Google DeepMind Institute: Economic policy for AGI

AI economic policy is often reduced to a false choice: trust markets to absorb the disruption, or replace lost wages with universal basic income. A new DeepMind Institute study argues that both positions overlook a more difficult problem. The right intervention depends on how deeply AI changes the relationship between labor, income, and capital.

“Economic Policy for AGI,” by Google DeepMind economists Julian Jacobs and Alex Imas, evaluates 11 policies across welfare, human agency, implementation feasibility, and durability under three increasingly disruptive scenarios. Its most useful contribution isn’t a single preferred benefit program. It is a state-contingent roadmap in which policy expands or changes when economic data crosses predetermined thresholds.

The DeepMind Institute Puts Economics Inside the AGI Debate

Researchers from Google and Google DeepMind launched the DeepMind Institute on September 16, 2026, as a forum for research and debate about artificial general intelligence. Its founding statement places economic institutions alongside safety, governance, philosophy, and questions about human purpose.

That scope matters because the AGI debate has largely concentrated on model capabilities, technical safety, and regulation. Those remain essential, but even well-controlled AGI could produce destabilizing outcomes if productivity gains accrue to a small group while wages, employment, or bargaining power deteriorate elsewhere.

The institute’s proximity to a frontier AI lab gives it access to relevant technical and economic expertise. It does not, however, turn its essays into corporate policy or democratic consensus. The institute explicitly says its publications represent their authors’ research and ideas, not Google’s official view. That distinction is especially important when the subject is redistribution, taxation, and public ownership.

Alex Imas introduced the project as an attempt to move beyond the familiar market-versus-UBI argument:

The Study Rejects a One-Policy Answer

Jacobs and Imas evaluate five targeted labor-market and wage policies:

• Active labor-market programs and retraining
• Wage insurance
• An expanded Earned Income Tax Credit
• A federal jobs guarantee
• Unemployment insurance

They compare those with six broader structural policies:

• A negative income tax
• Universal basic income
• A sovereign AI fund or dividend
• Universal basic capital
• Universal basic services
• Industrial policy

The framework judges each option by four criteria. Welfare covers living standards, psychological meaning, and macroeconomic stability. Agency considers economic participation, ownership of AI-generated gains, and democratic voice. Feasibility includes costs, public and political support, administrative capacity, and deployment speed. Durability asks whether the policy would remain useful as disruption moves from familiar automation to an economy in which human labor is no longer central.

According to the authors, the underlying research combines economic literature, a representative survey of roughly 2,000 Americans, and an experimental panel of 51 AI personas constructed from survey data about 51 real economists. The agent analysis was conducted using Expected Parrot’s EDSL, an open-source system for running surveys and experiments with language-model agents.

This mixed methodology produces a broad comparison, but the components answer different questions. The human survey measures public preferences. Literature reviews summarize existing evidence. The AI personas generate structured ratings under the researchers’ scenarios and rubric. Those ratings should not be presented as 51 economists personally endorsing the results.

The distinction is easy to lose because the tables assign precise scores to uncertain, hypothetical outcomes. A universal basic capital policy, for example, receives an agency score of 76.3, while unemployment insurance scores 43.7. Those numbers organize the assumptions and deliberations built into the experiment. They do not establish that UBC will improve agency by a measurable real-world amount.

DeepMind Proposes a Three-Stage Policy Sequence

Jacobs and Imas identify what they call “least-regret” interventions: policies that appear reasonable under current conditions but can lead into stronger systems if disruption intensifies.

Mild disruption calls for stronger existing institutions

In the first scenario, AGI behaves more like earlier waves of automation. It changes tasks, raises productivity, compresses some wages, and shifts demand between occupations without causing an economy-wide collapse in employment.

The recommended response combines expanded unemployment insurance, a modernized Earned Income Tax Credit, and employer-led retraining. These policies already have institutional foundations, making them less risky to deploy before the scale of the AI shock is clear.

Unemployment insurance would protect workers during transitions, although the authors argue that US programs may need wider coverage, longer durations, and better support for freelancers and gig workers. An expanded EITC would supplement low and moderate wages, including for workers without dependent children. Regular payments could also make the credit more useful during periods of volatility than an annual tax refund.

The retraining proposal is narrower than a generic call for more education. The essay favors employer-led programs and apprenticeships because training connected to real vacancies and workplace demand has generally performed better than classroom programs designed without strong employer involvement.

Broad displacement shifts support from wages to income

The second scenario assumes that automation begins outpacing job creation. Unemployment spells lengthen, median wages fall, and some workers remain underemployed even as the AI economy grows.

Here, the framework would gradually convert the EITC into a negative income tax. The distinction is important. The EITC primarily supports people with earned income, while a negative income tax creates an income floor for anyone falling below a specified threshold. Benefits decline as earnings rise, targeting support more tightly than a universal payment.

