"Economic policy for AGI" is an essay of roughly 5,300 words published on 16 September 2026 at the launch of the DeepMind Institute, by Julian Jacobs, an economist and research scientist at Google DeepMind, and Alex Imas, the company's Director of AGI Economics and Professor of Economics and Applied AI at the University of Chicago. It evaluates eleven household-facing economic policies against a four-dimension rubric and argues for sequencing them against observable economic thresholds rather than choosing one in advance. The essay summarizes a longer paper of the same title posted to SSRN, which it says contains the full analysis, methodology and an evaluation of fourteen potential funding mechanisms for each policy. Google DeepMind's site disclaimer states that Institute pieces "are intended to serve as conversation starters, reflecting the author's ideas and research, and should not be read as Google's official view."
Argument
The essay's starting premise is uncertainty about which economic scenario an AGI transition produces. It notes that economists are "currently not seeing definitive evidence of systemic employment or wage impacts," citing an Economist piece of 4 September 2026, while allowing at the other extreme that "the broad and dexterous nature of AGI capabilities could mark a break with prior technologies, reducing demand for human labor in most cognitive occupations." From that uncertainty it draws a design conclusion: policy responses should be "flexible" and "tied to clear empirical triggers," so as to "avoid acting too late to cushion real disruption or overreacting prematurely with interventions that produce a cure worse than the disease."
Three obstacles are named. The first is data: platform usage metrics "cannot tell us how AI alters net employment," workplace field experiments "rarely generalize to the macroeconomy," and administrative datasets "may lag real-world disruption." The second is not knowing which policies would actually distribute gains while preserving agency. The third is the absence of a common rubric for comparing interventions.
The authors argue that work carries non-monetary value across the income distribution and that employment serves a role in social cohesion, and treat preservation of economic participation as a dimension policy should be scored on, rather than as a constraint outside the analysis.
Method
Eleven policies were assessed using a combination of manual literature review, a United States survey, and a panel of 51 AI agent raters. Each agent's persona was built from survey data on 51 real economists, "capturing a diversity of economic and political attitudes," using a method the essay attributes to John Horton and identifies as EDSL personas. The stated rationale is that the approach "removes some degrees of researcher freedom, allowing us to test how agents 'rate' particular interventions based on the available evidence and their own deliberative process." The essay describes the methodology as "an initial attempt" and "exploratory," and says further experiments and empirical evaluations are needed.
The four scoring dimensions are welfare and resilience (standards of living; meaning and human value; macroeconomic stabilisation); agency and voice (economic participation; ownership of gains; democratic voice); feasibility and efficiency (political support; economic feasibility; popular support; administrative capacity; speed; author-coded operational readiness); and durability across three transition phases (mild disruption; broad displacement; AGI transformation).
The eleven policies are grouped into targeted labor-market and wage interventions — active labour market policies and retraining, wage insurance, the Earned Income Tax Credit, a federal jobs guarantee, and unemployment insurance — and universal floors, services and structural assets — negative income tax, universal basic income, a sovereign AI dividend, universal basic capital, universal basic services, and industrial policy.
Scores
Composite scores from the simulated panel, on 0–100 scales, as published. The essay's own tier labels are high (≥70), moderate (55–69), mixed (roughly 40–54) and low (below 40–42, varying by table).
| Policy | Welfare & resilience | Agency & voice | Feasibility | Durability |
|---|---|---|---|---|
| Negative income tax | 69.8 | 53.7 | 50.2 | 69.6 |
| Earned Income Tax Credit | 68.8 | 47.2 | 79.8 | 51.0 |
| Unemployment insurance | 66.7 | 43.7 | 54.3 | 62.0 |
| Universal basic capital | 62.0 | 76.3 | 33.1 | 74.2 |
| Universal basic services | 60.7 | 41.9 | 34.5 | 75.0 |
| Sovereign AI fund / dividend | 53.7 | 55.2 | 38.4 | 63.5 |
| Universal basic income | 51.2 | 53.1 | 35.5 | 57.7 |
| Directed industrial policy | 49.1 | 52.1 | 41.2 | 51.4 |
| Wage insurance | 47.4 | 44.7 | 48.2 | 35.9 |
| Federal jobs guarantee | 44.4 | 46.2 | 28.4 | 35.0 |
| Active labour-market policies | 36.9 | 37.0 | 48.0 | 23.0 |
The pattern the authors draw out is a trade-off: the interventions that score highest on feasibility (EITC at 79.8) score low on ownership of gains (20.6), while universal basic capital, which leads on ownership of gains (94.9) and on durability under full AGI transformation (93.5), scores lowest but one on feasibility (33.1) and lowest of all on operational readiness (30.0). Active labour-market policies score 5.9 on durability under AGI transformation.
Survey figures reported for the United States: 85% support publicly funded worker retraining, the most popular of the interventions; 72% support unemployment insurance; 54% favour a universal basic capital program.
The three trigger-based interventions
The essay proposes what it calls three "least-regret" interventions, each matched to a scenario and to the data that would trigger it.
