45 Tough Questions.
45 Honest Answers.

COAD never touches the Future Fund's capital. The AUD 269.1 billion corpus stays intact — only the annual investment returns (~7%, roughly $19 billion per year) contribute to COAD. The original purpose of covering public servant superannuation is also protected: the government has confirmed no corpus drawdown is required until at least 2032–33, and actuarial projections show superannuation liabilities are fully manageable from returns alone.

Think of it like a savings account: COAD uses the interest, not the principal. The fund is preserved indefinitely.

Source: Future Fund Act 2006 s.18; COAD INI-004 v5; Commonwealth actuarial statements 2024

The 7% assumption is actually conservative. The Australia Future Fund has averaged 8.4% over 18 years, Alaska's Permanent Fund 8.9% over 48 years, and Norway's sovereign fund 9.2% over 35 years. Even using the lower long-run figure, market volatility is addressed through the Three-Pillar design — the Future Fund is only one of three funding sources.

If returns fall, the AI Headcount Tax and bond financing absorb the gap. An 18% structural buffer is built into the model from Year 15 onwards, and the payment schedule is designed to ramp up gradually as funding capacity grows — so the system doesn't over-extend in bad years.

Source: Future Fund Annual Reports 2006–2024; Alaska APFC; Norway GPFG performance data; COAD INI-004 v5

The tax extends well-established principles — company tax already captures productivity gains through increased profits, and robotics/automation have always been subject to normal payroll and income tax. The UK, France, and Canada already operate Digital Services Taxes on the same principle. This is evolution, not invention.

The tax is designed as a headcount-based measure: companies that reduce their human workforce through AI adoption pay a levy proportionate to that reduction. This is measurable (ATO payroll data), auditable, and hard to restructure around without reversing the productivity gains that triggered it in the first place.

Source: UK DST; OECD Pillar One/Two framework; ATO payroll data infrastructure; COAD INI-004 v5

Australia's debt-to-GDP ratio is around 22.5% — compared to the OECD average of 68%, Japan at 266%, and the US at 129% (which lost its AAA rating only at that level). Even at peak COAD bond issuance, Australia's ratio stays well below the thresholds that trigger rating downgrades.

Critically, the bonds are funding productive economic investment — maintaining consumer spending and preventing the demand collapse that mass unemployment would create. Rating agencies assess debt sustainability, not just debt levels. A funded COAD is far less risky to credit markets than 2.54 million unemployed Australians with no income.

Source: Australian Treasury Budget 2024–25; OECD Fiscal Monitor; S&P rating methodology; COAD INI-004 v5

The 18% is the minimum buffer at Year 15 (peak demand). In earlier years the buffer is much larger: 105% in Year 1, 70% in Year 5, and 36% in Year 10. The system builds resilience while it has headroom, not when it needs it.

If displacement accelerates significantly, COAD has multiple response levers: extending bond issuance (within safe credit limits), adjusting the AI Headcount Tax rate, or temporarily reducing payment growth rates. The three-pillar structure is specifically designed so that no single shock to one source collapses the program.

Source: COAD Sensitivity Analysis INI-004 v5; Treasury scenario modelling methodology

This is the sharpest funding challenge and COAD addresses it on three grounds. First, the government's actuarial position is that corpus drawdown is not required before 2032–33 at the earliest — meaning returns are available. Second, the superannuation liability profile is well-understood and fully modelled: the Future Fund has surplus returns above the liability draw rate for the entire COAD program period. Third, if competing demands do emerge, COAD's bond pillar can increase proportionately — the three-pillar structure exists precisely to handle this kind of single-source constraint.

Source: Future Fund Act 2006; Commonwealth actuary projections; COAD INI-004 v5 Section 3 Sensitivity Analysis

The Australian Government borrows from global capital markets by issuing Commonwealth Government Securities (CGS) through the Australian Office of Financial Management (AOFM). Investors — superannuation funds, foreign central banks, and institutional investors — buy the bonds, the government receives the cash, pays interest over the bond's life, and repays principal at maturity. For COAD, Parliament would authorise a dedicated borrowing facility, potentially structured as a purpose-specific Social Bond, which is well-established practice internationally for social policy programs.

Australia is unusually well placed for this. A AAA credit rating from all three major agencies (S&P, Moody's, Fitch) means borrowing at very low interest rates. A debt-to-GDP ratio of 22.5% — against an OECD average of approximately 68% — means genuine fiscal headroom exists. And Australia's $3.9 trillion superannuation sector creates deep institutional demand for exactly the kind of long-dated, secure government bonds COAD would issue.

The bonds are designed as a bridge, not permanent debt. In the early years (2027–2030), while the AI Headcount Tax is being stood up and Future Fund returns redirected, bond proceeds fund the payments. As the tax matures from approximately 2029 onward, that revenue services bond interest and retires principal — replacing the borrowing rather than adding to it. By the mid-2030s, if both other pillars are fully operational, new bond issuance drops sharply and the program becomes largely self-funding. The interest cost is real — at current 10-year bond yields of approximately 4.5%, a $10 billion issuance costs around $450 million per year in interest — which is precisely why fast-tracking AI Headcount Tax legislation is a design priority.

On the "debt-funded welfare" framing: the distinction that matters is whether there is a defined revenue stream retiring the debt. There is. The COVID-era bond program funded JobKeeper at far greater scale with no defined retirement mechanism — COAD's three-pillar structure is more fiscally disciplined than that precedent, not less.

