180. In the model a land tax falls hardest on owners with little income, such as pensioners who own their homes outright, and the bottom three fifths of people dip to 0.976 of their 2030 level at 1%. The model also has no way for households to shift their spending towards housing as other goods become dearer, which is the distortion Korinek and Lockwood warn of. A land tax designed for an AGI economy, with deferral for owners who are rich in property but poor in income, is a question for the next draft.

Evasion and home production. Their point that evasion, untaxed goods and home production become the real limits on a consumption tax applies with force. In D18 only about 6.5 million to 7 million people are in paid work from the late 2030s, so most adults have time, and a wedge of more than half the price makes doing things yourself, swapping, and cash work far more attractive. The model treats collection as a fixed share: a twentieth of the levy uncollected passes, a tenth passes but fails if the world shares only two thirds of the AGI economy, and a quarter fails the fiscal test, with the worst five-year average debt at 76.7% of output. Their argument suggests that what goes uncollected would grow with the rate, and with the time people have outside paid work. The model does not show that.

The levy's path. Their equivalence holds for a consumption tax that stays constant. One that rises over time is a hidden levy on existing wealth, and it changes when people choose to spend. D18's levy rises through the build-out, and chapter 6 describes the reason that gives households to buy ahead of it.

The GDI's withdrawal. Their transfers are paid whatever people do, and they argue against tying transfers to work. D18's GDI is withdrawn at a dollar for each dollar of a day's wage, so on a working day it is tied to not working: a day's work adds nothing to income until the wage passes the GDI. The model can price lower rates, but not their benefit, because no one in the model works more when it pays to:

At 90 cents a person keeps a tenth of each dollar earned on a working day, and the package still passes in both worlds, at the cost of a little more debt. Below that it passes with the world at three quarters but fails, narrowly, with the world at two thirds, on debt in 2055. A true lump sum, never withdrawn, fails the fiscal test badly. The cost of a lower rate is in the table; its benefit, more part-time work, especially in the build-out years when the economy is short of capacity, is not, because the model has no labour supply that responds. Chapter 7 sets out the 90-cent variant beside D18.

The second stage, for a country that hosts AI but does not own the models. In the model Australia buys AGI and then runs it at home: 85% of what is spent on AI services stays in Australia, in local data centres, energy and firms, and the machines that do the displaced work are installed here. That plant is ordinary capital, owned by Australian firms, households and the fund, and by foreigners, and company tax already reaches its profits. What the model leaves out is the profit from owning the frontier models themselves, which would mostly be earned abroad. It can leave it out because it assumes AI is cheap: doing a displaced worker's tasks costs 8% of that worker's wage in 2030, falling to a tenth of that as the technology matures (chapter 3). History suggests that is plausible. Nordhaus estimated that innovators kept only about 2.2% of the total social surplus from American innovation between 1948 and 2001, and that most of the benefit passed to consumers.² The model assumes the same of the firms that use AI: 90% of what they save by automating reaches their customers as lower prices (Appendix E).

Korinek and Lockwood's second stage goes further: AI systems that own capital and spend it on their own purposes, beyond the reach of a tax on work or on people's consumption. The model has no such systems. If they arose abroad, Australia's stand-ins for a tax on their capital would be ownership, through the fund's 40% of assets overseas; the levy on imports; and the assumption that the rest of the world shares in AGI, which chapter 7 discusses. If the world moved into their second stage, with AI's owners reinvesting rather than paying out, the fund's overseas income would be the exposed part. D18 still passes with that income at 2% a year rather than 3%, with debt ending at 52% of output.

12.5 What the model adds

Korinek and Lockwood compare steady states, with constant taxes, in a closed economy at full employment. D18's hardest years are about the transition: building the AI plant while the economy is already full, financing a current account deficit, holding interest rates, and deferring some spending until there is room for it (chapter 6). A closed-economy steady state cannot see any of that.

The two are complementary. Korinek and Lockwood argue that the sequence, from taxing work to taxing consumption and eventually AI's own capital, is right. D18 tests whether Australia could get through the first step.

12.6 For the next draft

The next draft of this model plans to address four of their points:

• what goes uncollected of the levy, rising with its rate and with the time people spend outside paid work;
• households shifting their spending towards housing, and a land tax designed for owners who are rich in property but poor in income;
• people working more when the GDI is withdrawn more slowly;
• purchases brought forward ahead of a rising levy.

D18 funds a state that pays out 40% of output as the GDI and collects 65% of output in revenue by 2055. Which taxes carry that load is the most political question the package raises. This chapter explains the choices, and what the model can and cannot say about them.

13.1 A consumption tax that follows automation

The AGI levy is the package's largest tax: on its own it raises between 17% and 19% of output from 2040, net of the credits firms claim on what they buy, and with the levy on imports about 21% to 23% from 2037. Each industry pays a rate on its sales equal to a maximum rate, 120% from 2040, multiplied by the share of its human work that AGI has taken over (chapter 4).

Three things recommend it.

• It taxes AGI's output wherever it is produced. Imports pay it as well as local production. A tax on AGI profits can be avoided by booking them offshore; a tax on sales in Australia cannot.
• It is less visible than income tax. An income tax rate of 60% or 70% would be felt, and resented, on every payslip. A consumption tax is built into prices, and prices before tax are falling.
• It makes automation carry a cost. In the model the levy is the only thing that makes the cost of labour matter to a firm's decision to automate. A firm that automates past the point where the levy lifts its prices above those of a fully human competitor loses customers, so it stops short. This is close in spirit to the tax on automation that Falk and Tsoukalas find is the one remedy for the layoff trap (chapter 2), though theirs is charged per automated task and this one on sales.

It is also what the theory of taxation recommends once AI does most of the work. Korinek and Lockwood show that the most a tax on work can raise shrinks with labour's share of income, that a tax on consumption then becomes the main instrument, and that a high consumption tax paid back as an equal transfer to everyone moves everyone towards the same level of consumption without distorting saving (chapter 12). That is the levy and the GDI.

It is also less regressive than it looks. A consumption tax usually takes a larger share of poorer households' income, because they spend more of it. In D18 each fifth pays much the same share of its net income:

The model works out each household's consumption tax as its income after housing costs, times the share of that it spends, times the tax rate on what it buys. Three things differ between the poorest and the richest:

Poorer households spend more of what they have, and on its own that would make the tax regressive. But housing carries no levy, and half the poorest fifth rent, so more of their income goes on something untaxed. And richer households' spending leans a little more towards the heavily automated industries the levy taxes hardest. The three roughly cancel. In 2055 the top fifth pays 0.7 of a point less of its income than the poorest. The gap is wider while the levy is rising, 2.7 points in 2035, and narrows only later. Today's GST, by the same measure, is also close to proportional, and takes about half a point more from the poorest fifth than from the top. By 2055, relative to the tax paid, the levy's gap is much the smaller.

The levy is hardest on households in the years it is rising. Between 2031 and 2034 it lifts prices at the shops by 16.8%, while prices before tax fall. Chapter 6 follows those years.

13.2 A levy above 100%

The levy's maximum is 100% of a firm's price before tax in 2030, 115% in 2035 and 120% from 2040. No tax system levies a rate like that today, and it is fair to ask what it is for. D18 does not need it to pass the tests, but it uses what it buys. Capped at 110% throughout, the package still passes all four tests: government debt ends at 41% of output rather than 37%, and 6.6 million people are in work in 2055 rather than 7.0 million. Capped at 100%, it still passes, with debt's worst five-year average at 68.6% of output against a limit of 71.2% and debt ending at 47%; capped at 80%, it fails. The rate above 100% buys jobs, through the brake the levy puts on automation, and fiscal margin.

A rate above 100% also sounds more extreme than it is. Like the GST, the levy is stated as a share of the price before tax. Stated as a share of the price the buyer pays, as most countries state their consumption taxes, the maximums are 50%, 53.5% and 54.5%. Consumption taxes as a whole, about 53% of what households spend before tax from 2040, are about 35% of what they pay. Korinek and Lockwood estimate that replacing every American tax with a consumption tax would need a rate of about 37.5% before tax, which is 27% of the price paid (chapter 12). D18's is larger because its GDI does most of what wages do today.

13.3 Income tax within the self-imposed limits

Income tax raises 18% of output in 2030 and between 11% and 12% from 2035, because wages shrink and the GDI, though taxable, is a flat amount. D18's scale keeps to self-imposed limits: no rate above 50%, with no Medicare levy on top. Below the top rate it is heavy, 45% from $65,000 and 48% from $135,000, because those rates collect in the build-out, when incomes from work are still large. Heavier top rates also pass the tests, and would ease both of the package's thin margins (chapter 7), but by then wages are less than a quarter of output, so they mainly redistribute.

