The Bill Gates robot tax gets the diagnosis right
The Bill Gates robot tax is back, and this time it has company. In an essay on his blog on August 26, Gates proposed taxing robots and AI tokens and using the money for retraining and a stronger safety net. His core complaint is precise: "The tax system nudges you toward replacing people with machines." He is right about that. He is wrong about what to meter. A token counts computation. It does not count the job that left. Tax the margin that replaced the job, not the tokens that did the typing.
Gates first floated a robot tax in 2017, and it drew sharp criticism, including from former Treasury Secretary Larry Summers. Nine years later the idea returns with a friendlier audience. Six days after the essay, Fortune was reporting on a House bill that taxes AI companies on tokens or revenue and raises the rate as unemployment climbs. The robot tax has left the op-ed page. It now has sponsors, a named bill and a foreign minister arguing against it. That is the moment a campaign should say exactly where it agrees and where it does not.
The agreement is large. Gates says the tax would need to be targeted, so that it does not slow the purely beneficial uses of AI, like cheaper medicine and education. He says the money must reach the people who need it most. Tax The Agent says the same, and adds one design rule. The base has to move when the payroll moves. A levy that grows when nobody loses a job, and shrinks when many do, is a symbol. It is not a transition fund.
Payroll is a toll that only humans pay
Start with the asymmetry Gates describes, because it is written in statute and anyone can check it. In the United States, the employer pays 6.2% of wages for Social Security and 1.45% for Medicare, and the employee pays the same again, according to the IRS. That is 12.4% plus 2.9%, or 15.3% of wages. The Social Security part applies to earnings up to the wage base, which is $184,500 for 2026. An additional 0.9% Medicare tax is withheld on an individual's wages above $200,000.
Now run the same work through software. A company that hires a claims clerk sends the employer share, 7.65 cents on every wage dollar, into the system that pays pensions and hospital bills. A company that deploys an agent to process the same claims sends nothing through that channel. Gates put it plainly. Hire a person and you pay payroll taxes on the earnings. Buy a robot and you can usually write it off right away as a business expense. The work gets done either way. Only one version pays into the social contract.
This is why "robot tax" is a bad name for a real problem. Nobody needs to tax a machine for existing. The issue is that payroll contributions ride on wages, and wages are exactly what automation removes. Every task that moves from a salaried person to an agent takes its contribution with it. At one firm, that is a rounding error. Across an economy where agents draft, route, reconcile and answer at scale, it becomes a slow leak in the base that funds retirement and health care.
Congress turned the idea into a bill with an alarm attached
On August 7, Representatives Greg Casar, Valerie Foushee and Sara Jacobs introduced the AI Tax and Work Protection Act. It would tax large AI companies on the higher of two values: the value of the tokens they sell, or the revenue they earn selling AI products. According to Fortune, the rates would start at 2% on tokens and 3% on revenue while unemployment is 5% or less, and rise as unemployment increases. The money would fund a new Work Protection Administration that creates jobs in housing, infrastructure, child care and elder care.
The trigger is the smartest part. It ties the rate to the harm the tax exists to answer. If agents do not displace workers, the rate stays low and the argument about mass unemployment settles itself. If they do, the rate climbs without a new vote. Casar framed it without subtlety: "This bill says: we will not let AI billionaires get rich by putting you out of work." Jacobs added that if AI profits off human work, workers deserve job security and a share of those profits.
The weakness is where the bill looks. It taxes the companies that sell AI. The payroll that disappears belongs to the companies that deploy it: the insurer, the bank, the outsourcer, the agency. A lab can sell the same volume of tokens to a firm that empties a floor of claims staff, or to a firm that hires. The bill charges both sales the same way. The trigger reads the national labor market. The tax base does not read the firm where the job left.
A token measures effort, not the job it replaced
Tokens are attractive to lawmakers because they are countable. Every model call produces them, and providers already bill by them. That is also the trap. A token counts computation. A researcher can burn through enormous volumes screening papers without removing a single job. A support desk can replace a team with a workflow that uses far fewer tokens than the researcher. A flat token tax charges the curious scientist more than the company that emptied a call center.
The second problem is time. Every lab competes to make its models cheaper to run. If the price of a token falls while the number of jobs it replaces rises, a tax on token value collects less exactly when the displacement grows. The Casar bill softens this by taxing the higher of token value or revenue, and by raising rates with unemployment. A pure token tax has no such guard. Its base erodes on the same curve as the industry's costs.