Source: Australian Office of Financial Management (AOFM) — CGS issuance framework; S&P, Moody's, Fitch sovereign credit ratings for Australia; IMF Article IV Consultation, Australia 2025; AOFM CGS market data; COAD INI-004 v5.2 Section 3 bond pillar modelling

Yes — and this is one of the most important things to understand about how COAD is funded. The Future Fund's AUD 269.1 billion corpus is invested across a diversified portfolio: Australian and global equities, infrastructure, property, private equity, fixed income, and alternative assets. It has averaged approximately 7–8 per cent per year in returns over its history, generating around $19 billion per year. Under COAD, the corpus is never touched. Only the annual investment returns — the dividends, capital gains, and interest earned from those markets — are redirected to fund COAD payments.

So when the stock market performs well, Future Fund returns increase, and more money flows to displaced workers. The fund's equity holdings do the work; the underlying capital stays intact and continues compounding.

There is an elegant — and deliberate — logic to this arrangement. The companies driving AI adoption at scale — the technology giants, platform firms, and AI infrastructure providers whose share prices are rising rapidly — are precisely the companies whose market performance is generating those Future Fund returns. The corporations displacing workers through AI are, indirectly, funding the payments to the workers they displaced. COAD essentially recaptures a share of AI-driven corporate profit through the Future Fund's equity holdings and redirects it to the people bearing the human cost of that automation.

The other two pillars work differently. The AI Headcount Tax is a direct levy on businesses that reduce human headcount through AI — not investment returns, but a tax on the act of substitution itself. Sovereign Bonds are borrowing against future revenue, not market returns. Only Pillar 1 works through investment and the stock market — but it is the most philosophically coherent of the three, because it means the financial beneficiaries of AI displacement are structurally enrolled in funding its social consequences.

Source: Future Fund Annual Report 2024–25; Future Fund Act 2006 (Cth); Future Fund portfolio allocation data; COAD INI-004 v5.2 Section 3 Pillar 1 modelling; ASX and global equity market return data

Previous automation targeted physical tasks — assembly lines replaced manual labour, ATMs reduced bank tellers, but new cognitive jobs emerged to absorb displaced workers. The critical difference with AI is that it targets cognitive tasks simultaneously across all sectors — the refuge jobs that workers historically retrained into are the very jobs AI is now absorbing.

Every previous automation wave created enough new roles to absorb displaced workers because there was always a higher-skilled cognitive tier to move into. Agentic AI is now automating that tier as well. The refuge has gone.

As of May 2026, the evidence has moved beyond institutional projections to named corporate forecasts and measured employment data. Mustafa Suleyman, Chief Executive of Microsoft AI, stated publicly that AI will reach human-level performance on "most, if not all, professional tasks" within eighteen months — naming accounting, legal, marketing and project management specifically. The Stanford 2026 AI Index records a 20 per cent decline in entry-level software-developer employment in the United States since 2024. The Reserve Bank of Australia's May 2026 Statement on Monetary Policy expects the productivity inflection point in 2027 — the same year as COAD's Year 1. These are not projections; they are observed data and on-the-record forecasts from the institutions and executives directly inside the transition.

Source: IMF World Economic Outlook 2024; WEF Future of Jobs Report 2025; Acemoglu "Automation and New Tasks" 2022; Suleyman, Fortune, 16 May 2026; Stanford HAI 2026 AI Index Report; RBA Statement on Monetary Policy, May 2026

As of mid-2025 this is measured data, not projection. The Federal Reserve Bank of Dallas finds that AI is already creating a 0.5–0.7 percentage point annual drag on CPI, with long-run inflation expected to anchor at 1.8% rather than the traditional 2.5–3%. Software costs have fallen 40% in sectors with heavy AI adoption. Legal services, accounting, and radiology are seeing measurable price compression already.

For COAD's purposes, moderate deflation is actually positive: the real purchasing power of COAD payments grows over time without requiring higher nominal payments.

Source: Federal Reserve Bank of Dallas, June 2025; BLS sector-specific CPI data; IMF Fiscal Monitor 2024

The figure is triangulated from multiple authoritative international sources. The IMF finds 40% of global jobs are exposed to AI, with 60% of advanced economy jobs at risk. The WEF's 2025 Future of Jobs Report projects 83 million global jobs displaced versus 69 million created by 2027. Applied to Australia's 14-million-strong workforce with an occupation-level mapping (using the OSCA register), 2.54 million represents the cumulative displaced cohort who require income support — not all at once, but progressively over 14 years (2027–2041).

Source: IMF World Economic Outlook 2024; WEF Future of Jobs Report 2025; ABS Labour Force Survey; COAD INI-004 v5 assumptions log

If AI creates as many jobs as it destroys on net, COAD is a cheap insurance policy. Fewer displaced workers means lower demand, a larger funding buffer, and potentially an accelerated exit from the program. The three-pillar model scales down as easily as it scales up — unused capacity can be redirected to retraining programs or payment rate increases.

The most credible academic counterpoint is MIT economist Daron Acemoglu's published estimate that an upper bound of jobs meaningfully affected by AI and computer-vision technologies within the next ten years is "less than 10 per cent". COAD's model is calibrated to approximately 14.7 per cent of the Australian workforce at Year 15 — carrying an 18 per cent buffer above peak demand. The model is deliberately sized to be robust across the full range of credible estimates, including Acemoglu's conservative upper bound.

The deeper issue is distributional, not aggregate. A May 2026 MIT study by labour economist David Autor and colleagues finds that, across the postwar United States, new technology-enabled work was filled disproportionately by young workers under 30, university graduates, and urban workers — with university graduates 2.9 percentage points more likely than high-school graduates to be engaged in new work. Even on an optimistic assumption that AI creates as much new work as it displaces, that new work historically accrues to a different cohort than the mid-career, non-tertiary-credentialled, often regionally located workers COAD is designed to support. The distributional gap between who loses work and who gains it is the COAD value proposition — and it survives even the most optimistic net-employment forecast.

The asymmetry remains: if we're wrong to worry, the cost is manageable. If we're right and do nothing, the cost is a generation without income support.