Scenario 1, mild disruption — "wage polarization and shifting demand for skills throughout the economy, without a collapse in aggregate labor demand." The recommended package is expanded unemployment insurance, an expanded EITC, and employer-led retraining. On unemployment insurance the essay argues the US system "provides replacement but at a lower wage, expires quickly, and typically excludes non-traditional workers, such as gig workers and freelancers," and cites Denmark's flexicurity model, under which recipients receive up to 90% of prior salary to a maximum of $3,200 for up to two years. On retraining it states that public programs "have historically been disappointing" while "apprenticeships and employer-led models show the highest incidence of success by a considerable margin." On the EITC it proposes extending the credit to workers without dependents and to a broader band of low- and middle-income earners, and paying it in regular instalments rather than annually.
Scenario 2, moderate displacement and wage compression — triggered by "prolonged unemployment spells and falling median wages that outpace job reinstatement." The recommendation is converting the expanded EITC into a negative income tax. The essay argues against universal basic income here on four grounds: inflationary pressure; cost and bluntness as an instrument that "may fail to concentrate sufficient relief where it is needed most"; failure to alter the structural forces shaping participation; and leaving recipients "no direct stake in the automated parts of the economy." It argues a negative income tax is preferable to wage insurance because it conditions on income level rather than cause of loss, avoiding the problem of "untangling whether a worker was displaced specifically by AGI, or by standard market churn," and cheaper than UBI because it pays only below a threshold.
Scenario 3, structural labor-capital decoupling — "a hypothetical and extreme scenario" in which "the marginal productivity of human labor could approach zero in many sectors," triggered by macroeconomic data showing "a sustained decline in labor's share of GDP alongside increasing capital returns." The recommendation is universal basic capital: an ex-ante equity or asset endowment, such as shares in a publicly managed fund invested in equities and bonds, which the essay compares to proposals to invest Social Security and other trust funds in US equity markets and to past bipartisan proposals for personal Social Security accounts. It is preferred over a sovereign wealth fund on the ground that AGI productivity gains "should appear throughout the economy, not just in the AI sector," making AI "fundamentally different from oil, which is a discrete, identifiable income stream"; the Alaska Permanent Fund's $1,000 dividend in 2025 is given as the sovereign-fund example. The authors state it "may be premature to deploy UBC as a reaction to advanced AI systems right now" and that the near-term priority is designing "the institutional architecture — including governance, funding methods, and distributional mechanics — so that the policy could be ready to implement at the right moment."
The sequencing claim is that the three are compatible: expanded EITC and UI "could — both politically and economically — pave the way for a Negative Income Tax if displacement deepens," and an NIT "could soon after function as a universal income floor as earned wages declined," with a pre-designed UBC program paired alongside.
Stated limits
The authors describe the exercise as exploratory and call for "experiments, empirical evaluations, and research studying the global economic developmental impacts of AGI" to validate the options. They identify improved data collection on AI's economic impacts as a precondition for the trigger-based design to work, since the triggers are defined against macroeconomic indicators the essay says are not currently measured at the needed granularity or timeliness. The scores are the output of simulated economist personas rather than of the economists themselves, a distinction the essay makes in describing the method but does not treat as a limitation on the scores' interpretation.
Provenance
Fetched live on 17 September 2026 from institute.deepmind.com, the Institute's own site, served under Google's privacy and terms policies with assets on a Google Cloud Storage bucket, and converted from HTML. The launch and this essay's authorship are corroborated by Axios, "Google, DeepMind launch institute to explore AGI" (Maria Curi, 16 September 2026), and The Next Web, "Google DeepMind launches the DeepMind Institute to debate AGI" (Ana Maria Constantin, 16 September 2026), which lists the inaugural pieces with their authors; The Next Web independently reports the three least-regret interventions and the 85%, 72% and 54% survey figures. The page is dated to the public launch of 16 September 2026; its article:published_time metadata reads 9 July 2026, a content-management date earlier than the launch, and article:modified_time reads 16 September 2026. A Firecrawl copy fetched the same day differs in a few standfirst and spelling details, indicating the page was edited around launch; the capture follows the live HTML. The longer SSRN paper the essay summarizes was not retrieved.
Relationships
- related: Principles for a new utopianism (Cave, DeepMind Institute, 2026) — a companion essay from the same launch.
- related: Introducing the DeepMind Institute (Legg, Manyika and Hassabis, September 2026) — the Institute's own launch statement.
- depends-on: DeepMind Institute, Google DeepMind — publisher and the authors' employer.
- supports: AI Labor Disruption — supplies a scored comparison of eleven policy responses.
- contradicts: Universal Basic Income (in AI policy) — argues UBI is inferior to a negative income tax on cost, targeting and agency, and to universal basic capital on ownership of gains.
- related: AI Displacement vs. Augmentation, Middle-Manager Displacement, Labor Disruption Timelines: Who Predicts What and Why — the disruption scenarios the triggers are defined against.
- related: Artificial General Intelligence (AGI), AGI Timelines — the transition the policies are designed for.