What LeCun thinks the alarmists got wrong

On April 18, 2026, Yann LeCun, Meta’s former chief AI scientist and co-recipient of the 2018 Turing Award, posted on X a sentence that doubled as an instruction to his readers.

“Dario is wrong,” he wrote, of Anthropic chief executive Dario Amodei. “He knows absolutely nothing about the effects of technological revolutions on the labor market. Don’t listen to him, Sam, Yoshua, Geoff, or me on this topic. Listen to economists who have spent their career studying this.” He tagged five of them: Philippe Aghion, Erik Brynjolfsson, Daron Acemoglu, Andrew McAfee and David Autor.

The post was characteristic. LeCun has spent the better part of three years telling people that the figures most associated with warnings about artificial intelligence — the OpenAI chief executive Sam Altman, his Turing co-laureates Yoshua Bengio and Geoffrey Hinton, and Amodei — are speaking outside their competence. The April tweet did the unusual thing of including himself in that list, but only on labor markets. The instruction is straightforward: the people warning that AI will eliminate large numbers of jobs are the wrong people to ask. Ask the economists.

The instruction can be followed, and LeCun’s own list is the place to start. Daron Acemoglu of MIT, the third name on it, shared the 2024 Nobel Prize in economics with Simon Johnson and James Robinson for their work on how political institutions shape long-run prosperity, and has written for decades on automation and employment. In 2023, he and Johnson published Power and Progress, a 560-page book arguing that the gains from new technology reach workers only when innovation creates new tasks for them and institutions spread the gains, and that neither condition is automatic.

The book’s central claim cuts against a reassurance LeCun gave the day after the April post. Replying to a critic on X on April 19, he wrote: “AI is just another one of those technological revolutions. People like Dario present it as qualitatively different from previous revolutions. It’s not.” Acemoglu and Johnson do not argue that AI is a different kind of revolution. Their quarrel is with the premise that revolutions pay workers by default. The mechanism by which prior technological revolutions raised wages, they argue, was not automatic and was not a property of technology as such. It required a particular kind of innovation, the kind that creates new tasks for which human labor is suited, and a particular set of political conditions under which the gains from those new tasks were broadly shared.

The technical core of their argument is what Acemoglu has called “so-so automation.” In a 2019 paper with Pascual Restrepo of Boston University and in subsequent writing, he distinguishes between technologies that automate tasks while substantially raising productivity and technologies that automate tasks at roughly the productivity level of the human workers they replace. The second category, he argues, displaces labor without producing the surplus that funds replacement employment. Much of contemporary machine learning, in his account, sits in the second category: cheap enough to deploy, good enough to substitute, not transformative enough to seed a new wage-paying sector around it.

A 2020 paper of his in the Journal of Political Economy, also with Restrepo, used variation across U.S. commuting zones to estimate the labor-market effects of industrial robots between 1990 and 2007. The paper found a negative effect on both employment and wages in zones that adopted robots most heavily, with no clear offsetting employment growth elsewhere in the same regions.

That body of work is not a forecast of rapid, large-scale job loss. In a 2024 paper, “The Simple Macroeconomics of AI,” Acemoglu estimated that AI would raise total factor productivity by no more than 0.66 percent, cumulatively, over ten years, an effect he called nontrivial but modest. Nor does it support the reassurance that technological revolutions sort out the labor market on their own. A reader who followed LeCun’s instruction to the third name on his list would find an economist whose work undercuts LeCun’s premise.

There is a counter available, and LeCun’s list supplies it. Acemoglu is one economist, and the profession has not converged on Power and Progress as the consensus reading. David Autor at MIT, also on the list, has written on AI’s potential to lower the skill threshold for high-value work. Erik Brynjolfsson at Stanford, another, has emphasized augmentation. Anton Korinek at Virginia has modeled scenarios in both directions. The discipline is not unanimous. Treating the most pessimistic name on the list as the verdict would be the same move LeCun objects to when safety researchers cite Hinton.

The April post names the economists without saying which of their findings it relies on. His reply the next day, that AI is “just another one of those technological revolutions,” cites none of their findings.

