Acemoglu rejects both AI camps, and calls fear the costlier error

In an interview with Fortune published on August 15, 2026, MIT economist and Nobel laureate Daron Acemoglu declined to join either side of the AI argument, and said that giving in to fear is the worst kind of response available right now ^[1][2]. For anyone sizing capital spending against measured output, that is more useful than it sounds, because the two positions he rejects are the two standard ways an AI budget goes wrong.

The first camp, in his description, is the true believers, "quasi-moderate-friendly" types so convinced AI will be good for everybody with no exception that any challenge "drives them insane" ^[3]. The second treats the models as "stochastic parrots" generating plausible-sounding noise, a view he said parts of the left hold so firmly that any argument for genuine capability enrages them ^[4][5]. He called the dynamic "very unproductive" ^[6].

His own position is deliberately split. "I think you have to really have your head in the sand to think that AI is a stochastic parrot right now," he told Fortune, while adding that he is "also not willing to go along with some inchoate belief that everything will work out fine" ^[7][8]. He credits frontier models with genuine advances in comprehension, coding and scientific and mathematical discovery, and singled out the way proofs are being done ^[9]. Holding two conflicting ideas at once, he said, has become "sort of radical" ^[10].

The operational content sits in what he says about deployment. He borrows Wharton's Ethan Mollick's "jagged frontier": exceptional at some tasks, unreliable at others, sometimes an unusual combination of the two, which means "you need to do a lot of detailed babysitting" ^[11][12]. Babysitting is labor, and labor is cost; a workflow that only holds up under supervision does not clear a payback hurdle on the vendor's arithmetic. He also flagged that beyond coding there is not much evidence of wide adoption, noting that customer service employment has barely budged in recent years, and manufacturing likewise ^[13].

Set that against his long-running dispute with Stanford's Erik Brynjolfsson, a former colleague who argues for substantially greater productivity gains than Acemoglu projects ^[14]. Acemoglu says the two "actually agree on many things" and treats their public disagreement as a model for how policy debate should work ^[15]. Read as an input rather than a feud, that is a range, and a wide one. A capex case underwritten at the top of it is a bet on the believer camp rather than on measurement. Acemoglu added that he was pleased Brynjolfsson has increasingly called for redirecting AI in human-complementary directions, which he called "my bugbear for almost two decades" ^[16]. He has studied AI explicitly since at least 2018 ^[17], which puts the redirection argument roughly a decade ahead of his own AI work ^[18].

On why fear is the more expensive mistake, the interview gives the mechanism only indirectly: Acemoglu ties the escalation to a media environment that rewards attention, in politics and in AI alike ^[19]. The practical version for operators is that both errors are procurement decisions. Believers overpay for capability that needs babysitting; the frightened either abstain from a general-purpose technology or cut headcount ahead of any measured gain, on the strength of a narrative.

Worth watching: adoption breadth outside coding, starting with the customer service and manufacturing employment series Acemoglu cites ^[13]; Fortune's Canaries dashboard, built on ADP data, which it has used to track potential AI-linked job losses ^[20]; and whether human-complementary redirection shows up in product roadmaps rather than conference remarks. Acemoglu's own posture: "I wouldn't call myself an optimist, I would say I resolutely refuse to give up hope" ^[21].