The package's hardest income tax question is not a rate. It is the GDI's withdrawal on a working day. The GDI is withdrawn at a dollar for each dollar of the day's wage. So on a day's work that pays less than the day's GDI, $188 in 2040 in that year's dollars, a worker ends the day with exactly what they would have had without working. Only pay above that point adds to their income, and that is taxed. For low-paid work, and for a few hours' work in a day, there is no reward for working at all. A withdrawal of 90 cents in the dollar, which leaves a worker ten cents, also passes all four tests, at the cost of a little more debt (chapter 7), and Korinek and Lockwood argue against tying a transfer to work at all (chapter 12). The model has no response of labour supply to this (chapter 17), so it cannot say how many people would choose not to work, and in a world where comparable work is scarce it may matter less than it would today. But JobSeeker today withdraws 50 to 60 cents in the dollar for exactly this reason, and this is the first thing a more refined policy package would revisit.

13.4 Company tax at 35%, and imputation

Company tax is 35% from 2030. It is the main tax that reaches foreign owners of Australian business before their profits leave the country, and profits are a much larger share of the AGI economy, so it raises about 15% of output.

The case against a higher rate is relocation. Australian shareholders pay their own personal rate through dividend imputation, so a higher company rate mostly falls on foreign owners and on profits kept in the business. But a rate well above other countries' would give firms reason to move profits, or themselves, offshore. The model cannot represent that, so the package stays at a rate that is high but not an outlier. For comparison, a 40% rate from 2040 ends with debt 9 points of output lower. A 30% rate throughout only just passes the fiscal test, with debt's worst five-year average at 71.0% of output against a limit of 71.2%.

Korinek and Lockwood's theory prefers a tax on consumption to a tax on capital while people are still the consumers, but it makes an exception that applies here: when the consumption tax cannot be raised as far as it should be, a tax on the normal return to capital becomes worthwhile (chapter 12). The levy is limited, by what can be collected (below) and by the brake it puts on automation. Two further reasons fall outside their model. Company tax is the only tax that reaches the share of Australian profit that foreigners own, 29% of it in 2030 and 19% in 2055. And from 2038 the economy has idle capacity, so the saving a capital tax discourages is not what limits output.

D18 keeps dividend imputation and, as today, refunds unused franking credits in cash. The refunds matter most to members of the fund and to households whose tax rate is below the company rate.

13.5 Capital gains under deflation

The model has no capital gains tax base of its own. National accounts do not count capital gains as income, and the fund's growth is recorded as growth in its assets. What the model carries is today's proportion: capital gains are about 14% of personal income other than wages, taxed at marginal rates with the 50% discount.

A tax on nominal gains taxes too much when prices rise, because part of the gain is only inflation, and today's 50% discount stands in for indexing gains to inflation. The 2026–27 Budget proposes to return to indexation, as before 1999, from 1 July 2027.¹ In D18 the two measures of prices part company for a decade. Prices at the shops rise by 2% to 7% a year from 2031 to 2035, as the levy is phased in, and are then nearly flat until 2040; prices before tax fall from 2033 to 2038, by 4.0% a year at their fastest. An asset whose price held steady while producer prices fell would have gained in real terms by that measure, and a tax indexed to shop prices, which include the levy, would treat much of a real gain as inflation. How gains are measured would need a fresh look. Capital gains tax bears on the reward for taking risks, which a package that relies on private investment needs to keep.

13.6 Collection

A consumption tax at 53% of pre-tax prices creates strong reasons to evade it: cash sales, undeclared work and goods bought online from abroad. And the levy's own base can be gamed: its rate follows the fall in an industry's wages as a share of its revenue, which a firm can disguise by calling contractors employees or by shifting revenue offshore. Nothing in the model lets firms avoid it, though D18 already assumes that a fifth of the levy on imports goes uncollected.

The package tolerates some leakage. It passes if a twentieth or a tenth of the levy goes uncollected, if only half the levy on imports is collected, rather than D18's four fifths, and if shoppers switch harder away from heavily taxed goods. With a quarter uncollected it fails the fiscal test, its worst five-year debt at 76.7% of output against the limit of 71.2%, and with half the external test fails as well. With a tenth uncollected it also fails if the world shares only two thirds of the AGI economy. Korinek and Lockwood argue that what goes uncollected would grow with the rate, and with the time people have outside paid work; the model does not show that (chapter 12).

That is not a wide margin for a tax of a kind nobody has collected before. A levy of this size would need collection at the point of sale and at the border that is harder to avoid than today's GST, and the model can only test how much leakage the package tolerates, not how much would occur. A government that expected to lose a quarter of it would need to find the equivalent elsewhere before it began.

13.7 Land

The package taxes neither housing nor land. Housing is outside the levy, and the model's broad land tax is off. Korinek and Lockwood's theory points the other way: as other tax bases shrink, rent on land becomes more valuable to tax, and with a wedge of more than half the price on everything else, housing becomes by far the largest untaxed purchase. A broad land tax would ease the budget a great deal, but in the model it falls hardest on owners with little income, and fails the inequality test (chapter 12). A land tax designed for owners who are rich in property but poor in income is a question for the next draft.

13.8 Leave provisions

The model does not treat paid leave separately. Its wages are the national accounts' wages and salaries, which include paid leave, so employers' labour costs still carry it. Roughly, paid leave accounts for about a sixth of a permanent employee's wages, and casual workers receive a 25% loading instead.

In an AGI economy with the GDI, what happens to leave is a question of behaviour the model cannot answer. If employers stop paying for leave, wages or prices may fall. Or employers may end up paying what are in effect casual rates, which already include a loading for leave; then prices would not fall, and workers would in effect be paid for leave twice, once in the loading and again through the GDI for the days they do not work.

The AGI economy pays most of its income as profit. Who owns the businesses that earn it decides where that income goes: to Australian households through the fund and their own shares, or abroad. This chapter looks at the fund, at who owns Australian shares, and at the package's choice to buy Australian business back.

14.1 Members' money, not an endowment

The national fund starts as the superannuation pool, about $6,000 billion in 2030, in 2026 dollars. It would be tempting to treat that as a national endowment: pay out very little, let it grow, and use its earnings for the country. The package does not, because the pool belongs to its members. It was built from their contributions and their employers' contributions on their behalf, and a low mandated payout would rightly be called expropriation.

So each member's balance carries over at its value, and each member draws a fixed share of it every year, untaxed: 5% at 67 and over, 2.3% below. The pool pays out about 3% a year in all. People 67 and over draw about what today's rules require at 65 to 74, and younger members get an income from their balance that today's rules would not give them until retirement.

The fund's income is taxed each year at members' own marginal rates, between 41% and 44% of it from 2040, rather than at super's concessional 15%. Its growth, the retained earnings of the companies it owns and the growth of its overseas assets, is never taxed: not as it accrues, not when the fund sells an asset, and not when it is paid out. Against super today that is heavier on income and lighter on gains.¹ Korinek and Lockwood point out that today's exemption of retirement saving makes the income tax behave partly like a tax on consumption, which their theory favours; taxing the fund's income at members' own rates moves the other way, and is one of the package's capital taxes (chapters 12 and 13). In practice the fund hardly ever sells, because its income after tax more than covers what it pays out, so even a tax on realised gains would collect almost nothing from it. Taxing its growth as it accrued, at half members' rates, would raise between 1.3% and 1.8% of output a year from 2040. The fund reinvests all its earnings after tax, as super funds do. That is what lets it grow from $6,267 billion at the end of 2030 to $25,167 billion at the end of 2055, in the dollars of each year: from 190% of output to 216%. The package depends on that reinvestment: a fund that reinvested none of its earnings, or paid out 5% a year to every member, would fail the external and fiscal tests, and one that reinvested only half would fail the inequality test by a whisker (chapter 7).

When a member dies, the balance passes to their beneficiaries as a balance in the fund, so the fund is handed on between generations. That is a departure from super today, whose death benefits are paid out in cash, and the package relies on it, though it could bear the alternative. A pool paying out 4.41% a year, well over a point above D18's 3.01%, still passes all four tests. If death benefits left the fund as cash, they would add something like half a point to a point to what the pool pays out each year,² which is inside that margin. The cost is a smaller fund and more of Australian business foreign-owned: at 4.41% the fund ends at 165% of output rather than 216%, and foreigners own 26% of business profit rather than 19%.

Self-managed funds hold about a quarter of super today, and some hold a proportion of their assets in a family farm or business premises rather than term deposits, bonds or listed securities.³ Such assets cannot simply be pooled. They could stay as ring-fenced accounts inside the national fund, holding their own assets under its drawdown and tax rules, or be sold over a transition period. The model treats the pool as one portfolio and leaves the question open.

