Preparing for AI’s economic impact: exploring policy responses

economic policylabor marketstaxationworkforce trainingpolicy

A short Anthropic policy post that sorts nine economic policy ideas, gathered from economists and policy researchers, by how fast and how severe AI's effects on work might be, from workforce training grants and permitting reform to token taxes, sovereign wealth funds and a business wealth tax; Anthropic states that the ideas are starting points for study, not its own positions.

Anthropic · 2025-10-14 · original

In October 2025 Anthropic published a short post asking what tools governments could use if AI changes the structure of the economy. The occasion it gives is a trend in its own usage data: since the launch of the Anthropic Economic Index, users have been "increasingly likely to delegate full tasks to Claude", a finding the post links to the September 2025 report summarized in Anthropic Economic Index: Tracking AI’s role in the US and global economy. The post states uncertainty about the scale of the change and presents its ideas as material for further study rather than recommendations.

The post contains no data analysis of its own. Its value for this library is as a record of which policy ideas Anthropic chose to publicize at the time, how it framed them, and where it said its own interests were affected.

How the post is organized

The ideas came from work over the preceding year with economists and policy experts, including members of Anthropic's Economic Advisory Council and participants in its first Economic Futures Symposium; the post says it sought both non-partisan thinkers and people "from across the political spectrum" (introduction). It groups nine categories of ideas under three scenarios that differ in how fast and how deeply AI affects the labor market ("Matching policies to scenarios"):

  • Nearly all scenarios, including ones where labor market effects stay modest. These ideas, the post notes, have often been proposed in other contexts before.
  • Moderate acceleration, with measurable wage declines and job losses for large parts of the workforce, where more substantial fiscal support for displaced workers might be needed.
  • Faster-moving scenarios, with possibly dramatic job losses and worsening inequality, calling for more ambitious proposals.

The post says the proposals "don’t necessarily represent Anthropic's own policy positions" (end of "Matching policies to scenarios") and repeats the point in its closing note.

The nine ideas

Policies for nearly all scenarios.

  1. Workforce training grants. Abigail Ball and Oren Cass of American Compass propose public subsidies to employers who create formal trainee positions with structured training, suggested at $10,000 per year in the US. American Compass would fund it by redirecting higher-education subsidies; the post adds that taxes on AI consumption are another possible source.
  2. Tax incentives for retention and retraining. Revana Sharfuddin of the Mercatus Center argues that the US tax code favors physical over human capital, since AI systems can be expensed immediately while training costs face restrictions. Her proposals include removing the $5,250 cap on tax-free educational assistance and extending full, immediate expensing to job-related training.
  3. Closing corporate tax loopholes. David Gamage proposes closing the "partnership gap" that lets large businesses avoid entity-level taxes, and allocating business taxes by customer location with worldwide combined reporting, to limit profit shifting that AI might make more common.
  4. Faster permits for AI infrastructure. This is the one idea the post presents as a position Anthropic has "consistently advocated". It names three sets of US processes that delay data centers and power: land use and environmental permits, state reviews that can make transmission buildouts last "10 years or more", and grid interconnection approvals that typically take "4-6 years" for generation. Suggested steps include reforms to the National Environmental Policy Act and advance analyses of facility types such as data centers.

Policies for moderate scenarios.

  1. Adjustment assistance for AI displacement. Modeled on the US Trade Adjustment Assistance program. Suchet Mittal and Sam Manning outline an Automation Adjustment Assistance program funded at about the level of Trade Adjustment Assistance, "approximately $700 million annually", that could grow or shrink with the pace of displacement, possibly funded by taxes on AI-driven revenue of very large firms.
  2. Taxes on compute or token generation. Lee Lockwood and Anton Korinek propose studying taxes on tokens, robots and digital services. In their argument, a tax on tokens sold to end users suits a stage where humans remain the dominant consumers; if AI systems themselves become major consumers of resources, taxing compute and hardware might be more effective, despite distorting investment.

Policies for fast-moving scenarios.

  1. Sovereign wealth funds with stakes in AI, so that citizens and governments share in AI's returns, with a related UK proposal for an "AI Bond" from the Centre for British Progress.
  2. Value-added taxes. The post notes that six of the G7 and 37 of 38 OECD countries have national value-added taxes, the United States being the exception, and argues that a shift toward taxing consumption could become necessary if labor's share of income falls.
  3. New revenue structures, such as Gamage's "low-rate business wealth tax" alongside income taxes, which he compares to an asset manager's management fee paired with a performance fee.

Statements about Anthropic's own interests

Two passages bear on Anthropic's position. On token and compute taxes the post says Anthropic believes they "deserve serious study, even though they would directly impact Anthropic's revenue and profitability" (section 6). On infrastructure, by contrast, the permitting proposal is the company's own long-standing advocacy and is the only one framed that way. The closing section mentions a $10 million commitment to scale up the Economic Futures Program, announced earlier that autumn, and symposia in Washington, DC (September) and London (November).

Limitations and open questions

The post makes no forecasts, offers no estimates of the cost or effect of any proposal, and does not say how the nine categories were selected from the ideas Anthropic received. Each proposal is summarized in a paragraph or two; readers need the linked sources for the arguments. The three scenarios are described qualitatively, with no thresholds that would indicate which one is unfolding.

Research questions

  1. Which indicators would show which scenario is unfolding? The post ties each policy group to a scenario but gives no measurable markers. Comparing the scenario descriptions with labor market measures, such as wage and employment changes in occupations with high AI usage, would show whether the scenarios can be told apart in practice.
  2. How do Anthropic's later usage reports bear on the delegation trend cited here? The post's starting point is a rise in full-task delegation. Later Economic Index reports, such as Anthropic Economic Index report: Economic primitives (January 2026), would show whether the trend continued.
  3. How large would a token tax be relative to AI company revenue? The post concedes such taxes would affect Anthropic's revenue but gives no scale. Any rates proposed in the Lockwood and Korinek work, applied to public revenue figures, would answer it.
  4. Which of the nine ideas has Anthropic later endorsed as its own position? Only permitting reform is presented as Anthropic's advocacy. Comparing later Anthropic policy submissions with this list would show whether any other idea moved from "worth studying" to position.
  5. How was the set of nine chosen? The post says it received a broad range of proposals. A list of all proposals received, or the symposium programs, would show what was left out.

How it connects

References

  • Anthropic. (2025, October 14). Preparing for AI’s economic impact: exploring policy responses. https://www.anthropic.com/research/economic-policy-responses
  • The proposals by American Compass, the Mercatus Center, Gamage, Mittal and Manning, Lockwood and Korinek, and the Centre for British Progress are cited in the post and were not read for this note.