Tax the Tokens: Anthropic's Chief Economist Makes the Case for an AI Automation Levy

The most striking AI policy proposal of the season did not come from a legislator or a regulator. It came from inside one of the labs building the technology. At a Harvard Kennedy School forum moderated by economist Jason Furman, Anthropic’s chief economist Peter McCrory publicly endorsed a “token tax” — a levy on AI usage designed to curb excessive automation — and walked through a sweeping economic vision of what AI does to growth, wages, and the distribution of income between capital and labor by 2030.

The forum, held at Harvard’s Institute of Politics and covered by The Harvard Crimson on September 24, has been rippling through economic and policy coverage since, with fresh analyses still landing days later. Its core message is unusual in the AI debate: the chief economist of a frontier lab is making the case for taxing his own product.

The productivity puzzle — and Anthropic’s answer

McCrory opened with the contradiction defining the current macroeconomic moment: AI capabilities are compounding at a torrid pace — the range of tasks models can complete autonomously roughly doubles every four to seven months, he noted — yet total factor productivity has not surged, and US unemployment sits at a historically low 4.1%.

His explanation is diffusion lag compounded by a J-curve. Businesses must absorb steep restructuring costs — reorganizing data, dismantling internal firewalls, redesigning workflows around context the model can actually use — before any productivity dividend shows up in the statistics. Large enterprises in particular are wading through a costly experimentation phase that McCrory describes as the bottom of the J-curve for AI investment returns. The capability curve and the measured-output curve point in the same direction, but on different clocks.

Then he put a striking number on the table. By tracking the actual time users spend on tasks involving Claude, McCrory estimates that, based on current models and current usage patterns, labor productivity growth could rise by 1.8 percentage points. If those efficiency gains diffuse across the broader economy over the next decade, combined with capital deepening, US growth could revisit the high-growth era of the late 1990s and early 2000s.

For context, the Congressional Budget Office projects long-term US growth at 1.7%, and the Federal Reserve’s most optimistic scenario runs around 2.4%. Anthropic’s internal estimate — baseline plus 1.8 points — would tower over both. Furman pushed back with the classic bottleneck objection, citing research showing that after teams adopted coding agents like Claude Code, generated lines of code rose roughly twenty-fold while actual software release volume grew only about 30%. The unautomated “weak links,” he cautioned, can constrain aggregate expansion. McCrory’s response: even discounted for bottlenecks, a 1.8-point productivity boost is still large enough to force a re-evaluation of monetary policy and fiscal forecasting frameworks.

Labor loses 15 points in the tail scenario

On distribution, McCrory was blunt. AI “will most likely exacerbate inequality between capital and labor,” he said, and within labor itself the gap widens too. While randomized trials show AI can rapidly lift novices toward expert performance, the actual adoption data show the primary user base is high-wage, high-skill workers. Expertise acts as a multiplier, converting top talent into what he called hundred-fold engineers, rather than compressing the wage distribution from below.

The macro models are starker still. As AI-driven capital deepening accelerates, McCrory warned that in extreme scenarios labor’s share of income falls by roughly 15 percentage points from its historical 60% share of every dollar of income. The mirror image: capital owners — shareholders — capture returns from AI-driven growth far exceeding historical norms. The scenarios echo the modeling work Anthropic’s economics team published in September, which projected GDP outcomes ranging from a marginal 1.6% bump to a 32.4% expansion by 2030 depending on AI capability and adoption assumptions.

Three singularities, one bargaining-power experiment

Looking toward 2030, McCrory framed the endgame of the AI economy as three distinct “singularities”:

• The software singularity — AI systems making sustained progress at recursive self-improvement.
• The economic singularity — the question of whether the economy can achieve unbounded growth in finite time.
• The Coase singularity — named for economist Ronald Coase, whose theory of the firm was built on transaction costs. As AI agents negotiate increasingly complex transactions on behalf of users, those costs collapse, and firm boundaries and the structure of economic exchange fundamentally transform.

The Coase singularity is not purely theoretical for Anthropic. In an internal experiment the company calls the “Deal Project,” models negotiated against other models — and the results were unsettling. “More powerful models systematically extracted more rents when negotiating with weaker models,” McCrory reported, adding that people did not perceive the asymmetry in bargaining power. If agents become the standard interface for commerce, whoever controls the frontier model controls the pricing power.

Tax the tokens like carbon

The policy punchline followed from the diagnosis. Current tax structures in the US and most countries, McCrory observed, are heavily weighted toward taxing labor — payroll taxes, income taxes. If labor’s share of income declines, that base erodes, and governments will need alternative regimes, including heavier taxation of consumption.

And then the step beyond: a token tax. “In the economic policy framework, I view a token tax as an incentive mechanism to curb over-adoption,” he said. “A token tax can be analogized to a carbon tax or a tobacco tax.” The logic is Pigouvian: when an activity generates negative externalities — carbon emissions, tobacco consumption, or, in this framing, excessive automation that displaces workers faster than labor markets can absorb — a per-unit tax internalizes the social cost and slows adoption to a socially optimal rate.

McCrory cited research by UC Berkeley scholar Martin Béra showing that firms do not internalize the externalities of automation; in environments where displaced workers struggle to mitigate unemployment shocks, taxing automation carries both equity and efficiency justifications. The idea has intellectual company beyond academia — it rhymes with proposals floated across the political spectrum this year, and with the scenario modeling Anthropic itself published in September proposing study of taxes on “token generation, robots, robot services, and digital services.”

The confession of a lab economist

Two other moments from the forum deserve note, because they frame the whole enterprise.

First, on directed technical change: McCrory said his team is explicitly testing whether Anthropic can “operationalize” the economic concept of steering innovation toward complementing labor rather than displacing it. “I don’t actually know if it’s possible or not, but my team has an opportunity to see if it’s possible,” he said. It is a remarkable admission — a frontier lab attempting to aim its own research program using public-economics machinery.

Second, on humility. Furman closed by suggesting McCrory’s understanding of the subject “should rank in the top 10%,” and reminded the audience that honest forecasting requires the courage to admit error. McCrory, for his part, insisted that his team’s research is meant to “empower society to make these choices that we shouldn’t be the ones to unilaterally make.”

Whether or not a token tax ever reaches a legislative text, the fact that a frontier lab’s chief economist is designing the case for taxing his own tokens marks a shift in the AI policy debate — from whether AI has macroeconomic consequences, to who pays for them.

Sources

• The Harvard Crimson — “Anthropic Chief Economist Says Company Is Exploring Ways to Steer AI Toward Augmenting Workers” (Sept 24, 2026)
• BigGo Finance — “AI Labor Productivity Could Surge by 1.8 Percentage Points; Anthropic Economist Calls for Token Tax” (Sept 28, 2026)
• Anthropic — “Scenarios for our Economic Future”
• NPR — “Anthropic wants to test how AI could impact U.S. economy” (Sept 9, 2026)