14.2 Should the fund take new savings?

As designed, the fund is closed: after 2030 nothing more is paid in. People who were not super members in 2030, or had little in it, such as the young and migrants who arrive later, own little of it except by inheritance. Opening it to deposits would let them share in it. Whether that is worth doing depends first on whether the fund is a better place for a household's savings than the alternatives.

In D18 a dollar in the fund earns about 9% to 10% a year in real terms after tax in the 2040s: cash income taxed at the member's own rate, plus growth that is never taxed. The model's only alternative for a household's savings is a deposit, which earns −1.5% in real terms in 2040, when interest rates are at their lowest, and about nothing from the mid-2040s, when the rate is level with inflation. That comparison overstates the fund's advantage, because in reality a household could buy the same shares directly. Against that, the fund has three real advantages. Its growth is never taxed, while a household holding shares directly pays tax on its gains when it sells. A pool of this size costs far less to run than a personal portfolio or a retail fund. And it holds a diversified, partly overseas portfolio that no small saver could build. Its disadvantages are that money in it comes out only at the fixed rate, members cannot choose how it is invested, and a pool this large is exposed to future governments changing its rules.

So there is a financial case for opening the fund, with limits: without a cap on deposits, like today's caps on super contributions, most of the untaxed growth would go to those with the most to deposit. The national case is smaller than it might seem. In the model, households outside the fund always borrow for housing, as households outside super do today, about 1.6% of output a year. In the build-out, when deferred pay is being collected, they save instead; around 2040, when interest rates are lowest, they borrow more than usual; and by 2055 they have built up deposits worth 11.2% of output beyond that pattern (chapter 17). Paying 1% of output a year into the fund from 2031 uses those deposits up. It lowers debt in 2055 from 37% to 33% of output and the foreign share of profit from 18.6% to 15.5%, largely by turning households' deposits into the fund's assets. The model has no rule for who would pay in, so it cannot show the effect that matters most here, which is who comes to own a share of the fund. A starting balance for each new adult, paid from the budget, would do more for that than deposits, which help only those who can save. Both are design questions for the next version of the package (chapter 18).

14.3 Who owns Australian shares

Australian shares are owned by three groups, and the fund's buying moves them over time:

The market as a whole is worth barely more in dollars in 2035 than in 2030, and foreign holdings fall by more than a tenth. That is partly the model's measure, not a loss: shares are carried at book value, the capital behind them at the price of new capital, and AGI makes capital 7.0% cheaper in dollars in those years, almost as fast as retained earnings add to it (chapter 5). The shares of the market are what matter here, and they move steadily towards the fund. Chapter 5 sets out Australia's whole position with the rest of the world, of which foreign-held Australian shares are one part.

14.4 The buy-back and the foreign share of profit

The fund buys Australian shares back from foreign holders, year after year. As a result the foreign share of Australian business profit falls from 29% in 2030 to 19% by 2055. With no AGI it would stay at 31.5%.

The buy-back is a choice, and it has a cost. Without it D18 still passes all four tests, and by more: Australia ends a net creditor of 90% of output rather than 61%, and government debt ends at 33% of output rather than 37%. But foreigners then own 30.6% of Australian business profit in 2055, much as they do today. The buy-back trades foreign assets the fund could have held, and some of the budget's margin, for a larger stake in Australian business.

The case for it is that the AGI economy pays its income as profit. A country whose businesses are largely foreign-owned sends a large share of its AGI dividend abroad every year, taxed only by company tax on the way out. Owning more of its own businesses keeps that income, and the decisions that go with it, at home. The model cannot value control, so the package's choice here rests on a judgement the tests do not score.

14.5 The fund's equity mandate

D18's fund holds 55% of its assets in Australian shares and 40% overseas. Moving the balance changes the trade between external margin and ownership:

Keeping more of the fund at home lowers the foreign share of profit but leaves Australia with a weaker external position and more government debt. Sending half overseas raises the foreign share of profit to 28%, and on this model changes the external position and the debt hardly at all: more of Australian business stays foreign-owned, and so does more of its profit. Sending most of it overseas fails badly: the fund would have to sell Australian shares to foreigners to buy the foreign assets, and by 2055 foreigners would own well over half of Australian business profit.

55% is a middle setting. It gives as strong an external position and as low a debt as any of these settings, at the cost of a foreign share of 19% rather than 12% or 13%.

D18 collects 65% of output in revenue by 2055 and pays out 40% of output as the GDI. By one measure that is a state much larger than today's. By another it is a much smaller one.

15.1 A revenue take of 65% of output

Revenue rises from 42% of output in 2030 to 55% in 2035, 61% by 2037 and 65% by 2055. No market economy has run a tax take anywhere near that, and it is one of the package's largest reservations.

Most of it is paid straight back out. The GDI alone is 40% of output by 2055. What remains after the GDI, about 25% of output, pays for government services, interest and the public investment programme, and from 2052 leaves a surplus. The state in D18 is mainly a transfer system: it collects the income that AGI generates as profit and consumption and returns it to people as an income, rather than using it to buy things itself.

15.2 Services per person roughly flat

What the state buys for its citizens grows very little. Real government services per person rise just 5% between 2030 and 2055, while output per person more than doubles. As a share of output, services fall from 23% to 9%. In a no-AGI Australia, services per person rise 31%, in step with its much slower growth.

That is a setting, not a result. The model lets government services per person grow in step with output per person, not at all, or anywhere in between, and D18 sets them near the bottom. Moving the setting shows how much room there is:

There is some room, though not a great deal, because D18 spends so much of its budget on the GDI. Services per person could grow by 33% by 2055, about as fast as they do in a no-AGI Australia, and D18 would still pass all four tests, with debt ending at 51% of output rather than 37%, not far inside the limit of 54.8%. At 46% the fiscal test fails, on debt in 2055. At 76% the inequality test fails as well. What cannot be afforded at all is keeping the state's share of output. A state that grows its services with output spends much of the AGI dividend on its own purchases, with the same taxes and the same GDI, and its budget and the external account both fail badly.

So the smaller state is D18's choice, not a necessity. The package spends the room on a larger GDI, a lower debt and a surplus at the end of the period; a government could spend it on services instead, and match what a no-AGI Australia would provide. That choice belongs to politics rather than to the model.

The argument over whether government should take 22% or 28% of GDP looks less important when output per person more than doubles in 25 years. A state whose services per person hold steady still provides what it does today, for a growing population, in an economy where AGI should make many of those services cheaper to provide. One whose services grow as a no-AGI Australia's would provides nearly a third more. The model's result is that the state can pay everyone an income and grow its services as fast as it would without AGI, but not keep its share of an economy that AGI has made so much larger.

15.3 A transfer state

A GDI of 40% of output would be the largest income support system any market economy has run. No country has tested what that does to how people work, save, vote and see each other, and the model cannot tell us. It shows the money can be made to add up. Whether a society can be built around it, with most adults outside paid work, is a larger question than this commentary can answer (chapter 17).

From 2038 D18's economy uses between 81.7% and 88.5% of its capacity, 84.8% on average. About a sixth of what AGI could produce is never made, because nobody buys it. That capacity is mostly machines and AI, not people, and in the model's world it costs very little to use. This chapter asks what it would take to use it, and what kind of world that would need.

The short answer is that the rest of the world would have to buy more of what Australia makes. The budget cannot fill the capacity on its own within the limits the package sets, and cutting taxes to let households spend more fills little of it and costs jobs. With the world buying more, and a larger public investment programme aiming at full use, the economy runs at 93.5% to 98% of capacity from 2038 and the package passes all four tests with more room than D18. That result is economically feasible in the model. Whether it is feasible in practice depends on assumptions the model can state but cannot test.

16.1 Why the capacity sits idle

Output in the model is limited by spending as well as by capacity (chapter 3). From the late 2030s spending falls short, for three reasons.

• Households. They spend a share of their cash income that does not rise as they grow richer, and the fund reinvests all of its own income (chapter 17).
• The state. Government services per person barely grow, so the state buys a shrinking share of output (chapter 15).
• The world. D18 assumes the rest of the world gains three quarters as much from AGI as Australia does. Its demand for Australian exports grows more slowly than Australia's capacity: ordinary exports fall from 23% of Australia's capacity in 2030 to under 20% by 2055.

The public investment programme fills part of the gap. It aims to keep the economy at 95% of capacity, but it cannot spend enough to do so. It first spends in 2039, and until 2047 it grows as fast as its deficit limit allows, borrowing about 3% of output a year. From 2047 it is held at its cap of $1,000 billion a year, at 2026 prices for capital goods, while the economy keeps growing, and the share of capacity in use falls, to 83% by 2050.

