8 October 2026 ยท Breaking
Daron Acemoglu's 'Will AI Replace Workers?'. The Same Tool Can Do Both.
Daron Acemoglu, who shared the 2024 Nobel Prize in economics, has a simple message: AI doesn't have to replace workers. Writing in The Humanist Review of AI1, he argues that the industry is building the wrong thing. It is chasing machines that imitate people and take over whole jobs, when it could build tools that make people better at theirs. He calls this "pro-worker AI" and says it "is within the frontier of existing models and can be developed easily".
Much of this is right, and worth saying loudly.
Direction is a choice. Nothing in the technology says it must replace the radiologist rather than help her. Acemoglu names four forces pushing towards replacement. Business models are built on software licences and advertising. A handful of giant firms buy up anyone who might do it differently. Firms and toolmakers each wait for the other to move. And the industry believes, as far back as Alan Turing, that the goal is a machine as clever as a person.
His best point is about tax. He says American workers face an effective marginal tax rate above 25%, while machines and software face close to zero. If that's right, the tax code pays firms to swap people for equipment. You don't have to believe anything about AGI to want that fixed. Treat a pound spent on a worker the same as a pound spent on a machine, and a lot of "inevitable" automation stops paying.
Here is the problem. His own best example is a classroom tool that shows a teacher where each pupil is stuck, so she can regroup the class and give individual attention that used to cost a private tutor. That is a lovely tool. It could also be used to let one teacher handle a class that used to need two.
The split between "replaces" and "helps" works for a single task. It blurs for a whole job. Almost every tool that makes a worker more productive means that, at the same output, you need fewer of them. Spreadsheets helped accountants, and firms hired fewer bookkeepers. Whether a tool ends up pro-worker depends less on how it's built than on who pays for it and who keeps the savings. A school that buys the tool to raise results keeps its teachers. A council that buys it to cut costs doesn't. Same tool.
So the proposals that will matter most are the ones that change who gains, without needing anyone to be convinced. Equal tax on labour and capital. Paying people for the data that trains specialist tools. Competition, so that tools built for workers have someone to sell to besides the few giant companies that sell to their bosses.
The proposal that will matter least is his call to change the conversation away from AGI. He wrote in July that model progress had outrun his forecasts but real-world use hadn't. Since then OpenAI has published 722 new maths results2 from a model nobody can buy yet, and the White House has ordered its agencies to call AI "Super Intelligence"3. The labs and the state have chosen their goal, and they chose it on purpose. An essay won't turn them round. A tax code might change what the rest of the economy does with what they build.
Acemoglu ends by saying "a better future lies in cultivating human skills, not sidelining them." Agreed. What his essay doesn't settle is who gets paid for the extra work that skilled human now does.
Update โ 8 October 2026. Added after publication, from correspondence. The two paragraphs below are not part of the original argument above, and the second corrects it.
David Autor pointed out in correspondence that higher productivity alone doesn't tell us who benefits. A key question is whether technology makes workers' expertise more or less scarce relative to demand. It can expand what skilled workers can do, or make their skills widely available โ with very different effects on pay. His paper with Daron Acemoglu and Simon Johnson, Building Pro-Worker AI4, sets this out. Credited with permission. This is WE's paraphrase, agreed with him, not a quotation, and it does not mean he endorses this essay or WE.
WE's note: the original post put too much weight on who controls the savings and too little on how technology changes demand for expertise and the number of people able to supply it. Ownership and bargaining power still matter, but they do not settle the outcome by themselves. One correction too: the post originally said the classroom tool "lets one teacher handle a class that used to need two", as if that were established. It isn't. Cutting staff is one possible use of such a tool, and the text above now says so.