Source: Acemoglu, D. — standing published position; MIT News — What Makes New Work Different, 21 May 2026 — news.mit.edu; COAD FIN-001 v2.6 buffer analysis; Treasury risk management framework principles

COAD's case has never rested on a simple "exposure equals displacement" equation. Three points apply.

First, the OpenAI framework is consistent with COAD's design. COAD is sized for a phased displacement scenario maturing over fifteen years, not for an instantaneous capability-to-displacement transition. The framework's "capability leads, usage lags" finding is one of the reasons COAD adopts a fifteen-year build-out rather than a near-term shock response.

Second, the OpenAI framework reinforces, not weakens, the need for an Australian institutional capacity to measure AI usage in real time. That capacity is the AI Agency proposed in INI-001 / INI-002. The Anthropic Economic Index data-sharing arrangement under the Australian Government–Anthropic MOU (signed 1 April 2026) is precisely the kind of usage-measurement feed required to track the capability-to-deployment lag as it closes.

Third, the framework's logic cuts in COAD's favour on timing. The alternative to acting on exposure signals is acting only after displacement has occurred — but by then the lead time to legislate, capitalise, and operationalise the fund has already been lost. Consistent with PMBOK 8th Edition's Uncertainty performance domain, policy infrastructure for a high-impact risk must be in place before the risk crystallises, not after.

Source: The AI Jobs Transition Framework — OpenAI, April 2026 — cdn.openai.com; Australian Government and Anthropic MOU, 1 April 2026 — anthropic.com

This is the most common objection to guaranteed income programs — and the empirical evidence directly refutes it. In the Stockton SEED pilot, recipients were 28% more likely to find full-time employment than the control group. Alaska's Permanent Fund has paid universal dividends to 731,000 people for 48 years, and Alaska consistently maintains higher workforce participation than the US average.

COAD is explicitly designed at 70% of minimum wage so that employment always pays meaningfully more. Someone who takes a job at minimum wage earns $15,000+ more per year — a 43% income premium. The evidence shows people use income security to search for better jobs, not to avoid work entirely.

Source: Stockton SEED Evaluation 2021; Alaska Permanent Fund Corporation; Damon Jones & Ioana Marinescu, JPE 2022

Stockton is one data point among many. Alaska's Permanent Fund has operated at scale — 731,000 recipients, 48 consecutive years — with no evidence of work disincentive. Finland's 2017–2018 basic income experiment (2,000 participants) showed improved employment and wellbeing. Kenya's GiveDirectly program (tens of thousands of participants over a decade) shows consistent positive employment effects across diverse populations.

The concern about small pilot generalisation is valid, which is why COAD's evidence base draws from multiple large-scale real-world examples rather than relying on any single trial.

Source: Alaska APFC annual reports; Finland Kela basic income study 2020; GiveDirectly longitudinal evaluation 2024

This compares the wrong things. Work incentives are based on nominal dollars — employers pay in nominal dollars, workers receive COAD in nominal dollars, and the spending choice happens in nominal terms. The work incentive is the $15,000+ additional nominal income from minimum wage employment.

The PPP figure adjusts for international comparison purposes (to show what $35,000 buys relative to other countries), not for domestic spending decisions. A displaced Australian deciding whether to take a job is comparing $35,000 (COAD) versus $50,284 (minimum wage) — not PPP-adjusted figures. The 43% income premium for working remains real and meaningful.

Source: FWC Minimum Wage Order 2024; COAD INI-004 v5 payment schedule modelling

Constitutional analysis supports COAD under existing Commonwealth powers. Section 51(xxiii) grants power over social welfare, unemployment benefits, and similar payments — COAD falls squarely within this. Section 96 grants power to make financial assistance to states. Section 81 provides appropriation authority for expenditures on government purposes, which has been broadly interpreted.

The High Court's interpretation of social welfare powers has expanded significantly since the 1940s — Medicare, Family Tax Benefit, and JobSeeker all operate under the same framework. COAD is structured as targeted income support for a specific displacement event, which is well within established precedent.

Source: Commonwealth Constitution ss.51, 81, 96; High Court welfare power jurisprudence; COAD legal analysis

COAD transcends the traditional welfare debate because it's fundamentally not welfare. For Coalition values: it's funded through investment returns and a market-based AI productivity mechanism (not redistribution), it preserves work incentives, and it protects consumer demand — maintaining the economy that businesses depend on. For Labor values: it provides dignity and economic security for workers displaced through no fault of their own, preventing the poverty trap of inadequate JobSeeker.

The analogy is Medicare — initially controversial, now untouchable across party lines. Structural economic protection for citizens displaced by forces beyond their control is not a left/right question once the displacement is real and visible.

Source: COAD political strategy analysis; Medicare political history; Treasury consultation framework

The counterargument to business is their own self-interest. Consumer spending is 55% of Australian GDP. If 2.54 million workers lose their income with no support, consumer demand collapses — which is bad for every business, AI-adopting or not. COAD maintains the spending capacity of displaced workers, directly benefiting the businesses whose AI investments caused the displacement.

The AI Headcount Tax is also a predictable cost that can be modelled and planned for — far preferable to the regulatory uncertainty of multiple ad-hoc government responses. Progressive businesses understand that social licence for AI adoption depends on visible evidence of shared benefits.

Source: ABS National Accounts; RBA consumption data; Business Council of Australia AI policy submissions

Services Australia has demonstrated it can scale massively when required. During COVID-19, it processed 1.6 million JobSeeker claims in four weeks, scaled from 800,000 to 2.4 million recipients in 90 days, and maintained 99.7% payment accuracy while handling 10× normal demand. The system can scale — the challenge is political will and preparation, not technical capacity.

Importantly, COAD is structurally simpler than JobSeeker: eligibility is occupation-based (linked to the OSCA register), not means-tested, which removes the most complex and error-prone assessment processes. Fewer decisions means fewer errors.