16.2 Three ways to use it

More spending on Australian output can come from three places: the budget, households, or the rest of the world. The model can try each.

The budget. A bigger programme aiming at 99% of capacity, with its cap raised to $1,500 billion, lifts average use only from 84.8% to 87.2%, and government debt ends at 50% of output rather than 37%. There are two reasons it does so little. Until the mid-2040s the programme is already borrowing up to its deficit limit, so a higher target or cap cannot make it bigger then. And an AGI economy's public investment adds to capacity almost as fast as it uses it: every project adds to what the economy can produce, and about a sixth of the export capacity the programme builds finds no buyer at the prices the model allows (chapter 4). Loosening the deficit limit fills a little more, at a cost the fiscal test does not allow. Loosened only from 2038, after the build-out, to 4% of output, with the cap at $2,000 billion, the economy runs at 91.1% on average, but debt ends at 62% of output, above the limit of 54.8%. Loosened for the whole period, the programme spends during the build-out, when the economy is already near full, much of the extra spending goes on imports, and the package fails the fiscal test badly, with debt ending at 81% of output.

Households. Cutting the levy's maximum rate from 120% to 100% from 2045 lets households buy more, but it lifts average use only to 85.3%, and debt ends at 42%. It also costs jobs. The levy is what makes automation expensive for a firm, so a lower levy means more automation: employment in 2055 falls from 7.0 million to 6.2 million. Raising the GDI faster still, at 150% of the growth in output per person rather than 125%, does no more (85.5% on average). Households spend part of any extra income on imports, and in an economy with this much spare capacity the rest does not go far.

The world. If the rest of the world gains as much from AGI as Australia does, its demand for Australian exports keeps pace with Australia's capacity: ordinary exports stay at 23% of it rather than falling below 20%. With nothing else changed, the economy then runs at 91.3% to 95% of capacity from 2038, and government debt falls to 11% of output by 2055. The model sets the world's demand for exports and its companies' profits separately (chapter 7). Raising export demand alone to 100 uses almost as much of the capacity, 91.8% on average; raising world profits alone uses little more than D18, 86.2%. So the capacity is filled by what the world buys. The fund's higher returns help the budget and the external account, not the use of capacity.

16.3 The full-capacity variant

Combining the world with a bigger programme uses nearly all of the capacity. The full-capacity variant is D18 with three changes:

• the rest of the world sharing all of Australia's gain from AGI rather than three quarters, in both its demand for exports and its companies' profits;
• the public investment programme aiming at 99% of capacity rather than 95%;
• the programme's cap raised from $1,000 billion to $1,500 billion a year, at 2026 prices for capital goods.

Everything else is D18, including the programme's 3% deficit limit, the $150 billion a year by which it can grow, and the limits on its returns and on the prices its exports fetch. The variant is the published model's "Capacity-saturating programme" preset.

It runs at 93.5% to 98% of capacity from 2038, 95.4% on average, and at about 96% in 2045 and 2055. The programme reaches its $1,500 billion cap by 2050. All four tests pass, with more room than D18: net foreign liabilities peak at 30% of output and end at −77%, and government debt falls to 22.4% of output by 2055, against D18's 36.9%.

16.4 What changes when the capacity is used

Output, not jobs. By 2055 the economy produces 23% more than D18's. Employment rises by only 0.5 million, to 7.5 million. In an AGI economy the idle capacity is capital, not people, and using it employs machines. The familiar case for public works in a slump is that they employ the unemployed. Here the benefit reaches people only through what the works produce and earn.

Households. They are about a fifth better off than in D18 by 2055: the 10th percentile, the median and the 90th percentile end 2.11, 2.29 and 2.35 times as well off as under today's system, against 1.73, 1.84 and 1.89. Most of the gain comes through the GDI, which rises with output per person and by 2055 buys what $314 a day buys today rather than $237. The build-out years change little.

Interest rates and house prices. With the economy fuller, the central bank's rule sees less spare capacity. The interest rate falls only to 2.6% in 2040, where D18's falls to 1.0%, and is between 2.4% and 3.5% from 2045, up to a point above inflation. The variant pays more on its debt than D18 does, and its debt still falls faster. Higher rates also hold house prices down: against prices at the shops they end 1.89 times their 2030 level rather than 2.20 times.

The external account. More exports turn the current account to surplus by 2045, and Australia ends a net creditor of 77% of a year's output. The foreign share of Australian business profit falls to 16.7% rather than 18.6%.

16.5 What has to be believed

The world shares all of the AGI economy. This is the assumption that does the work, and it is stronger than D18's. D18 assumes the rest of the world gains three quarters as much from AGI as Australia does. The full-capacity variant assumes it gains as much, and keeps buying the same share of what Australia can make as it does today. With the same programme settings and the world's two settings lower, the economy uses less:

The bigger programme is a worse bet than D18's own if the world turns out cautious: at two thirds D18 passes all four tests (chapter 7), and with the bigger programme it fails the fiscal test.

There are reasons to think the world could do as well as Australia. The incentives to adopt AGI are much the same everywhere, and Australia's main trading partners may adopt it faster. There are also reasons to doubt that the world would want more of what Australia sells. AGI makes manufactured goods cheap everywhere, so what Australia can sell is what AGI cannot reproduce: coastline, space and climate, education in English, a stable and safe society, and minerals and energy processed with abundant sun and wind (chapter 4). Whether demand for those grows as fast as Australia's capacity is the question the variant answers by assumption.

A programme larger than any country has sustained. Public investment reaches 17.5% of output at cost by 2050, against 4% in 2030, and gross investment 43% of output. For comparison, China's gross fixed capital formation was about 40% of GDP in 2023, and Australia's mining investment peaked at just under 8% of GDP in 2012.¹ By 2055 the public capital the programme has built is worth 1.49 times a year's output, on top of the public capital Australia already has. About a fifth of the export capacity it builds goes unsold. The model has no limit below the economy's total capacity: in reality, finding good projects, planning approvals, materials, energy and specialised skills would bind well before the last unit of idle capacity was used.

The programme's returns and markets hold up. The variant keeps D18's limits: the return on public capital halves once the productive stock reaches a year's output, and the programme's export prices halve once its exports reach 12% of output. With the world buying more of Australia's ordinary exports, those limits are enough.

It is sturdier only because the world is kinder. Of the 118 single changes in chapter 7, 12 break the variant, against 21 for D18, and every one of the 12 also breaks D18. Among the nine D18 breaks it survives are firms passing on only 60% of their cost savings, a risk premium on government debt, a higher dollar, every member drawing 5% of their balance, and a quarter of the levy uncollected. But the stress table holds the world at 100. The variant's extra margin comes from the world assumption, not from anything the package does.

16.6 What to make of it

D18's idle capacity is mainly a statement about the world. The package assumes a rest of the world somewhat less transformed than Australia, whose demand for Australian exports grows with only three quarters of Australia's gain from AGI. A government cannot spend its way to full capacity within prudent limits, because in an AGI economy public investment adds to capacity almost as fast as it fills it, and cutting the levy to let households spend more costs jobs.

The late decline is also worth noticing. From the late 2040s the deficits of both D18 and the variant narrow: the variant is in surplus by 2050 and D18 from 2052, while D18's capacity use falls. With the programme at its cap, nothing in the package turns the surplus into spending on Australian output until the surplus rule starts lifting the GDI, in 2053; by 2055 it has raised the GDI by 6.6%. A government in that position would have choices the package does not make: a larger programme later, services that grow faster (chapter 15), or lower taxes.

If the world does as well from AGI as Australia, most of the capacity can be used, households are about a fifth better off, and the budget and the external account both end stronger. That is a reason to watch the rest of the world's adoption of AGI as closely as Australia's own, and to build the export industries AGI cannot make cheap. It is not a reason to plan on it. The full-capacity variant is a possibility the package could grow into, not the base case.

Every model leaves things out. This chapter lists what this one leaves out that a reader should weigh, and the places where its simplifications could change the story.

