Preparing for AI’s economic impact: exploring policy responses
How will the arrival of powerful AI systems change the structure of the economy? We are uncertain, and so are external experts. But as AI systems continue to improve, and are adopted at an ever-larger scale, it’s crucial there is more discussion about the tools policymakers could use to respond to AI's economic impacts—whatever their nature. To help with this, we’re sharing several economic policy ideas that merit further study.
Since launching the Anthropic Economic Index, we've observed an important shift in AI use. Users are becoming increasingly likely to delegate full tasks to Claude, “collaborating” with Claude less. As AI models continue to work independently for longer periods of time, and as more employers adopt AI to improve their productivity, we expect this trend to accelerate. The implications for the workforce are uncertain.
How should policymakers respond? This is not an easy question, nor is it one that any single actor can answer. There is great uncertainty about the scale of the transition ahead, and a wide range of views about how to manage it. But it is imperative to begin formulating ideas now for the economic scenarios we might find ourselves in.
Over the past year, we’ve worked with economists and policy experts from around the world (including members of our Economic Advisory Council and participants in our first Economic Futures Symposium) to move this discussion forward. To generate a broad range of ideas, we’ve engaged with both non-partisan thinkers and those from across the political spectrum.
Below, we briefly explore nine of these categories of ideas, covering workforce development, permitting reform, fiscal policy, and social services.
While we don’t know what the optimal policies will prove to be, we’re committed to sharing ideas in the open, and to being transparent about the economic effects of advanced AI.
Matching policies to scenarios
The rate, scale, and form of AI's economic effects will determine the policy responses that are necessary across the world. Accordingly, we've organized these initial ideas into three broad categories:
Policy ideas for nearly all scenarios, including those where negative effects on the labor market remain modest. These are policies that their advocates argue merit consideration almost regardless of how significant the disruption of AI proves to be. Given this, many of these proposals have been suggested in other contexts before. They include upskilling workers and students for emerging jobs, and reforming permitting processes to enable the construction of energy and computing infrastructure to improve productivity.
Policy ideas for scenarios with moderate acceleration, where AI leads to measurable wage declines and job losses for large portions of the workforce. Here, more substantial fiscal support for displaced workers might be needed. To offset negative externalities imposed on displaced workers from rapid automation, taxes on automation might be considered in this scenario.
Policy ideas for faster-moving scenarios, potentially involving dramatic job losses and worsening inequality. These proposals are much more ambitious, and are designed to respond to a starkly different economic picture. So far, ideas include using sovereign wealth funds to give citizens stakes in AI revenues, and finding new ways to generate government revenue.
The proposals below don’t necessarily represent Anthropic's own policy positions. But we’re excited by the breadth of proposals we’ve received, and we hope they encourage further research and debate.
Policies for nearly all scenarios
1. Invest in upskilling through workforce training grants
At our DC Symposium, Abigail Ball, Executive Director of American Compass, presented the Workforce Training Grant—a proposal she developed with colleague Oren Cass to direct public resources toward on-the-job training.
Under this model, governments would provide substantial annual subsidies (Ball and Cass suggest $10,000 per year in the US) directly to employers who create formal trainee positions with structured training programs. This training could take multiple forms: programs operated by individual employers, by employer consortia or industry associations, through partnerships between employers and organized labor, or by technical schools and community colleges working alongside employers.
American Compass proposes redirecting existing higher education subsidies to fund this program. But a range of other funding mechanisms might also deserve consideration—including the possibility of using taxes on AI consumption to support workforce development initiatives.
2. Reform tax incentives for worker retention and retraining
Tax policy can, on the margin, incentivize employers to retrain and retain employees rather than reducing headcount.
Revana Sharfuddin of the Mercatus Center argues that the US tax code creates a bias favoring physical capital investment over human capital investment. Businesses can immediately expense AI systems through bonus depreciation, yet face numerous restrictions when deducting worker training costs. She proposes reforms to the Internal Revenue Code, including eliminating the $5,250 cap on tax-free educational assistance, and extending full and immediate expensing to all job-related training.
These changes would aim to reduce the cost of retraining relative to the cost of layoffs, helping those workers whose positions might otherwise be on the margins of workforce reduction decisions.
3. Close corporate tax loopholes
Tax policy expert David Gamage has outlined reforms designed to prevent AI transformation from straining government budgets. Several of his proposals involve closing the “partnership gap” that allows large businesses to avoid entity-level taxes, and modernizing tax allocation to combat profit shifting and better capture value from digital and intangibles-based business models.
The second reform would allocate business taxes based on customer locations through market-based apportionment, while requiring worldwide combined reporting to treat multinationals and subsidiaries as single entities. This approach is designed to limit artificial profit-shifting to tax havens—a practice that could become more prevalent as AI potentially further increases the economic importance of profits derived from intangibles.
Gamage argues that "governments that act first will solve their fiscal challenges and better position residents to thrive in an AI economy. Those that wait will face resource constraints when flexibility is most needed."
4. Accelerate permits and approvals for AI infrastructure
Anthropic has consistently advocated for reforming permitting and power procurement processes in the United States and allied nations. Accelerating these processes is needed to develop the infrastructure to train and deploy frontier AI—that is, large-scale data centers, transmission infrastructure, and power generation facilities. Reforms will also unlock investment, economic growth, and job creation in the places where AI is built. Failing to accelerate AI infrastructure development will slow productivity and job growth, and it could introduce national security risks from vital AI infrastructure moving offshore.
Three overlapping sets of U.S. regulatory processes delay building large-scale AI infrastructure for years. The first category is permits. These include a series of land use and environmental approvals at the federal, state, and local levels. Second, state regulatory reviews for transmission projects can cause buildouts of new lines to last 10 years or more. Finally, approvals to interconnect facilities to the electric grid typically take 4-6 years for generation resources.
