Wealth

AI could wipe out the UK’s income tax base: Here’s what replaces it

Ryan Brothwell 4 min read
AI could wipe out the UK’s income tax base: Here’s what replaces it

Key Points

  • A new IPPR paper sets out three tax options if AI shrinks the UK's income tax base.
  • A progressive consumption tax would pair higher VAT with a universal basic income.
  • An AI token tax would work like VAT on the output large language models produce.
  • A token tax needs international agreement, as frontier models sit in the US and China.
  • An unearned rents tax on AI capital applies only if AI wealth concentrates severely.

A new paper sets out three taxes Britain may need if AI removes enough jobs to shrink income tax revenue.

The Institute for Public Policy Research published the analysis, written by Ben Ansell, professor of comparative democratic institutions at the University of Oxford, who argues that large language models raise real questions about whether many jobs still exist in a decade.

The fiscal problem runs in two directions at once. Income tax revenue falls as employment falls, while demand for state support rises among the people made redundant.

The losses would come partly from people who had been substantial contributors, including graduate and service sector professionals.

A progressive consumption tax

The first option shifts taxation away from income and towards spending, an approach Ansell attributes to AI economist Anton Korinek.

Consumption taxes usually count as regressive, because poorer households spend a larger share of their income, though the UK already zero-rates food, children’s clothes and public transport under VAT.

The logic changes if wages fall as a share of national income while returns to capital rise, because much of that capital income gets spent on consumption. Higher VAT combined with lower income tax then moves some of the burden onto wealthier citizens.

Protecting poorer households requires a tax-free consumption allowance, working like the personal allowance does for income tax. Rather than asking vendors to check customer income, the paper suggests delivering it as a negative income tax or a universal basic income.

An AI token tax

The second option taxes the output of AI models rather than the companies running them.

Economists have debated robot taxes for a decade, charging firms for labour-replacing machines, sometimes at the equivalent of the national insurance that employing a displaced worker would have cost.

Two problems undermine that approach. Companies have every incentive to misrepresent how much labour-replacing technology they use, which gets harder to police as invisible AI agents replace physical robots, and making capital costlier than labour tends to reduce productivity if competitor economies decline to follow.

A token tax targets the tokens models produce instead, operating much like VAT on token production. Businesses and consumers using AI pay a proportionate charge, and exempting business-to-business intermediate sales lets the tax capture a share of the value the models create.

A floating point tax on compute itself offers an alternative, though revenues from it would fall as processing efficiency improves.

The token tax carries its own obstacles. It requires AI companies to permit auditing of token production, and a UK-only version would leave British AI users paying charges their foreign rivals avoid.

Frontier models sit largely in the US and China, so Britain would be asking foreign companies to audit continuously for British fiscal purposes. Ansell concludes that the tax would need agreement among like-minded countries.

An AI unearned rents tax

The third option applies in the most extreme scenario, where AI wealth concentrates severely.

Korinek’s position is that a sufficiently concentrated AI sector justifies shifting taxation entirely onto capital in the form of unearned rents, harvesting some of the returns AI produces.

Economists normally treat capital taxation as a deterrent to investment. That objection weakens if machines make the allocation decisions and the proceeds flow either to a small group of AI owners or straight back into further AI rollout, leaving direct taxation as the route to sustaining consumption for everyone else.

Ansell draws the parallel with the late 19th century, when Georgist arguments targeted unearned rents on land. The control at stake now is compute rather than land.

None of the three amounts to government policy, and the paper presents them as preparation for a scenario rather than measures for the current parliament.

Now read: Good news for hybrid workers: Britain’s return-to-office push has stalled