17.1 Behaviour

• No labour-supply response. Nobody in the model decides to work more or less because of the GDI or the tax scale. The GDI is withdrawn at a dollar for each dollar of a day's wage, so a day's work that pays less than the day's GDI leaves a person no better off (chapter 13). In a world where comparable work is scarce this may matter less than it would today, but the model cannot say.
• No new work from rising incomes. The human share of each industry's work is set by how far AGI has reached into it, so employment grows only with the population (chapter 5). Richer households may want more of what people do best, and some of the displaced would find that work. Chapter 2 shows the package still passes if many of them do.
• No relocation. Firms do not move profits or themselves offshore in response to company tax or the levy. That is why the package keeps company tax at 35% rather than testing higher rates the model would reward (chapter 13).
• No avoidance of the levy. Firms pay the levy in full, although a tax of up to 120% of the price before tax would be avoided, by cash sales, by buying abroad, or by gaming the wage ratio its rate is based on. The package passes with a tenth of the levy uncollected, and fails the fiscal test with a quarter uncollected (chapter 13). Korinek and Lockwood argue that what goes uncollected would grow with the rate and with the time people have outside paid work (chapter 12).
• No migration response. Population follows its projected path. A country paying every adult a GDI would face questions about who is eligible that the model does not address.
• Saving does not rise with prosperity. Households spend a share of their cash income that depends on where they stand relative to the median, not on how well off everyone is, so they save no more as their purchasing power nearly doubles. Two things make them save less in D18: the GDI makes a precautionary buffer less necessary, and interest rates below 5% let borrowers spend more, although that second effect fades as the cut is bid into house prices (chapter 3). The share of cash income spent rises from 86% in 2030 to 89% in 2040, and is 88% in 2055. The fund's income is all reinvested and its payout is assumed to be spent in full. If households saved more, the package would do worse, not better: the economy has spare capacity from the late 2030s, so more saving means less spending, less output and less levy. Without the first effect the package still passes, with debt's worst five-year average at 66.5% of output against a limit of 71.2%; without both it passes, at 69.2%; without the second alone it passes with more room.
• No response of the leave system. What happens to paid leave when the GDI pays people for days they do not work is left open (chapter 13).

17.2 The unemployment payment

• Everyone displaced is paid, without conditions. The model pays today's unemployment payment to every worker displaced after 2030, for as long as the phase-out runs. It applies the test on a partner's earnings and nothing else: there is no waiting period, no assets test, no requirement to look for work and nobody who fails to claim. Today's system has all of these. The budget's figure therefore could be more than today's rules would cost, and the households' figure more than they would receive.
• JobSeeker and Rent Assistance only. Parenting Payment and the rise in Family Tax Benefit when a family's income falls are left out. Both would pay displaced parents more, and cost the budget more (chapter 4). On the other hand, every child is paid a child GDI.
• A wage well above the payment is not replaced. The payment is today's safety net, not income insurance. Households that lose a high wage are still well behind today's system in the build-out (chapter 5). Countries with earnings-related unemployment insurance would do more for them; the package does not.
• The household layer has fewer workers than the economy. The model's household types contain fewer earners than the economy has workers, so from 2033 the household layer pays displaced workers 74% to 84% of what the budget pays in total, though the same amount for each. The fiscal test uses the budget's figure. The counts of who is worse off, and the Gini coefficient, use the smaller one.

17.3 Prices and interest rates

• Prices follow costs, never demand. Firms set prices from their costs, and when spending exceeds what the economy can produce the excess is imported. That works for goods that can be imported. It does not work for meals out, haircuts or building work, whose prices, and the pay of the people who provide them, would rise if demand outran supply. In the build-out years that could mean higher inflation than the model shows, and households squeezed by prices rather than by imports: better for the current account, worse for purchasing power in 2031–35.
• No second round from the levy. The levy lifts prices at the shops by 16.8% between 2031 and 2034, and the central bank "looks through" it. In the model it can, because wages follow productivity and the 2.5% trend, today's payments follow wages, and nothing is indexed to shop prices. If workers and pensioners won compensation for the levy's price rises, inflation would be higher in the build-out, a real central bank would probably hold rates higher still, and the budget would pay more on its debt in the years it is highest. The Reserve Bank made the same assumption about the GST in 2000, for a far smaller and single step in prices (chapter 3).
• The central bank answers capacity only. Its rule no longer cuts for AGI's falling costs, which was the weakness of earlier versions. It still does not look at the current account, and in the build-out the economy is importing what it cannot make. A rule that did would hold rates a little higher again.
• The rate's floor follows a lagged measure. The rate is never below underlying inflation as measured over the two years before. From 2040 to 2045, as prices return to their trend, the measure lags, and the rate is below that year's shop inflation by 0.2 to 0.6 points. From 2050 it is level with inflation. That is no longer a rate below inflation, but a real rate of about zero in the 2040s and 2050s is still low: the model's own neutral rate is 2.5% in real terms. With the rate never below 3% or 4% the package still passes, unless the world is also at two thirds (chapter 7).
• Ordinary inflation is assumed, not modelled. The model assumes the Reserve Bank meets its 2.5% target on average. With about a sixth of capacity idle from the late 2030s, and prices before tax falling in the early and mid 2030s, that could be hard to do in practice (chapter 2). At 1.5% the package passes by 0.6 of a point, and at 1% it fails (chapter 7).

• No domestic business debt, and no business failures. Corporations in the model owe the private part of Australia's foreign debt and pay its interest from their profits, and small-business debt is in the household sector's interest bill. What the model lacks is debt that firms owe to Australian lenders. For the economy as a whole that mostly cancels: banks and firms are one sector, and interest paid to Australian households is a transfer between Australians. Two things it hides. If firms borrowed more heavily to pay for the build-out, their interest deductions would grow faster than the model allows and company tax would be lower. And in the build-out, prices before tax fall while interest rates stay near 5%, so a firm with debts and falling prices is squeezed; firms undercut by automated competitors could fail, and their lenders with them. The model has no defaults and no banking losses.

17.4 Deferred pay and the no-loss rule

• Deferred pay earns no real interest. Balances are indexed to prices, so they keep their value but grow no further. The money deferred is household saving held in the savers' own names. In the model it sits with households' other savings, as deposits that lend to the government, so it neither goes abroad nor lies idle. The household sector as a whole earns interest on it, or borrows that much less, but the model does not credit that interest to the individual accounts, which are only indexed.
• Repayment by household type. The model returns deferred pay to each household type in proportion to what that type paid, not to each person.
• The no-loss rule protects the average recipient, in dollars. Today's transfers are withdrawn no faster than the GDI rises for the average recipient. That does not protect every recipient, or protect anyone against rising prices. Households that live mostly on the pension barely gain for nine years from 2038, and working-age households already on today's payments lose a tenth of their purchasing power between 2038 and 2047 (chapter 5). Single pensioners, who receive more than the average per person today, have most to lose. A supplement for single adults would address part of this. The package does not include one.
• Years of loss in the build-out. The package reports, without scoring it, whether any of the bottom three fifths loses more than 0.5% of its purchasing power in a year. In D18 the second and middle fifths do in 2033, by 1.7% and 1.4%, and the poorest fifth, just, in 2031, by 0.5%. A faster start to the GDI would avoid it, at the cost of the fiscal margin (chapter 4). The bottom three fifths together are never more than 0.5% below their 2030 level.

17.5 The fund

• Death benefits stay in the fund. A member's balance passes to their beneficiaries inside the fund. If death benefits were paid out in cash, as super's are today, the fund would pay out more each year. A pool payout of 4.41%, well over a point above D18's, still passes all four tests, so the package could bear it, at the cost of a smaller fund (chapter 14).
• Self-managed funds are not treated separately. The pool is one portfolio; how self-managed funds holding farms or business premises would move into it is left open (chapter 14).
• The fund is closed to new savings. Nothing is paid in after 2030. Opening it is a design question for the next version (chapter 14).
• Taxing the fund's income, not its growth. The fund's income is taxed each year at members' rates. Its growth is never taxed, whether it accrues, is realised by a sale or is paid out. That treats members more generously on growth than households holding shares directly, who pay tax on gains when they sell, and than super funds in accumulation, which pay about 10% on realised gains (chapter 14). Taxing the growth as it accrued, at half members' rates, would raise 1.3% to 1.8% of output a year from 2040.

17.6 Stocks and prices

• Shares at book value. Australian shares are carried at book value, the capital behind them at the price of new capital, growing with retained earnings; the fund carries its holdings at cost. Book values fall in the build-out years because AGI makes capital cheaper faster than retained earnings add to it (chapter 5). That is why, in dollars, the fund is worth less at the end of 2034 than at the start of 2030, although its income has grown. For the external test this is the cautious choice. At market values, which follow profits, Australians' holdings abroad, which are the larger, would rise by more than foreigners' holdings here, and net foreign liabilities would be lower.
• A blunt instrument for the financing limit. When the government's limit on foreign borrowing binds, the model slows AI adoption, and that often leaves the external position worse rather than better, because an economy held back imports what it no longer makes. A government choosing the limit would more naturally pause the public investment programme, or draw on the fund's overseas assets; the model has the first as an option, off in D18. The limit never binds in D18, and breaches are reported beside the external test rather than scored in it (chapter 3).
• Housing. House prices follow rents and interest rates, and rents drift up slowly; higher house prices do not feed back into rents.