Source: Services Australia COVID-19 operations report 2020; ANAO performance audit; COAD PDB-001

COAD solves the attribution problem by shifting it from the individual to the occupation level. Rather than asking "was this person's job eliminated by AI?" (which is impossible to prove), COAD asks "is this occupation type structurally at risk from AI?" — which is assessable using occupation-level data, employment statistics, and workforce modelling.

The Occupation Standard Classification for Australia (OSCA) register lists which occupations qualify. If your occupation is listed and you're unemployed, you're eligible — no individual causal dispute required. This is the same principle as workers' compensation: we don't require proof of exactly which action caused an injury; we assess based on occupation-level risk profiles.

May 2026 provided a concrete illustration of why employer self-declaration cannot be the basis for attribution — in either direction. Intuit's Chief Executive publicly stated that the company's approximately 17 per cent workforce reduction "had nothing to do with AI" — while simultaneously reorienting the company toward AI-first operations. Cloudflare's Chief Executive, in the same week, published a detailed public thesis explicitly attributing a 20 per cent workforce reduction to AI automation of coordination, finance and middle-management roles. Two firms making structurally identical workforce decisions: one denying AI causation, one asserting it. An eligibility system reliant on employer self-declaration would produce opposite outcomes for workers in identical situations. COAD's occupation-and-task-exposure design is robust under exactly this kind of attribution failure.

Source: COAD PDB-001 eligibility framework; OSCA register design; Workers' Compensation Act precedent; Intuit CEO, CNBC, 20 May 2026; Cloudflare CEO, Fortune, 21 May 2026

COAD has structural fraud resistance built in from the design stage. The citizenship and residency requirements alone eliminate the non-citizen fraud that accounts for the majority of welfare fraud attempts. Employment status is verified in real time through ATO PAYG data — a payment stops automatically when employment income is reported. The occupation-based eligibility (OSCA register) means eligibility is determined by ABS-maintained occupational data, not individual self-declaration.

There are no "cash in hand" payments and no complex means-testing — the simplicity that makes COAD administratively lean also makes it harder to game than a system with hundreds of means-test thresholds and conditional requirements.

Source: ATO PAYG data architecture; Services Australia fraud framework; COAD PDB-001

This is a genuine and important policy gap. COAD's primary eligibility framework is designed for workers who held an OSCA-listed occupation and lost that role to AI-driven structural change. A person entering the workforce for the first time — with a qualification or trade training in an AI-affected field but no employment record — cannot satisfy a displacement test, because displacement requires a prior employment state to have been disrupted.

The underlying harm is real but different in kind: it is structural labour market entry failure — where AI has reduced or eliminated entry-level vacancies in a field before the person ever had the chance to enter it. A data entry graduate, a paralegal completing their degree, or a logistics trainee finishing their certificate may find that the occupation they trained for no longer generates hire volume, yet they have never been employed and therefore cannot be "retrenched."

INI-004 v5.2 flags this cohort under the Graduate AI Displacement Bridge (GADB) concept — a proposed supplementary pathway for new labour market entrants whose target occupation is OSCA-listed and where ABS vacancy data shows the entry-level hire rate has declined materially relative to graduation volumes. Under GADB, eligibility would be assessed on the basis of: (a) a completed qualification or vocational credential in an OSCA-listed occupation; (b) documented, unsuccessful job market entry attempts over a defined period; and (c) a structural vacancy decline threshold confirmed by ABS Labour Account data, rather than individual displacement evidence.

GADB has not yet been legislatively designed. It represents the next layer of policy development required before COAD can be considered complete for all affected cohorts. Until that pathway is formalised, first-time entrants in AI-affected occupations would access existing JobSeeker Payment and Youth Allowance arrangements — a temporary gap that the COAD project design team has identified as priority work for the next phase.

Source: COAD PDB-001 eligibility framework; INI-004 v5.2 ASM-S03 (structural labour market assumptions); OSCA register design principles; ABS Labour Account, Australia (cat. 6150.0)

This is a fair and important challenge. $20,000 per year is below the Henderson Poverty Line for a single adult in Australia (approximately $33,200 per year in 2024–25 terms), and the COAD design team does not dispute that. The $20,000 figure is the floor of a payment that rises progressively to $35,000 per year by 2041 — and it is intended to sit alongside existing Commonwealth income support, not replace it.

COAD's design intent is that recipients retain access to applicable Commonwealth payments they already qualify for. A displaced worker not engaged in paid work would typically also be eligible for: Commonwealth Rent Assistance (up to approximately $4,900/yr for singles renting privately); the Energy Supplement; a Health Care Card providing concessional medicines, bulk billing, and public transport discounts; and Family Tax Benefit where relevant. These supplements materially increase the effective income floor above the COAD payment alone.

The more complex question is how COAD interacts with JobSeeker Payment under Centrelink means testing. If COAD is treated as assessable income under the Social Security Act — as most regular payments are — it would likely taper out most or all of a concurrent JobSeeker entitlement at the standard 50 cents-in-the-dollar reduction rate above the income free area. The COAD policy design team has identified this as an open legislative design question: a specific COAD income-test exemption (analogous to exemptions already in place for NDIS and some veterans payments) would need to be legislated to allow genuine stacking. That work has not yet been completed.

What is settled is the direction: COAD is a floor, not a ceiling. Combined with non-cash concessions, Rent Assistance, and a resolved means-testing framework, the policy aims to bring displaced workers to an adequate — if modest — income during the transition period. The adequacy gap at the $20,000 starting rate is real, and it is the strongest argument for fast-tracking the means-testing resolution as a priority legislative task before the 2027 commencement date.