Concrete steps to address these challenges include reforms to the National Environmental Policy Act (NEPA), which requires federal agencies to review many projects’ environmental effects. Advance analyses of certain kinds of facilities, such as data centers, could help speed reviews of future projects. Other reforms could include leveraging federal authorities to expedite critical transmission buildouts and upgrades and collaboration with utilities to identify opportunities for fast interconnections.
As Tyler Cowen, faculty director of the Mercatus Center and member of our Economic Advisory Council, notes: "I am all for permitting reform—the energy sector included."
Policy ideas for moderate scenarios
5. Establish trade adjustment assistance for AI displacement
Several economists are exploring how the Trade Adjustment Assistance (TAA) model – in which affected workers are given opportunities to obtain new skills, or receive other support – might be adapted for labor market disruptions in an era of powerful AI. Ioana Marinescu of the University of Pennsylvania, a member of our Economic Advisory Council, views TAA-like "AI insurance" as a mechanism "to support those who lose jobs due to AI."
Along these lines, Suchet Mittal and Sam Manning have outlined a potential Automation Adjustment Assistance (AAA) program. They describe how funding AAA at levels similar to TAA—approximately $700 million annually—could be an initial option, with mechanisms built in to increase or decrease the size of the program in line with the pace and scale of AI-driven displacement.
Mittal and Manning note that if such a program needed to expand in the future, it could potentially be funded through taxes on AI-driven revenues from firms above a certain high level of market capitalization, creating a direct mechanism for the AI sector to support workers displaced by the technology.
6. Implement taxes on compute or token generation
University of Virginia economists Lee Lockwood and Anton Korinek (a member of our Economic Advisory Council) propose studying of a range of taxes on “token generation, robots, robot services, and digital services.”
These taxes offer different potential benefits – and distortionary risks – depending on the stage of AI’s development within the economy. A tax on AI-generated tokens sold to end users (a “token tax”) might be desirable when humans remain dominant consumers in the economy, even if powerful AI reduces the relative economic role of labor.
Korinek and Lockwood argue that, if the economy reaches a stage where powerful AI systems become themselves major consumers of the economy’s resources, taxing AI resource accumulation–e.g. via taxes on compute and other hardware–might be more effective than token taxes on human end users. Although these taxes on computational resources distort investment along an AI-transformed economy’s trajectory, they could become the only remaining mechanism to capture some of the windfall generated by AI if the role of both labor markets and human consumption in the economy declines.
We believe taxes in this broader category deserve serious study, even though they would directly impact Anthropic's revenue and profitability. These taxes could provide crucial revenue for vital fiscal programs—including several others discussed in this post.
Policy ideas for fast-moving scenarios
7. Create national sovereign wealth funds with stakes in AI
A growing set of proposals aims to give citizens and governments greater stakes in AI's economic returns. Sovereign wealth funds could enable states to acquire positions in AI-related assets. In scenarios where the AI sector captures an outsized share of economic wealth, government investment could both shape the sector's behavior and "distribute AI-derived wealth more equitably."
Writing for the Centre for British Progress, Emma Casey, Emma Rockall, and Helena Roy have proposed a related concept for the United Kingdom: an AI Bond. The AI Bond would aim to ensure adequate investment in "the AI stack" to capture AI's benefits and then distribute its returns more evenly across Britain—even as AI research roles concentrate in a few cities, like London.
8. Adopt or modernize value-added taxes
Six out of the G7 countries have national value-added taxes (VATs), as do 37 out of 38 OECD countries. The United States is the exception.
As AI transforms the economy, labor's share of the production of value might decline significantly. A shift toward taxing consumption (as through a VAT) could become necessary to fund core government activities. VAT collection also provides governments with fine-grained information about the economic production network—which could be particularly valuable during this potential period of rapid technological and economic change.
"Value-added taxes are non-distortionary and to an extent, self-enforcing," notes John Horton of MIT's Sloan School of Management, a member of our Economic Advisory Council.
9. Implement new revenue structures to account for AI’s growing share of the economy
If AI is responsible for a large share of economic output (causing labor’s share to decline), governments might require new revenue streams to complement income tax. Another of David Gamage’s proposals is exploring a "low-rate business wealth tax" as a complement to income taxes. His reasoning: "Income taxes face accounting manipulation; wealth taxes face asset valuation challenges. Using both makes the system harder to avoid" for highly profitable enterprises.
Gamage analogizes this system to the fee structures that certain asset managers charge clients: "the wealth tax functions as a management fee for providing legal infrastructure protecting accumulated capital, while the income tax serves as a performance fee for profits generated in state markets." This idea represents one way that governments might adapt to changes in the value of human labor, although we think there are many more ideas to be explored in this area.
Continuing the conversation
Earlier this fall, Anthropic announced a $10 million commitment to scale up the Economic Futures Program. This investment will support rigorous empirical research on AI's economic impacts and policy ideas, as well as expand our Symposia series—beginning with an event in London this November, which follows our September event in DC.
None of the ideas outlined here represent definitive recommendations. They are starting points for deeper research, policy development, and public debate. The economic effects of AI remain uncertain in both timing and magnitude, and different scenarios will require different responses.
What's clear, though, is that proactive engagement between researchers, policymakers, and the AI industry is essential. By exploring these options now—before we know the shape of AI’s economic effects—we can better prepare for a range of possible futures, and ensure that workers and communities are well-placed to benefit from the full potential of AI.
Most of the policy ideas discussed in this post have emerged from proposals from or conversations with members of Anthropic's Economic Advisory Council, participants in our Economic Futures Symposia, and independent researchers. They do not all necessarily represent Anthropic's policy positions.