17.7 The world

• Two settings for the world. How much of the AGI economy the rest of the world reaches is set twice: once for Australia's export demand, once for the profits of the world's companies, and so the fund's overseas returns (chapter 7). Both are assumptions, not results. Nothing in the model links them to each other, or to what Australia does.
• No crisis. Beyond the financing limit, the model has no sudden stop in foreign lending, no currency crisis and no sovereign risk premium unless one is set with the "Risk premium on government debt" setting (chapter 7).
• Fixed overseas returns. The fund's overseas assets earn 3% income and 3% growth a year whatever happens in the world economy, except in the market-shock stress cases.
• Who makes AI. The model knows only what AI costs and how much of the spending on it stays in Australia. It does not represent any market power that may exist among AI providers, which is what would make AI expensive (chapter 7).

17.8 Politics

The model says nothing about whether any of this could be legislated, or survive an election. It does not model the social effects of a society in which most adults are not in paid work: loss of purpose, the status that attaches to work, or the politics of a state that pays everyone. Those may matter more than anything the model can measure, or they may not.

17.9 Where the model's own accounts are weaker

The model's accounting checks run at the base case, at every preset and at a dozen other settings. Three settings D18 does not use are weaker than the rest.

• Everything sold. Run with the published studies' assumption that everything the economy can produce is sold, households buy more than they can pay for, and their accounts do not close. Figures from that comparison (chapter 7) are indicative.
• A broad land tax. The budget charges households a share of a broad land tax set by who pays today's state land taxes, while the household layer charges each household the tax on its own land. D18 has no land tax, but a package that used one, as Korinek and Lockwood's argument suggests (chapter 12), would need the two made consistent.
• Far corners. At settings far from D18, such as a land tax of 2.6% or a GDI of $25 a day, the model occasionally fails to converge or leaves part of the fund unplaced. None of the stress changes in chapter 7 does.

17.10 For a later version

• Using the idle capacity late in the period. From 2051 the budget runs into surplus while the share of capacity in use falls; the package has no rule that turns that surplus into spending on Australian output (chapter 16).
• Inequality in the 2040s. The Gini rises from 2035 to a peak in 2047, as the transition's payments are withdrawn, and falls back only when the budget's surplus lifts the GDI, from 2052 (chapter 5). A withdrawal of today's payments that is slower but paid for elsewhere, a GDI tied directly to income from capital, or a tax on the fund's growth (chapter 14), would hold it down more surely.
• A reward for working. A GDI withdrawn at less than a dollar in the dollar, and what that would cost (chapter 13).
• Prices and wages that respond. Prices that respond to demand for what cannot be imported, wages that respond to prices, and a central bank that watches the current account as well as capacity.
• Opening the fund to capped deposits, or to a starting balance for each new adult (chapter 14).

18.1 What the exercise shows

If AGI arrives on the scale its builders describe, and displaced workers do not find comparable work, today's tax and transfer system does not survive it. Within ten years nearly half the jobs have gone, and with them much of the wage base that income tax rests on. Today's safety net would keep most households afloat, but only by borrowing: the budget runs deficits above a tenth of output from 2036, and debt passes 100% of output in 2038 and 150% by 2045. Without the safety net, the poorest households would lose more than half of their purchasing power and inequality would double. Either would happen while the economy produces more than ever (chapter 2).

It does not have to be that way. D18 shows that one internally consistent set of policies can carry Australia through: an income for everyone, today's unemployment payment for those who lose their jobs before that income is large enough to live on, a consumption tax that follows automation, a national fund built from superannuation that owns more of Australian business and pays its members an income from their balances, restraint by the highest earners in the build-out years, and public investment timed to use capacity that would otherwise sit idle. Under the model's assumptions it passes four demanding tests at once. It leaves each fifth of people between 1.65 and 1.87 times as well off as the same fifth under today's system, protects most people 67 and over from the start, leaves nobody in the fifth who are poorest under today's system more than a tenth worse off than under it in any year, holds inequality below today's system throughout, and ends with Australia a net creditor and government debt at 37% of output. It does so with a central bank that holds interest rates above 4% until 2037 and never lets them fall below underlying inflation (chapters 3 and 4). And because its income and its taxes follow the jobs AGI actually takes over, it passes at the pace of any earlier technology, from the smartphone's to electricity's, and does little if AGI does little (chapter 7).

18.2 What it does not show

It does not show that D18 is the right package, or a likely one. It is one response to one scenario, and others pass too (chapter 7). The package relies on the rest of the world going through the same transition, on AGI adding to output as well as taking jobs, on a levy that can exceed 100% of the price before tax and has to be largely collected, and on tax and transfers larger than any market economy has run. Its smaller state is a choice rather than a necessity (chapter 15). It has two thin margins, the budget in the 2030s and inequality in the 2040s, and they pull against each other: 21 of 118 single changes to its assumptions break a test, 14 of them the fiscal test and 8 the inequality test (chapter 7). Its transition protects the poorest two fifths and is paid for by people in the upper half, and by people on today's pensions and benefits who gain more slowly (chapters 5 and 13). If AGI turns out to be only a strong productivity improvement, as the published studies and Treasury expect, today's system copes, and D18 does little: few jobs go, so its GDI phases in only a small part of the way (chapters 8 and 9). What it cannot survive is AGI that takes the jobs without adding much to output (chapter 7). And it says nothing about how people would behave, or live, in an economy where most adults do not work.

18.3 What comes next

• The two margins. D18's budget and its inequality test are both close to their limits, and most ways of easing one tighten the other. The next version should look for settings that ease both: the schedule on which today's payments are withdrawn, a tax on the fund's growth, or top income rates above the 50% cap.
• The households that stand still. Households that live mostly on the pension barely gain for nine years from 2038, and working-age households on today's payments lose ground over the same years (chapter 5). A gentler path for them, and a supplement for single adults, are the first things to test.
• A reward for working. The GDI is withdrawn at a dollar for each dollar earned on a working day. A 90-cent rate also passes (chapter 7); what a lower rate would buy needs a model of how people respond to it (chapter 12).
• Korinek and Lockwood's challenges. What goes uncollected of the levy as its rate rises and people have more time outside paid work; spending shifting towards housing, and a land tax designed for owners who are rich in property but poor in income; and purchases timed ahead of a rising levy (chapter 12).
• Using the idle capacity, and the late surpluses. Chapter 16 shows that the world, not the budget, decides how much of the capacity is used. What the package lacks is a rule for the 2050s, when the budget runs into surplus while capacity sits idle.
• The design questions left open. How capital gains are measured when prices before tax fall; how the levy would be collected; how self-managed funds move into the national fund; what conditions, if any, attach to the unemployment payment.
• Opening the fund. Capped deposits, or a starting balance for each new adult, so that people who were not super members in 2030 can own a share of it.
• Prices and wages. Prices that respond to demand for what cannot be imported, and wages that respond to prices, so that the levy's effect on inflation is a result rather than an assumption.
• The next breakthrough. The technology clock gives AGI one S-curve, slow, then fast, then slow. A second wave of capability after the first has matured is the obvious next scenario to test.

The model and its settings are published so that anyone can change an assumption or two, or take the entire model, tinker with it and see what follows. The point of this exercise is not D18. It is to spark ideas and start discussions, and to move towards having analysed, credible options lying around before they are needed.

A.1 The four tests

A package passes when it passes all four together. The limits were set on GDP at market prices. The model measures output before consumption taxes, so each limit is restated on that basis using today's gap between the two, which is fixed and does not change with the package's own consumption taxes.

The limits are judgements. The external limits allow a larger foreign debt than Australia carries today, for the few years a large build-out would need, and require it to fall back to 30% of GDP by the end. The fiscal limits allow debt near today's level for the whole government sector, and require it to be falling by the end. They are set on gross debt: everything the three levels of government owe, before netting off their financial assets. National net debt is more than 20 points of GDP lower than gross at the start (chapter 3). The purchasing-power test requires that people across the distribution end the period no worse off than today's system would leave them. People are ranked by their household's purchasing power per equivalent adult (chapter 3). The inequality test requires that the transition not make the country less equal than it is.

The external test does not require that there be no year the country could not finance: no year in which the current-account deficit exceeds one sixth of the room left below a ceiling on net foreign liabilities of 80% of output, plus the deficit growth alone would absorb, even after AI adoption has been slowed as far as policy can move it in a year. That limit is a policy the government chooses, not one the world imposes, and its instrument, slowing AI adoption, often leaves the external position worse. So such years are reported as breaches of the government's rule, beside the test rather than in it. The model records the deficit as borrowed and flags the year; what would give in reality is not modelled (chapter 3). D18 has no such year.

