Raghuram Rajan Wants To Tax The AI Tokens That Replace Workers

Raghuram Rajan, the former Reserve Bank of India governor who now teaches at the University of Chicago's Booth School of Business, wants to tax the small units of text and data that AI models process. The pitch is to fix a distortion in the tax code that pushes firms toward automation.

In a Project Syndicate commentary published on August 14, titled "How Corporations Can Mitigate an AI Jobocalypse", Rajan argues that companies pay payroll taxes, Social Security contributions, and other levies on human workers but face no equivalent charge when they swap a worker for an AI system. That asymmetry, he writes, tilts the math toward replacing people even when the broader social cost falls on the displaced worker.

Rajan, who led India's central bank from 2013 to 2016 and is one of the most cited central-bank voices on financial crises since, is pitching a small, ramping tax on the AI tokens a company consumes. The rate would start low and rise as adoption grows. Revenue would, in spirit, be earmarked to cushion the workers AI displaces: retraining, placement, and a buffer while they find the next role.

The proposal is best read as a Pigouvian nudge: a charge designed to push a private decision (how much AI to use) toward a social optimum (how much AI should be used once the cost of displaced workers is counted). It is the same logic that put taxes on cigarettes and carbon.

Project Syndicate is a global opinion outlet for policy commentary, and his piece is a proposal, not a government plan. The Indian press cluster, LiveMint, trak.in, NDTV Profit, and the Economic Times, has been re-reporting the op-ed as a policy idea worth taking seriously, not as imminent legislation.

The asymmetry Rajan points to is concrete. In the US, employers pay 6.2% in Social Security tax on each worker's wage up to the cap, plus Medicare, federal unemployment, and matching state charges. AI vendors charge per token, the small chunks of text or data an AI model reads, writes, or transforms, and the buyer pays sales tax or VAT the way they would on any software service. Nothing in the standard tax stack targets the act of replacing a worker with a model.

A firm that doubles its AI token use pays twice as much into the displacement fund. A firm that hires a worker instead pays nothing extra, because the payroll system is already charging the employer for that choice. Over time, as more firms build AI into core workflows, the rate climbs and the fund grows.

The numbers he leans on suggest AI adoption is still patchy, which is part of his case for starting the levy low. A recent US Census Bureau Business Trends and Outlook Survey, cited in his piece, shows that only 20% of firms with fewer than 20 employees currently use AI, compared to 37% of businesses with at least 250 employees. Big companies are pulling ahead; small ones are not yet there. A small initial rate keeps the small firms from being priced out of useful productivity tools while the displacement fund accumulates where the displacement is actually happening.

He leans on two well-worn counterweights to the AI-takes-all-the-jobs story. The first is the Jevons effect: when a resource gets cheaper, more of it gets used, and the demand for the thing it complements rises. Cheaper AI, on this view, expands the market for the people who build, deploy, and maintain it. The second is the work of MIT economist David Autor, who has spent two decades showing that new tools tend to augment moderately skilled workers, from nurse practitioners with diagnostic AI to accountants with language models to paralegals with document review, rather than erase them. Rajan also notes that brand-new roles, from prompt engineers to AI safety reviewers, are already being hired at scale.

Because, in Rajan's read, even a milder version of the doomer scenario still leaves the same fiscal problem. If AI replaces a tenth of the workers it could replace, the displaced workers still need retraining, income support, and time. The tax is a way to fund that without waiting for Congress, Parliament, or a multilateral body to agree on a broader AI policy.

Where is the line between taxing AI tokens and taxing any other piece of software? If a company uses open-weight models in-house, who measures the tokens? Would a US state, an EU member, or a coalition of adopters move first, and would the revenue stay with the jurisdiction that collected it? Rajan acknowledges these as open problems in the same piece. He is naming a mechanism, not closing a legislative file.

The proposal moves the AI-policy conversation from whether the models should be safer and more capable to who pays when the cheaper model replaces a worker, and through what lever. That is a question the existing tax code already has an answer to, and Rajan's bet is that the answer is currently tilted the wrong way.

The next milestone is whether any government picks the idea up. India's finance ministry has not commented. The US Treasury has not commented. The European Commission's AI Office, which has been the most active regulator on model rules, has not commented. The proposal is, for now, exactly what Rajan says it is: a starting point for a conversation the wire has been avoiding.