The same shape recurs on the question of catastrophic risk. On X on April 2, 2023, he wrote: “The only reason people are hyperventilating about AI risk is the myth of the ‘hard take-off’: the idea that the minute you turn on a super-intelligent system, humanity is doomed.” To Time, in February 2024: “It’s not going to happen in one day… That’s the preposterous scenario.” The argument is that every prior technology has rolled out incrementally, with prototypes, fixes, wider rollout, and regulation in parallel.

This is a coherent argument against a specific scenario. The scenario is the one Nick Bostrom laid out in his 2014 book Superintelligence, in which a sufficiently advanced AI system bootstraps itself to vastly superhuman capability over a period of days or weeks, escapes meaningful human oversight, and pursues objectives misaligned with human interests before any regulatory response is possible. That scenario, the hard takeoff, framed much of the safety writing that followed Bostrom’s book.

Much of the safety case made since 2023 does not depend on it. Several prominent arguments are slow-takeoff and gradual-disempowerment arguments, claims that danger does not require a single recursive self-improvement event but emerges from the gradual capability growth that everyone, including LeCun, agrees is happening. Some of these arguments sketch worlds in which advanced AI systems progressively automate consequential decisions across institutions, with no individual handoff that looks dangerous, until human oversight has eroded past the point of meaningful correction. The argument is that the relevant threshold is not superintelligence but sufficient automation of economic and political processes that human agency over them weakens past recovery.

Bengio makes one version of this case. He signed the Future of Life Institute’s pause letter in March 2023, has reorganized his work toward AI safety research, and with coauthors published “Managing Extreme AI Risks Amid Rapid Progress” in Science in May 2024. The paper describes paths to catastrophe, from misuse to loss of control, that do not require a sudden takeoff. It rests on the proposition that capabilities are continuing to scale, that current alignment techniques are not robust to that scaling, and that the gap between what models can do and what their developers can verify they will do is widening. None of these claims requires recursive self-improvement.

LeCun has argued with Bengio directly: in June 2023 they took opposite sides of the Munk Debate in Toronto on whether AI research poses an existential threat, with Max Tegmark, the MIT physicist who leads the Future of Life Institute, beside Bengio and Melanie Mitchell beside LeCun. That was before the Science paper. He has rebuked Bengio, Hinton, and Tegmark for x-risk advocacy; “You, Geoff, and Yoshua are giving ammunition to those who are lobbying for a ban on open AI R&D,” he wrote to Tegmark on X in October 2023. His quoted rebuttals treat the safety case as a continuation of the hard-takeoff discourse. Bengio has built a research program around the slow-takeoff case, anchored by the 2024 Science paper. LeCun’s public response has been to say the safety community has fallen for a myth that the slow-takeoff case does not depend on.

A second counter is available, and it is stronger than the first. LeCun could reasonably reply that the slow-takeoff arguments rest on capability projections he believes are technically wrong, that the current generation of large language models is, on his view, structurally incapable of the kinds of planning, world-modeling, and goal-directed behavior the gradual-disempowerment scenarios assume. Engaging the slow-takeoff literature in detail would, on this view, be engaging a hypothetical built on a technical premise he has spent most of his recent academic work arguing against. It is a real reply, and LeCun gestured at it in the October 2023 post to Tegmark: “Current Auto-Regressive LLMs are not driven by objectives, so let’s not extrapolate from their current weaknesses.” His other quoted statements are pitched against the hard-takeoff version.

The two positions hang together as a single move. On x-risk, the rebuttal is aimed at a scenario much of the current safety case does not need; on jobs, the reassurance rests on a premise that Acemoglu, one of the economists LeCun names, has written a 560-page book against. Both arguments tell the worriers they are listening to the wrong experts. On jobs, the experts LeCun names include Acemoglu, who undercuts him. On risk, the experts he dismisses have already moved past the scenario he rebuts.

The economists LeCun names do not all agree with him, and one of them, Acemoglu, rejects the premise that technology’s gains reach workers on their own. Bengio’s published case rests on a class of risk that LeCun’s quoted rebuttals do not address. The reader is left to decide whether the instruction to consult those people is best honored by reading what they have written.