A.2 The reported measures

Five further measures are reported beside the tests but are not part of the pass mark:

The no-loss measure counts a fall of more than 0.5% in a year as a loss. On that measure D18's second and middle fifths each have a loss in 2033, of 1.7% and 1.4%, the year jobs start to go in large numbers while the GDI is still small, and the poorest fifth has one, just, in 2031, of 0.5% (chapter 5). The inequality trend, whether the Gini ends no higher than it was in 2032, is reported but not scored. D18's Gini ends higher than in 2032, 0.167 against 0.161 (chapter 5).

Every setting of D18 is in the published model; loading the model with no changes loads D18. The tables below list the ones this commentary refers to. Dollar settings are in 2026 dollars, as on the model's page.

B.1 Taxes, by five-year marker

Each rate and threshold is set at 2030, 2035, 2040, 2045, 2050 and 2055, and moves in a straight line between them. Thresholds are in 2026 dollars and rise with 55% of the growth in output per person, measured in dollars. The rates are the whole rates: there is no separate Medicare levy.

B.2 Income, transfers and deferred pay

B.3 The fund, the programme and the state

B.4 The world and the technology

C.1 The checks

The model is stock-flow consistent: every dollar one sector receives is paid by another, each sector's income less its spending equals what it lends or borrows, and every stock changes only by the flows into and out of it. About twenty-five automated checks are run after every change. Among them:

• the flow of funds: every sector's receipts and payments, built as a double-entry table and reconciled to the model's own lending and borrowing;
• the stocks: government debt, the fund, household deposits, external debt and each group's shareholdings, each restated as an identity and checked year by year;
• the external account: every flow that crosses the border appears in the current account, and every stock it builds appears once;
• the no-AGI baseline: every setting is stated rather than inherited from the package, and the baseline's first and last years are pinned;
• the tests themselves: defined once and checked for the price basis they use;
• the household layer against the budget: each household's capital income, fund payouts and taxes, per person, against the budget's own totals, and the rate of tax on what each band of households buys against the rate the budget collects on household spending;
• the page's controls: every default and every value a preset sets lies within its slider's range and on its slider's steps, so the page can show what the model runs.
• the technology clock: in every year, the growth AI adds to the economy is exactly the change in the clock's level, held back by the adoption brake, plus the diffusion point; and with the clock off, the fixed path it replaces behaves as before.

The accounting checks run at the base case, at every preset and at a dozen stress settings, not only at the base case. Three settings D18 does not use are known not to close, and are run so that any change in how they fail is seen (chapter 17).

Many of the checks are tested in turn by deliberately injecting the fault they are meant to catch and confirming that they catch it. The frozen model passes all of them except one check of the household transfer layer that is left open by design.

C.2 What the audits found

The model was rebuilt and audited several times during 2026. The audits found errors that changed earlier results, and every result published before them is superseded. They were of a few kinds.

• Money created or destroyed. An early version counted the public investment programme twice in the budget. Later, the measure of national income subtracted the whole GST and levy from business profit, understating profit, company tax and dividends by up to nearly a fifth of output, and the levy on imports was counted as revenue that nobody paid. Once every receipt had to have a payer, a dozen such flows were found and corrected.
• The external account. Foreigners had been charged interest on their own reinvested profits, interest had been charged twice on offshore government bonds, and the government's foreign-asset income had been left out. The fund's overseas return had been carrying a hidden extra 3% a year of growth.
• The base year. Calibrating the base year to the 2024–25 national accounts, including owner-occupiers' imputed rent and households' interest flows, showed that the economy starts at full capacity, as Australia does. That changed the story of the transition from a demand slump to an AI build-out that has to be financed.
• The household layer. The comparison with today's system was put on one basis: the same households, the same units, and super paid out as today's system pays it. A check against the budget found that the household layer taxed the fund's payout, which the budget does not, gave households no franking credits, spread capital income over a fixed population rather than a growing one, and charged households the levy only on what is made in Australia, where the budget collects it on imports as well. All were corrected, and checks now hold the household layer to the budget.
• Stocks and trade. A stock audit found households borrowing at no cost, shares traded at book value while profits tripled, and exports and imports that did not respond to the economy. Exports and imports now follow the economy and the world.
• An adversarial review. The model was attacked rather than used. Six mechanisms were corrected: the GDI and the tax thresholds follow output per person in dollars rather than in volume; the part of the price level that reflects AI takes the productivity the economy actually gets, in the same year; the extra spending low interest rates allow fades as the cut is bid into house prices; deferred pay restrains spending at its payers' own rate of spending; a risk premium does not raise the dollar; and the surplus waterfall does not read 2030's one-off surplus.
• Interest rates. A question whether the package worked only because lenders accepted 1% while inflation ran at 2.5% led to a central bank that does not cut for falling costs, a rate never below underlying inflation, and the fund's return measured on one price basis (chapter 3).
• The dollars shown. An earlier unit, 2026 dollars with ordinary inflation taken out, could not show inflation, made debts appear to shrink, and put two price bases into the national fund's table. The page and the commentary show the dollars of each year for anything that adds up, and what an amount buys at today's shop prices for anything a person receives or pays (chapter 2). New checks confirm that the budget, debt, external accounts and fund add up in every year in those dollars, and that the built page shows what the model runs.
• The figures in this commentary. Every figure quoted here comes from a pack generated once from the frozen model. The pack's generator was checked independently on an earlier package: with no sight of the code that produced the pack, the check recomputed about 5,000 of its figures from the model and compared several hundred with the built page. It found four reporting errors, among them an average wage that had been divided by a rounded count of workers, and all were corrected.

D.1 Data behind the model

ABS Australian System of National Accounts 2024–25; ABS Balance of Payments and International Investment Position (June 2026); ABS external accounts (5204043); ABS Taxation Revenue and Government Finance Statistics; ABS fiscal incidence study 2015–16; ATO Taxation Statistics; APRA Annual Superannuation Bulletin 2024–25; RBA background on the listed equity market (2021); Jobs and Skills Australia's occupation-level exposure scores for automation and augmentation; the 2026 Intergenerational Report for trend growth.

D.2 Works discussed

• Brett Hemenway Falk and Gerry Tsoukalas, The AI Layoff Trap, arXiv 2603.20617 (version 3, June 2026). arxiv.org/abs/2603.20617
• James A. Lennox and Janine M. Dixon, Occupations, Tasks and Generative AI: A Computable General Equilibrium Analysis, Centre of Policy Studies working paper G-367 (May 2026). copsmodels.com
• EY, The economic upside of Artificial Intelligence for Australia (5 August 2026). ey.com
• Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter and Peter McCrory, Economic Scenarios for Transformative AI, The Anthropic Institute working paper 2026-02 (September 2026). anthropic.com
• Anton Korinek and Lee M. Lockwood, Public Finance in the Age of AI: A Primer, Brookings Center on Regulation and Markets working paper (January 2026). brookings.edu
• William D. Nordhaus, Schumpeterian Profits in the American Economy: Theory and Measurement, NBER working paper 10433 (April 2004). nber.org
• Treasury, Intergenerational Report 2026: Australia's future to 2066. treasury.gov.au
• Jobs and Skills Australia, Our Gen AI Transition: Implications for Work and Skills (2025). jobsandskills.gov.au
• John Maynard Keynes, How to Pay for the War (Macmillan, 1940).
• Reserve Bank of Australia, "The Economy and Financial Markets", Bulletin, August 2000 (on looking through the GST's effect on prices). rba.gov.au
• JobSeeker Payment and Rent Assistance rates from 20 September 2026, taken from secondary summaries of the indexation and not yet checked against Services Australia's own tables. pay-calculator-australia.com
• Wynne Godley and Marc Lavoie, Monetary Economics: An Integrated Approach to Credit, Money, Income, Production and Wealth (Palgrave Macmillan, 2007).
• Australian Bureau of Statistics, Measuring What Matters, "Income and wealth inequality". abs.gov.au
• Australian Government, Budget 2026–27, proposed capital gains tax changes, as summarised by Chartered Accountants ANZ. charteredaccountantsanz.com
• Australian Taxation Office, self-managed super fund statistics, as reported by SMS Magazine (10 December 2025). smsmagazine.com.au
• Australian Taxation Office, "How SMSFs are taxed". ato.gov.au
• Colonial First State, "Federal Budget 2026-27" (super's tax treatment and Division 296). cfs.com.au
• Dario Amodei on AGI by about 2028 (interview). youtube.com
• Demis Hassabis on AGI by about 2030, Fast Company. fastcompany.com
• Dario Amodei, "Cancer is cured, the economy grows at 10% a year...", as reported by CNBC (28 May 2025). cnbc.com
• Cory Doctorow, "Ideas lying around" (2023), for the Milton Friedman quotation. doctorow.medium.com

The published model opens on D18. This appendix describes every control on its page, in the order the page shows them: six groups, with the page's own headings inside them. For each control it gives the value D18 uses, the range the page allows and what the control does. Appendix B lists the same package by what it does rather than where it sits on the page, and chapters 4 and 7 explain the settings that matter most.