Source: Melbourne Institute Henderson Poverty Line, March Quarter 2025; Social Security Act 1991 (Cth) income test provisions; Services Australia Commonwealth Rent Assistance rates; NDIS income-test exemption (Social Security Act s. 8(8)(y)); COAD INI-004 v5.2 payment adequacy notes

Welfare expansion was formally analysed and rejected for structural reasons. JobSeeker's means-testing creates high effective marginal tax rates above 60% — it actually penalises recipients for taking low-paid work. It also creates permanent fiscal burden with no programmatic end point, and the stigma and conditionality of "welfare" reduces compliance and dignity.

COAD, by contrast, pays at a flat rate with no means test, is explicitly time-limited (2027–2041), and is funded through dedicated sources rather than consolidated revenue. It is an income offset for a structural market failure — categorically different from welfare.

Source: COAD Business Case Option 2 analysis; ACOSS welfare reform analysis; Henry Tax Review effective marginal rates

Full UBI is fiscally unviable at any meaningful payment level. 20 million adult Australians × $20,000 per year = $400 billion annually — equivalent to 62% of total Commonwealth tax revenue. This would require either quadrupling taxes, eliminating all other government services, or printing money. None of these is realistic.

COAD is not UBI-lite — it's a targeted response to a specific structural problem: AI displacement. By limiting eligibility to workers in verified AI-displaced occupations, COAD achieves a meaningful payment level ($20K–$35K) at a fundable cost ($76B at peak) rather than an inadequate payment spread thinly across everyone.

Source: ABS adult population data; Commonwealth Budget 2024–25 revenue figures; COAD Business Case Option 3 analysis

Retraining is necessary — but it's insufficient as the only response. The scale problem: 2.54 million people cannot all become AI engineers or data scientists. The speed problem: AI advances faster than retraining cycles (see Challenge 6.5). The demographic problem: older workers face real and acknowledged skill ceilings. The targeting problem: what do you retrain people into when AI keeps advancing?

The Davos 2026 retraining pledge covered only 5% of globally affected workers. Historical retraining programs (US Trade Adjustment Assistance: only 37% found comparable jobs; coal miner retraining: most never transitioned) show the limits. COAD includes retraining — recipients can study while receiving payments. COAD + retraining is the answer, not retraining alone.

Source: US TAA Program Evaluation 2023; WEF Reskilling Pledge 2026; Brookings Coal Transition Analysis; COAD INI-004 v5

A Job Guarantee has theoretical appeal but serious practical problems. The government would need to create 2.54 million meaningful positions — not fake or make-work jobs. Geographic mismatch (jobs where government needs them, not where workers are), skills mismatch (displaced office worker assigned to road maintenance), and a massive administration cost make this extremely difficult at scale.

Cost comparison: a Job Guarantee at minimum wage would cost $128 billion annually, rising to ~$180 billion with supervision and administration — 2.4× the cost of COAD. It also requires accepting assigned work, whereas COAD preserves worker autonomy to find better employment on their own terms. Evidence shows people use income security to find better jobs, not to avoid work.

Source: Levy Economics Institute JG proposal; Centre for Full Employment and Equity Research; COAD Options Analysis

Yes — and this is the defining structural problem that makes retraining-only policy fundamentally inadequate in the agentic AI era. Challenge 6.3 shows retraining fails at scale. Challenge 6.5 shows it also fails in time.

Pre-agentic automation (2010s) displaced a role, but workers could retrain for a stable target. Agentic AI (2025+) displaces the original role and absorbs the retraining destination role during the same 12–24 month training period. AI capability is approximately doubling every 12 months — faster than any certification can be completed.

Examples: a telemarketer retraining as a project manager over 12 months arrives to find agentic AI already performing PM coordination at scale. A data entry clerk retraining as a data analyst over 18 months emerges into a market where AI handles 80%+ of standard analysis. The target keeps moving.

COAD's response: it doesn't promise retraining will restore employment — it provides a stable income floor regardless of outcome. The 14-year program horizon is explicitly designed for sustained, rolling displacement rather than a one-time transition. COAD is the bridge to a new economic structure, not a promise that the old one returns.

Source: Stanford HAI Index 2025; McKinsey "Generative AI and the Future of Work" 2024; WEF Future of Jobs 2025; COAD INI-004 v5

The funding source is different — but the policy lessons remain valid. Alaska proves that universal payments to large populations do not create dependency (68% workforce participation, above US average). It proves that such programs can achieve permanent bipartisan political support (48 years with no meaningful rollback attempt). And it proves the administrative mechanics can work at scale with high accuracy.

The Australian equivalent isn't oil revenue — it's the sovereign wealth accumulated through Future Fund investment returns, which provides the same "common wealth returns to citizens" principle. The source is different; the logic and the lessons are the same.

Source: Alaska Permanent Fund Corporation 2024 Annual Report; APFC workforce participation data; COAD precedent analysis

COAD doesn't try to replicate Norway's model. Norway's fund is universal — designed to fund all government services for all citizens indefinitely, as a complete replacement for tax revenue. COAD is targeted — it supports a specific cohort (AI-displaced workers) for a defined period (2027–2041) at a specific payment level.

The comparison is like saying Australia can't have Medicare because Norway's healthcare system is funded differently. COAD draws on Norway's lessons about sovereign fund governance and intergenerational equity — not its scale or coverage design. Australia's AUD 269.1 billion Future Fund is appropriately sized for COAD's targeted purpose.

Source: Norway GPFG annual report 2024; Future Fund Act 2006; COAD design principles INI-004 v5

This is accurate — and it's a feature, not a bug. Countries that adapt to structural economic change first gain competitive advantage. Australia positioning as a "responsible AI jurisdiction" can attract AI investment from companies that need social licence. The COAD framework can be exported and referenced by other nations facing the same challenge.