The third problem is geography. Tokens can be bought from a provider in another country, routed through a subsidiary, or bundled into a software subscription that never itemizes them. A tax on a technical unit invites the relabeling that tax lawyers are paid to find. None of this makes the token useless. It makes it a good audit signal and a poor tax base. Use it to check a claim about how much work agents did. Do not use it to decide what is owed.
Taxing the machine punishes the wrong decision
The robot half of the Gates plan has the opposite problem. It taxes the purchase, which is the moment of investment, not the moment of displacement. A hospital that buys a robot to lift patients and keeps every nurse would pay the same as a warehouse that buys the same robot and cuts a shift. Gates wants the tax targeted, and he is right to worry. A tax on machines as machines cannot tell a tool that helps a worker from a tool that removes one.
Gates has answered the efficiency objection head on. According to Going Concern, he wrote that "we'll be able to afford a little inefficiency as the price for keeping people employed." That is an honest trade. It is also the sentence opponents will quote in every hearing, because it concedes that the tax is a brake. For a tax on purchases, they would be partly right. It slows the investment whether or not anyone loses work.
The campaign position is narrower, and harder to caricature. Do not slow the machine. Do not tax the purchase order. Charge the margin the machine produces once that margin stops paying into the systems that payroll used to fund. A firm that automates and keeps growing its workforce would owe little. A firm that automates and sheds the payroll would owe on the output that moved. That is not a brake on innovation. It is a receipt for the transfer.
The loser narrative objection deserves a straight answer
The sharpest reply to Gates came from Australia. In remarks reported by The Epoch Times on September 16, at the Australian Strategic Policy Institute's Sydney Dialogue, minister Tim Ayres rejected the token tax. He said he is always slightly wary when large technology companies argue that their own technology will cause large-scale job losses. He warned of "a kind of loser narrative" that "suits some interests and not others." His answer was investment, employment and growth in Australia.
He has a point, and the campaign should concede it. When the people selling a technology announce that it will end careers, they are also announcing that it works. Fear is a sales pitch. A tax designed around a prophecy of mass unemployment lets the prophets set the terms. If the unemployment never comes, the policy looks foolish. If it comes slowly and unevenly, sector by sector, a national trigger may never fire at all.
But the objection argues for a better base, not for no base. An attributable-margin contribution does not require anyone to believe a forecast. It asks a measurable question inside each firm: did output that used to carry payroll contributions move to systems that carry none? If the answer is no, the contribution is zero. That is the opposite of a loser narrative. It pays only where the transfer happened, and it stays silent everywhere else.
Tax the margin that left payroll
The campaign's alternative is the agent contribution, a proposal built on what we call the agentic margin. It is not law anywhere, and no government has adopted it. It would apply a rate to the share of a firm's revenue that is attributable to autonomous work, the output that agents now produce where people once did. The deploying firm pays, not the model maker. The base is revenue, which already exists and is already audited. The attributable share is the part a firm would have to defend, with the kind of documentation tax authorities already demand for transfer pricing and research credits.
Compare it with the proposals on the table. Against the robot tax, it charges the result, not the purchase, so the hospital that keeps its nurses owes nothing extra. Against the token tax, it follows value, not computation, so cheaper models do not shrink the base. Against the Casar bill, it lands on the firm where the job left rather than the lab that sold the model. It keeps the best idea in that bill: the rate is a dial that a legislature could tie to what happens in the labor market.
The shape is easiest to see in a sector that sells hours of process work. In our hypothetical business process outsourcer scenario, nearly half the revenue is attributable to autonomous work, and the headcount it once required is the headcount that leaves. A token tax would charge that firm for its compute bill. A robot tax would charge it for almost nothing, because it bought no robots. The agentic margin charges it for what actually changed.
Run your own numbers before the next hearing
The AI Tax and Work Protection Act is a bill, not a law, and the Gates essay is an essay. The debate they started will outlast both. Candidates will be asked whether AI should pay, and "yes" is not an answer. A useful answer names a base, a payer and a trigger. Gates named tokens and robots. Casar, Foushee and Jacobs named tokens or revenue, with an unemployment alarm. The campaign names the attributable margin of the deploying firm.
One move for this week. Pick one company you know well, your employer or a client. Estimate its annual revenue and the share of that revenue now produced by autonomous systems rather than people. Then set the rate at a level you would defend in public. Run your own numbers in the illustrative calculator linked below, and set the result beside the 7.65 cents per wage dollar that the same work used to send through the employer's payroll.
If the figure looks too small, the rate is a dial, and the argument is about the dial. If it looks too large, ask what the displaced payroll contributions would have been. Either way, you will hold what most of this debate lacks: a base, a payer and a number you can explain. Then ask the people running for office which of the three bases they would tax, and why.