Dollar settings are in 2026 dollars, as the page states them. Amounts a setting produces in a later year are given in the dollars of that year, with what they buy at today's shop prices where a person receives them (chapter 2). "Output" is the model's measure of national output, before consumption taxes, so its shares run a little higher than shares of GDP at market prices (chapter 1). The help text beside each control on the page often says more than this appendix; where it quotes a different default, the value here is D18's.

Some controls change nothing in D18, usually because another control has switched off what they act on. The tables say so, and they are listed together in the last section but one.

E.1 The presets

Seven buttons at the top of the page load complete scenarios. Each first puts the technology settings back to the base case's and then applies its own changes; four also change policy. "Restore all defaults" returns every control to D18.

At the foot of the page, "Current settings" lists every control's value and holds them all as a block of text. Copy it to keep a run; paste one in and press Set to reproduce it exactly.

E.2 The AGI scenario

The page describes this group as AGI's capability, cost, speed of deployment and effect on labour. These settings describe the technology and the world it arrives in. None of them is policy.

Key AI scenario settings

The technology clock

The page describes the clock as one S-shaped path for the AI frontier: slow, then fast, then slow. What AI can do, how much it adds to productive capacity and what it costs all follow it, and firms take AI up along an S-curve of their own (chapter 7).

How exposure becomes displacement

What the AI costs, and for how long

The plant it takes, and the world it lands in

E.3 The redistribution

The page describes this group as the package being tested. These are the policy settings: the payment, what it replaces and what pays for it.

The payment

What it replaces

What pays for it

The taxes, by year

Every tax rate and threshold is set at six markers, 2030 to 2055, and moves in a straight line between them. The page shows them as three tables with a row for each marker and a column for each tax, under two headings, "Consumption and company tax, by year" and "Personal income tax, by year". A value in bold blue differs from the row above; to hold a tax steady, give it the same value in every row. D18's values:

• Base GST (0–30%) is charged on everything households buy, rent included, with no exemptions for food, health or education. D18 keeps it at 10% in 2030 and ends it in 2035.
• Max AGI levy (0–200%) is the most any industry pays. Each industry pays this rate times the share of its human work that AGI has taken over, measured by how far its wages have fallen against its sales, on its prices before tax (chapter 13). Housing is never levied. Above 100% the rate is more than the firm's own price before tax. D18 does not need the rate above 100% to pass: capped at 100% it still passes all four tests, with fewer people in work and higher debt; capped at 80% it fails the fiscal test.
• Company tax (20–50%) is charged on profits. The model has no profit shifting or offshoring that grows with the rate, so it scores high rates too kindly; the page's help text puts Australia's realistic range at 30% to 40%.

The thresholds (tax-free $0–120,000; bracket 2's ceiling $30,000–200,000; bracket 3's $60,000–300,000; bracket 4's $60,000–500,000) are stated at 2030 incomes, so $25,000 is about $21,900 of a 2026 income (today's tax-free threshold is $18,200). From 2030 they rise with output per person at the bracket indexation rate below.

The rates (bracket 2's 0–50%, the others 0–60%) are whole rates: there is no separate Medicare levy. Bracket 2 is where most people with a job and the GDI sit, so its rate does more to their purchasing power than the top rate does. The bracket 3 and 4 rates, 45% and 48%, are set high to collect more in the build-out (chapter 4).

Under the same heading:

How the base responds

E.4 The external account

The page describes this group as who funds the deficit, and for how long.

The limit on foreign borrowing

What is owed, and what it costs

Trade

Reaching the constraint without braking

E.5 Public investment

The page describes this group as the mechanism that absorbs capacity to build investments that can earn an income at home and from abroad.

The programme

Setting it year by year

E.6 The fund

The page describes this group as the pool of about $6 trillion that replaces superannuation (chapter 14).

Return and drawdown

The portfolio

Market shock

E.7 Model mechanisms

The page describes this group as how the model works, rather than what policy does. These settings are the model's structure. They are on the page so that a reader can see what each one contributes, not as choices a package makes.

Model closure

Spending and saving

Investment and the brake

What does not get cheaper

Run options

E.8 Settings that change nothing in D18

Each control was moved one step from D18's value, or switched, and also set to each end of its range, with everything else left at D18. These change nothing in the run anywhere in their range, except for some a figure the page shows beside them. Most act on something another setting has switched off, and matter once it is on:

• The AGI productivity bonus, the post-2035 productivity bonus, the capability freeze, the AI cost decline and the year it stops: the technology clock sets capability, productive capacity and cost instead. They act when the clock is off.
• Productivity of reallocated labour: nobody is re-employed. It matters in the three study presets, which reallocate everyone.
• Debt level where the risk premium starts: there is no premium.
• Dollar response to the rate gap, and external debt in foreign currency: with the world's rates following Australia's and the world rate at the 5% neutral rate there is no gap, so the dollar never moves.
• Trade elasticity: import prices fall as Australia's do and the dollar does not move.
• Land tax phase-in: there is no land tax.
• Pausing the programme at the external limit: the limit never binds.
• The older financing ceiling's level: the older rule is off.
• The four market shock settings: there is no shock.
• Voluntary withdrawal: the page shows what it would cost, but the model does not apply it.

Six change nothing one step from D18 but change the run further along their range, because they are limits D18 does not reach:

• The external position ceiling and its approach period. The rule limits each year's deficit, not the stock, and D18's deficit is never within 2.1 points of output of what it allows. The ceiling can fall to 70% without the rule binding; at 65%, or with an approach period of 8 years, it binds in 2036 alone.
• The programme's capacity target. The 3% deficit guard, not the target, limits the programme until it reaches its cap.
• Household consumption: below 50% it starts to ration households' spending; at 45% D18 fails the fiscal test.
• The surplus waterfall's debt floor and its limit on the GDI's rise. D18's surplus lifts the GDI from 2053, but a small move in either changes nothing; lowering the limit to 0, or moving the floor far, does.

The external rule's switch changes only figures the page reports in D18's own run. Under stress the rule binds more often; its breaches are reported beside the external test rather than scored in it (chapter 7).

The surplus waterfall acts in D18: its budget is in surplus with debt below the floor from 2052, and the waterfall lifts the GDI from 2053.

A few more change who gains but not the economy's totals: ease of changing jobs, foreign firm share, foreign incidence to consumer, housing costs split by tenure and housing share of spending. The deferred-pay threshold changes the rate charged above it and what each household can spend while deferred pay is collected, and, a little, the economy, because it decides whose spending the collection restrains.

E.9 Controls that change the view, not the run

The rest of the page has a few controls of its own. They change what is shown, never the run: the year chosen for the detailed tables, and the personal scenario, which applies the settings above to a household the reader describes.

This model started as a spreadsheet and grew. As someone who has long enjoyed writing programs as a way of thinking and who finds value in creating models to identify and try to clarify at least some assumptions, I started writing this one too.

Having begun last year to replace some day-to-day manual programming with AI coders, and knowing my macro-economic, tax and transfer background was woefully inadequate to finish this model in any reasonable time, I began using Gemini, Copilot, ChatGPT and mostly Claude to explore the domain and implement various approaches as code. The speed was liberating: ideas are cheap to try, evaluate, iterate, discard, evolve, and an AI helps keep track of what worked, what needs more more, what can never work. Adding a new graph to visualise what X does to Y, or turning everything on its head is never too much trouble, and this lack of friction encourages deeper and ever deeper exploring and never being afraid to ask "what if?".

The model's goals are mine, its philosophy is mine, its methods are mostly mine, but its code is entirely written by AI. This commentary was entirely mine for the previous published draft, but was rewritten and expanded by AI, then edited by me.

All mistakes, naiveties and poor judgements remain entirely my own.

I realise it is a bit "meta" using an AI to model the possible effects that its imminent successors will have on how we think and work, but to me at least, it seems to become increasingly natural as the capabilities of AI improve. Even over the months of this exercise, the assistance provided by AI has markedly improved. Five months ago, mistakes were to be expected and genuine insights were rare. Last week, Claude Opus 5.5 suggested of its own volition that an interesting way to cap domestic demand in a period during which large investment was required for building productive assets would be to try an idea Keynes suggested in 1940: deferred pay, which, to my surprise, had even been picked up and promoted by RG Menzies during his war-time addresses. I'd never heard of it.

If AI does indeed reach AGI capabilities by 2030 and the implications for replacement of human labour are as vast as its builders and financiers are presuming, then our response will need to resemble a war mobilisation - not to engage with an enemy but to prepare our society to make best use of the opportunity AI presents.