This transition also avoids requiring agencies to prove that AI caused a particular layoff. If a person’s income falls below the threshold, support activates regardless of whether the immediate cause was automation, a recession, corporate restructuring, or ordinary market churn.

Full transformation requires ownership, not only transfers

The third scenario is deliberately extreme. Human labor loses much of its economic value, capital captures a growing share of output, and traditional job creation no longer restores broad participation.

The proposed backstop is universal basic capital. Rather than sending everyone a recurring cash payment, UBC would give citizens an ownership stake through mechanisms such as publicly managed investment accounts or shares in a diversified social wealth fund. People would then receive returns as the underlying capital appreciates.

This explains why UBC performs particularly well in the study’s agency analysis. Its simulated score for ownership of gains is 94.9, and its durability score under full AGI transformation is 93.5. The policy is intended to address a structural problem that income transfers alone do not solve: who owns the productive assets in an economy where machines perform most economically valuable work.

The authors do not recommend launching UBC immediately. They argue for designing its governance, funding, and distribution mechanisms now, then activating it only if the data indicates sustained labor-capital decoupling.

Public Support Favors Policies Built for Familiar Shocks

The study’s sharpest finding is the tension between public support and resilience under more transformative scenarios.

Publicly funded retraining received support from 85% of surveyed Americans, while unemployment insurance received 72%. Yet the agent panel gave active labor-market policies a durability score of only 5.9 under full transformation. Retraining cannot restore employment if the economy stops creating enough economically competitive human jobs.

Universal basic capital presents the reverse problem. It received support from 54% of respondents and scored only 33.1 for implementation feasibility. Under full transformation, however, it was among the strongest options because it would distribute capital ownership rather than depend on employment.

The results also complicate the case for UBI. The authors regard universal payments as expensive and poorly targeted in moderate scenarios. In a more radical transformation, UBI could protect consumption but leave recipients without ownership, workplace influence, or a claim on AI-driven capital growth.

That does not settle the UBI debate. Universal basic services recorded the highest composite durability score, while a negative income tax led the welfare rankings and the EITC led feasibility. The tables reinforce the study’s main argument: changing the evaluation criterion changes the apparent winner.

Better Economic Data Must Determine When Policies Change

A trigger-based system only works if governments can detect the transition between scenarios. AI adoption statistics alone are insufficient. A company can deploy AI without eliminating jobs, while another can restructure its workforce before new technology appears in standard productivity measures.

The US already has pieces of the required infrastructure. The Census Bureau’s Business Trends and Outlook Survey provides nationally representative, biweekly data about business AI use. The Bureau of Labor Statistics publishes employment, wage, occupation, and unemployment-duration data. The Federal Reserve’s Distributional Financial Accounts track household wealth quarterly across different population groups.

An operational AGI policy dashboard would need to connect and extend these sources. Useful indicators could include:

• AI adoption by business function, occupation, and task
• Layoff and hiring rates in highly exposed sectors
• Wage growth and unemployment duration by occupation
• Movement into lower-paid or lower-productivity work
• The pace at which displaced workers find comparable jobs
• Labor’s share of national income
• Capital returns and the distribution of equity ownership
• Changes in output, prices, working hours, and household consumption

The trigger for a negative income tax should not be a new model release or a prediction about when AGI will arrive. It should be sustained evidence that unemployment and wage losses are outrunning labor-market recovery. Likewise, UBC would require stronger evidence that economic growth was becoming structurally detached from human work.

The Agent Panel Is a Starting Point, Not Expert Consensus

The use of economist-inspired AI personas is methodologically interesting. John Horton’s updated “Homo Silicus” research argues that language models can serve as configurable economic agents, allowing researchers to explore behavior under different preferences and scenarios. Such simulations can expose trade-offs, test alternative framings, and make assumptions easier to compare.

They cannot resolve the unprecedented uncertainty surrounding AGI. Results may depend on model choice, prompts, persona construction, scoring rules, source material, and how the three scenarios are described. Public support also says little about fiscal sustainability or long-term macroeconomic effects.

The DeepMind essay acknowledges these limits, describing the framework as an initial taxonomy that needs experiments, empirical evaluation, and research beyond the United States. Its policy rankings are best treated as hypotheses for further study, not forecasts or instructions ready for legislation.

Final Thoughts

The most defensible part of DeepMind’s economic policy framework is its refusal to legislate today for one confident AGI prediction. Expanded unemployment insurance and wage support are sensible if AI produces familiar disruption. A negative income tax becomes more relevant if employment recovery weakens. Capital ownership enters the picture only if labor genuinely stops providing a dependable share of national income.

The hard work is therefore not selecting a fashionable acronym. Governments need to design escalation paths, specify triggers, and build measurement systems before a crisis forces decisions using delayed or incomplete evidence. If AGI’s economic effects arrive quickly, the difference between preparation and improvisation may matter as much as the policy ultimately chosen.