It's also worth noting that every major social innovation was unprecedented when introduced — Medicare, superannuation, compulsory voting. Australia has a history of successful policy innovation that other nations subsequently adopted. The risk of acting first is real; the risk of not acting is a generation without income support.

In April 2026, OpenAI — the world's most prominent frontier AI developer — published a policy blueprint titled Industrial Policy for the Intelligence Age: Ideas to Keep People First, explicitly proposing three mechanisms that are structural analogues to COAD pillars: automated labour taxes as a funding source (analogous to COAD Pillar 2), a Public Wealth Fund distributing AI-productivity gains to citizens (analogous to COAD Pillar 1 and the sovereign return mechanism), and adaptive safety nets with threshold triggers (analogous to COAD's FIN-001 section 6A recalibration mechanism). When the developer of the technology is independently converging on the same policy architecture, "unprecedented guinea pig" is no longer the right frame — it is mainstream policy thinking that Australia would be among the first to implement at sovereign scale.

Source: OpenAI, Industrial Policy for the Intelligence Age, 8 April 2026 — openai.com/index/industrial-policy-for-the-intelligence-age/; OECD AI Policy Observatory; Productivity Commission innovation policy analysis

The Bores model and COAD share a funding logic — a dedicated levy on AI economic activity, with proceeds reaching citizens — but differ on three design questions that matter for Australian conditions.

First, trigger logic. Bores is contingency-triggered: payments switch on when displacement indicators register. COAD is scheduled and anticipatory: graduated payments begin in Year 1 and scale over fifteen years. The Bores contingency design carries an institutional risk that the trigger threshold is set too tightly — either firing too late (after displacement has caused sustained harm) or becoming a political target for industry lobbying to raise the trigger. COAD's anticipatory schedule removes that single point of failure.

Second, asset base. The Bores model includes equity stakes in AI companies. COAD's Pillar 1 leverages the existing Australian Future Fund — an AUD 269.1 billion sovereign-asset vehicle already operating under the Future Fund Act 2006, with 18 years of proven governance. Building a new equity-stake mechanism from scratch carries implementation risk and political exposure that the Future Fund architecture avoids entirely.

Third, jurisdictional fit. The Bores design is calibrated to US federalism, the US tax base, and a US-citizen recipient pool. COAD is calibrated to Australian constitutional heads of power (section 51(xxiii) and section 51(ii)), the Future Fund Act 2006, and the Australian Government's existing fiscal architecture. The two designs are not interchangeable across jurisdictions.

The broader point: the two proposals are complementary in advocacy terms. The Bores plan demonstrates that AI-dividend mechanisms are now live policy options in comparable Western legislatures — strengthening, not challenging, the case for COAD.

Source: Alex Bores rolls out "AI dividend" plan — Axios, 20 April 2026 — axios.com; AI tax proposal: public ownership and governance — The Hill — thehill.com

The Sanders proposal — an AI Sovereign Wealth Fund Act teased in 28 April 2026 but not yet formally introduced — would impose a one-time 50 per cent tax, paid in the stock of the largest AI companies, with the government acquiring voting shares and board representation. This is compulsory equity acquisition. COAD is categorically different: it preserves an existing corpus (the Future Fund, built over 18 years under the Future Fund Act 2006) and draws only the investment returns on that corpus; it acquires no equity stakes and compels no share transfers. COAD is funded by three independently modelled pillars, targeted only at workers displaced by AI, conditional (eligibility ceases on retraining or re-employment above the income threshold), and time-limited. The Sanders proposal confirms that pairing AI-value capture with public distribution has entered mainstream legislative discourse — which supports COAD's mechanism as orthodox rather than novel — but the instruments are entirely distinct: sovereign-corpus-return-redistribution is not equity nationalisation.

Source: Senator Bernie Sanders, op-ed — sanders.senate.gov; Fortune, 3 June 2026 — fortune.com. Note: bill teased, not yet formally introduced.

Three points apply.

First, the market reaction is a signal of credibility, not unsoundness. Equity prices re-price the expected after-tax earnings of affected firms when a credible national mechanism for sharing AI-derived corporate surplus is publicly floated — that is exactly what economic theory predicts. The same re-pricing occurred when the United Kingdom introduced North Sea oil taxation in 1975 and when Australia introduced the Petroleum Resource Rent Tax in 1987. Both are now mature, accepted features of their respective fiscal landscapes, and neither prevented long-run investment or growth. A market that took the proposal seriously enough to re-price is a market that understands the proposal is substantive.

Second, COAD's Pillar 2 is narrowly calibrated. The marginal-impact analysis in FIN-001 sizes Pillar 2 at a five-to-fifteen per cent levy on measurable AI-derived productivity gains — not on total profits, share capital, or dividend distributions. The Australian sharemarket impact of a precisely scoped Pillar 2 will differ materially from the untargeted AI-tax framing that spooked the KOSPI. COAD's design is deliberate on this point.

Third, the Korea episode is a communications lesson, not a policy veto. Kim Yong-beom's proposal became public without a legislative framework, without industry pre-consultation, and without a marginal-impact analysis for affected sectors. The Presidential Office immediately distanced itself, characterising it as his personal view rather than a Government commitment. COAD's communications plan should therefore publish the Pillar 2 marginal-impact analysis ahead of any formal political announcement, and sequence the Future Fund partnership and sovereign bond pillars first — so Pillar 2 is the third, not the first, element the market encounters.

Source: Korea Roils Markets by Floating "Citizen Dividend" From AI Tax — Bloomberg, 12 May 2026 — bloomberg.com; Presidential official proposes "public dividends" from AI-driven boom — UPI, 12 May 2026 — upi.com

On 10–11 June 2026, Anthropic published an economic policy framework identifying basic income, sovereign-wealth mechanisms and AI-firm taxation as appropriate responses to severe AI-driven unemployment, backed by a USD 350 million Economic Futures Program (including 100 "Claude Corps" retraining fellowships at USD 85,000 per year). This is significant institutional corroboration — it shows the category of mechanism that COAD uses is gaining serious, industry-level acceptance.

