Why not just tax the AI companies?

In February 2017, Bill Gates said the robot that takes your job should pay taxes. The same month, the European Parliament looked at exactly that idea and voted it down, 396 to 123. Nine years later the debate has not moved an inch: this August, three US House Democrats proposed taxing AI companies' revenues to fund public jobs, and the objections raised were the ones from 2017.

Nine years of stalemate usually means a question is badly posed. The instinct behind it is perfectly sound: these companies are building enormous wealth on public foundations, and the public should share in it. The problem is the instrument. Here is why the obvious tool keeps failing, and what actually grips.

What happened when countries tried

Taxing an asset's owners is not a new idea, and Europe ran the experiment at full scale. In 1990, twelve OECD countries levied a tax on net wealth. By 2017, eight had abandoned it. When France converted its version into a property levy in 2018, three remained. France's experience shows the mechanics: in a single year near the end, 780 households liable for the wealth tax moved abroad while 300 came back. One can argue about how much revenue that lost, and economists do. What nobody disputes is the direction: the tax stood still and the base walked away.

None of this means taxation is wrong or futile. It means a tax on something mobile is a bet that the something will stay put, and wealthy owners keep winning that bet.

The most portable asset in history

Now apply that lesson to AI. A wealth tax at least chases yachts and mansions. The core asset of an AI company is a model: in physical terms, a large file. It can be copied to a server in another jurisdiction over a weekend. The revenues are booked wherever the corporate structure finds it cheapest, which is a game multinationals had already mastered before this industry existed. The world's answer to that game, a global minimum corporate tax of 15 per cent, took years of diplomacy to agree and years more to phase in. Tax coordination moves at the speed of treaties. This industry moves at the speed of compute.

So the robot-tax idea keeps failing for a structural reason, not a lack of nerve. Whatever the rate, the levy lands on things that can leave: profits, revenues, the corporate shell. The gains quietly compound elsewhere, and the public collects a toll on an ever-shifting road.

The part the debate forgets

There is also something backwards about framing this as the state reaching into private pockets. The public already paid, up front, at the entrance.

The chips these models run on are subsidised by $52.7 billion of American public money plus a 25 per cent investment tax credit, and by more than €43 billion of public investment in Europe. The engineers were trained in publicly funded universities. The legal order that makes a model an ownable, defensible asset at all is public infrastructure. The book's phrase for this is that the public is the uncompensated lead investor: it funded the foundations of the venture and holds no equity in it. Seen that way, the question is not whether to confiscate something private. It is why the lead investor was never given shares.

What actually grips

Which points at the answer. If everything the company has can move, tax nothing that can move. Attach the public's claim to the one thing that cannot: the market.

A frontier AI firm can host its models anywhere on earth. What it cannot relocate is 450 million European customers, or 330 million American ones. Access to a market that size is worth more than any subsidy, and it is entirely within a democracy's gift. So make a stake the price of the licence to sell: operate here, and the public holds equity in what you are building, real shares, held in a fund, paying every citizen a dividend. The weights can move. The buyers cannot.

That single change of grip is what separates this from the robot tax. A tax is a claim on this year's flows, renegotiated every budget and dodged every year in between. A stake is a deed to part of the machine itself, wherever it earns and whatever it earns, compounding for the public the way it already compounds for private owners. It is the difference between billing a tenant and owning part of the building. And it is the first of the four tests this series keeps returning to: fund the public's share from assets, not flows; make it universal; armour it against raids; and do it now, while the wealth is still forming.

The honest reading

To be clear about what this argument is not. It is not anti-tax: schools, hospitals and pensions run on taxes, and nothing about public ownership replaces them; people who cannot work will always need a welfare state funded the ordinary way. Nor would an equity toll be easy: it needs the market power of a bloc like the EU or the US; it invites fierce lobbying; and a version watered down into symbolism would be worse than honest failure.

But after nine years, the robot-tax debate deserves a verdict, and the verdict is that it has been asking the machine for a wage substitute when the real question was always about title. The companies should indeed pay. The lasting way to make them pay is not a bill they can relocate away from. It is the answer to this series' standing question, applied: not how much the machine owes. Who owns the machine?

---

Where the evidence lives. Why taxes on mobile wealth keep failing is chapter 5 of The Horse Is Here to Stay; the case for the market-access stake, and why the public is already the machine's lead investor, is chapter 7. The tax bills and ownership proposals are scored as they appear in the Stable.

One of a series of free guides answering the questions people are actually asking about AI and work. The book carries the evidence, the sources and the counter-arguments in full. These guides are published under a Creative Commons Attribution-ShareAlike licence: translate them, teach from them, quote them, build on them.