But corroboration of the instrument family is not validation of COAD's specific figures. Anthropic's framework does not endorse COAD's AUD 35,000 annual payment, the 1.22–2.03 million displaced-worker planning baseline, or any of the three COAD funding pillars specifically. Each of those figures rests on independent Australian modelling set out in INI-004, FIN-001 and OPT-001 — and that modelling stands or falls on its own evidence base, not on whether a US AI developer agrees that "something like this" is warranted.

Anthropic's framework is also explicitly not a UBI: payments are targeted at workers displaced by severe AI unemployment, conditional on eligibility criteria, and backed by a dedicated Economic Futures Program — the same conditionality structure as COAD. That distinction matters for parliamentary scrutiny: COAD is not a universal handout, and neither is Anthropic's framework.

Source: Anthropic, "Policy on the AI Exponential" / Economic Futures Program — anthropic.com, 10–11 June 2026; COAD AI News Briefing, 15 June 2026, sections 2.1 and 5.

This is the most important fairness challenge COAD faces and it deserves a direct answer. Ideally, no one should receive less support than they need. But the correct response is not to dilute COAD — it's to address the inadequacy of standard unemployment support separately.

The principle of cause-contingent support levels is already well-established in Australian social policy: Workers' Compensation pays more than general sick leave for the same injury, because circumstances matter. Defence Service Pension, Disability Support Pension, and industry-specific structural adjustment payments (automotive, textiles, steel) all provide differentiated support based on the specific nature of displacement. AI displacement creates structurally different circumstances from cyclical unemployment — the displaced jobs are not returning.

Source: Social Security Act 1991; Workers' Compensation precedent; Steel/automotive adjustment payment history; COAD INI-004 v5

This concern is directionally correct and honest. As AI capability expands, more occupation categories will satisfy the register listing criteria. COAD's design anticipates this through a two-stage approach: the register lists occupations where AI has materially substituted labour (not merely supplemented it), and eligibility requires actual unemployment — not just occupation membership.

If AI genuinely displaces most occupations, then a payment program that scales to cover most displaced workers is the correct policy response — that's not a flaw. The program is designed to end in 2041; if by then AI has absorbed most occupations, the policy conversation will be about a post-labour-market income structure, not retraining.

Source: OSCA register methodology; COAD program sunset provisions INI-004 v5; ABS occupation classification framework

This is a real and important design consideration. A displaced accounts clerk in Broken Hill and one in Melbourne receive the same COAD payment — but their situations are structurally different. The Melbourne worker can access dozens of employers, multiple TAFE campuses, and a dense labour market to pivot into. The Broken Hill worker may have one major employer, limited local training infrastructure, and redeployment options that require leaving their community entirely.

The counterintuitive point is that COAD's flat payment is actually more protective in regional areas, not less. Regional cost-of-living indexes consistently run 10–20% below capital city averages for non-housing essentials, meaning the real purchasing power of a $20K–$35K COAD payment stretches further outside major cities. More importantly, when the alternative is zero income in a town with few employers, a guaranteed income floor is more critical than in a city where casual work and gig income can partially bridge a gap. COAD prevents the forced population drain from regional communities that unmitigated displacement would cause.

The genuine gap is on the retraining side, not the payment side. A regional recipient cannot easily access metropolitan TAFE or university campuses, and online delivery — while improved — is not equivalent for hands-on or laboratory-based programs. COAD's design acknowledges this through two mechanisms: first, recipients may study while receiving COAD payments (so income support doesn't expire during retraining); and second, INI-004 v5.2 explicitly flags a regional retraining loading as a recommended supplementary measure — additional Commonwealth funding for distance and online retraining access for COAD recipients outside major metropolitan areas, modelled on the existing Regional Education Support Package framework.

The flat payment is the right base — it ensures no one in a regional area receives less income support than their city counterpart for the same displacement. The regional loading addresses the asymmetric cost of accessing retraining, which is where the genuine inequality sits.

Source: Regional Australia Institute Regional Workforce Report 2024; ABS Regional Price Index; COAD INI-004 v5.2 regional equity provisions; DESE Regional Education Support Package framework

A CPU (Central Processing Unit) is the main brain of a computer. It handles everything your computer does: running your operating system, opening apps, browsing the web, sending emails, and doing calculations. Every computer, laptop, phone, and server has one.

CPUs are designed to be extremely fast and versatile — they can handle almost any task thrown at them, one after another, in rapid sequence. A modern CPU typically has between 4 and 64 processing cores, each extraordinarily powerful. Think of it as a small team of expert problem-solvers who can tackle complex, varied tasks in quick succession.

The CPU handles all the logic that controls AI software — loading the model, interpreting instructions, managing memory, sending results back to you. But when it comes to the heavy mathematical lifting that AI requires, CPUs hit a wall. That's where GPUs come in.

Source: Intel Corporation — CPU Architecture Overview; AMD Processor Design Guide

A GPU (Graphics Processing Unit) was originally designed for one specific job: generating the thousands of pixels that make up a video game image, many times per second. To do this, it performs millions of identical small calculations — colour, lighting, shadow — all simultaneously. The hardware design that emerged for this job turns out to be almost perfectly suited to running AI.

A GPU contains thousands of smaller, simpler processing cores working in parallel. NVIDIA's flagship AI chip (the H100) has over 16,000 individual processing cores on a single card. None of those cores are as powerful as a single CPU core — but together, doing the same calculation 16,000 times simultaneously, they can process AI workloads 10 to 100 times faster than a CPU.

Training a large AI model requires performing the same type of matrix multiplication billions of times. GPUs were built exactly for this. Modern AI would not exist at its current scale without the GPU — which is why NVIDIA's stock price increased by over 700% since 2022, and why access to GPU computing time is now a geopolitical resource.

The world's largest AI data centres — operated by Google, Microsoft, Meta, Amazon, and dedicated AI companies — contain hundreds of thousands of GPUs. Training a single large frontier AI model can consume more electricity than a small Australian town uses in a year.

Source: NVIDIA H100 Technical Specifications; IEEE Spectrum "The GPU That's Eating the World" 2024; International Energy Agency — AI and Energy Report 2024

The simplest way to understand the difference is few powerful vs. many simple. CPUs are built around a small number of extremely capable cores that can handle almost any task. GPUs are built around a massive number of simpler cores that excel at doing the same operation over and over in parallel.

In practice, AI systems use both. The CPU manages everything — loading data, accepting your question, sending you the answer. The GPU does the actual AI computation in between. A data centre running a large AI model might have 1 CPU for every 8 to 16 GPUs.

Source: NVIDIA Annual Report 2024; MIT Technology Review "The Chip That Changed Everything" 2023; Goldman Sachs Global AI Infrastructure Report 2024

Parameters are the numbers an AI model learns during training. They are the stored knowledge — billions or trillions of numerical values that encode every pattern, relationship, and fact the model absorbed from processing vast amounts of text, code, images, and data. When you ask an AI a question, those parameters are what determine the answer.

Think of a neural network as an enormous web of connections — similar in concept (though not in biology) to the neurons in a brain. Each connection has a weight, which is a parameter. During training, those weights are adjusted millions of times until the model reliably produces good outputs. When training is finished, the weights are frozen — and that frozen set of numbers is the model.

Early AI models had millions of parameters. By 2020, GPT-3 (the model that first demonstrated convincingly human-like language) had 175 billion. Today's frontier models are estimated to have hundreds of billions to several trillion parameters — though most AI companies do not publicly disclose exact figures.

To store and run a model with one trillion parameters, you need roughly 2 terabytes of memory just to hold the numbers — far exceeding what any single GPU can hold. This is why large models are split across dozens or hundreds of GPUs working in concert, and why the computing infrastructure for frontier AI costs billions of dollars to build.

Source: OpenAI GPT-3 Technical Report (Brown et al., 2020); Google DeepMind Model Architecture Papers; Anthropic Model Card documentation; Kaplan et al., "Scaling Laws for Neural Language Models" (OpenAI, 2020)

AI models do not read text the way humans do — word by word. Instead, they break language into small chunks called tokens. A token is typically a word fragment, a common word, or a punctuation mark. In English, one token is roughly 4 characters, or about three-quarters of a word.

For example, the sentence "AI is transforming the workforce" becomes approximately 7 tokens: AI /  is /  transform / ing /  the /  work / force. Notice that "transforming" and "workforce" are each split — because the model recognises those parts separately.

Tokens matter for several practical reasons:

• Context window: An AI model can only "see" a limited number of tokens at once. Early models had 2,000–4,000 tokens (about 3 pages of text). Modern models can handle 100,000 to over 1 million tokens — the equivalent of entire books or extensive document sets. This is why newer AI can summarise a 300-page report in seconds.
• Speed and cost: AI companies charge by the token. Processing 1 million tokens costs between a few cents and a few dollars depending on the model. When a business deploys AI to handle thousands of customer queries simultaneously, token efficiency directly affects whether automating jobs is financially attractive.
• Quality: More tokens processed = richer context = better answers. A model given 100 tokens of context will give a worse answer than one given 10,000 tokens.

For AI displacement, the token context window is critical. Early AI could only process short snippets — enough to answer a simple question, not enough to read a contract or analyse a financial report. As context windows have grown from thousands to millions of tokens, AI has crossed the threshold needed to automate the kinds of knowledge-work roles previously considered safe from automation.

Source: Anthropic Claude Technical Documentation; OpenAI Tokenization Guide; Vaswani et al., "Attention Is All You Need" (Google Brain, 2017); AI and Compute, OpenAI Research 2018

This is the most important question — because the answer explains why AI displacement is happening now, at this speed, at this scale, when previous automation waves were manageable.

In 2020, researchers at OpenAI published a paper on what they called scaling laws. Their discovery: AI capability does not improve gradually as models get bigger. Instead, it improves in jumps — and beyond certain scale thresholds, models suddenly develop abilities they did not have at all at smaller sizes. These are called emergent capabilities.

Here is what scale means in concrete terms:

• More GPU cores → faster training → models trained in weeks instead of years → AI capabilities advance faster → the pace of displacement accelerates.
• More parameters → the model stores more knowledge and handles more complex reasoning. A 7-billion-parameter model can answer simple questions. A 700-billion-parameter model can draft legal submissions, write financial analysis, or diagnose medical images — tasks that previously required years of professional training.
• Larger token context windows → the model can read entire documents, case histories, financial reports, or code repositories at once → tasks that required a human professional's sustained attention over hours or days can be completed in seconds.

The critical insight for understanding AI displacement: there is no stable floor. In previous automation waves, machines replaced physical tasks while cognitive tasks remained safe. In the current wave, cognitive capability scales with compute — and compute is doubling roughly every 12 months. There is no cognitive refuge that remains permanently out of reach as scale increases.

This is why COAD is designed as a 14-year program rather than a short-term adjustment measure. The displacement is not a one-time event. It is an ongoing process driven by continued scaling — and the income support needs to match that duration.

Source: Kaplan et al., "Scaling Laws for Neural Language Models" (OpenAI 2020); Wei et al., "Emergent Abilities of Large Language Models" (Google Brain, 2022); Stanford HAI — AI Index Report 2025; IMF World Economic Outlook 2024 (Chapter 3: AI and the Labour Market)