In this article
When Geoffrey Hinton left Google in May 2023, the Spanish press dispatched him with a headline: "the godfather of AI regrets it". I went looking for the line that supposedly proved it and got a silly surprise: it doesn't exist. What exists is something more uncomfortable, less cinematic, and which almost nobody bothered to quote in full.
The quote anyone can check
I'll start where it can be checked, which in this trade is the decent thing. Hinton — one of the fathers of deep learning, 2018 Turing Award winner — left Google at seventy-five, and he did it in plain sight, in an interview with the New York Times that Cade Metz wrote and the paper published on 1 May 2023. It's in the NYT archive, behind a paywall, accessible with a subscription or from a university library. Whoever wants to go to the source has it there, dated and signed. No need to take my word for it.
To Metz he said this: "The idea that this stuff could actually get smarter than people — a few people believed that. But most people thought it was way off. And I thought it was way off. I thought it was 30 to 50 years or even longer away. Obviously, I no longer think that."
That's the line that got translated as regret. It's worth reading again, slowly, because it doesn't contain a single word of blame nor any reference to half a century of work he'd now lament. There's a man correcting an estimate: he thought the thing would take decades and it turned out it wouldn't. Turning that into "regret" is preferring the headline that sells to the one that's true, a cheap substitution that costs little and yields a lot in clicks. And the timeframe Hinton corrected isn't a calendar detail: it's the variable on which any political decision about regulating this hangs, because you don't legislate the same way against a problem arriving in fifty years as against one looming in five.
The honest chronology
Hinton has spent more than half a century inside neural networks. He started in the seventies, when a good part of the artificial intelligence community wrote them off as a dead end, and he stayed there in the eighties, when almost no one funded them. He co-developed the backpropagation algorithm alongside David Rumelhart and Ronald Williams. In 2012, his Toronto group — with Alex Krizhevsky and Ilya Sutskever — presented AlexNet, the network that won the ImageNet competition and split the history of computer vision in two. Before that, deep networks were a university-department promise hard to defend in a thesis committee. After it, they were a business bought at auction price.
Google took him in March 2013, by acquiring DNNresearch, the small company Hinton had set up the year before with Krizhevsky and Sutskever following the ImageNet triumph. He worked there for a decade, splitting his time between Toronto and the company, until that May 2023 departure.
It's worth noting what this chronology rules out. The departure wasn't the late disillusion of someone who looks back and renounces their work; it was a reaction to something very recent. Between 2022 and 2023, the improvement of the large language models — GPT-3.5, ChatGPT, GPT-4 — ran faster than he'd anticipated, and what he placed at thirty or fifty years began to look like a matter of less than ten. The recalibration he found so hard to digest came from the data in front of him, not from an examination of conscience.
Why leave Google to say that? Hinton was clear, and here the translation matters because it's circulated deformed. In the tweet he posted on 1 May 2023 he wrote, literally: "I left so that I could talk about the dangers of AI without considering how this impacts Google. Google has acted very responsibly." He didn't say he was leaving to "speak freely", as if he'd been gagged; that's the paraphrase that later slipped into half the press. He said he was leaving so as not to have to weigh the impact of his words on his employer while he was on its payroll. And in passing he defended that employer, which spoils any narrative that needs a corporate villain. The nuance evaporates the moment someone prefers the version with a film baddie.
The argument that goes the other way
What Hinton has held since 2023, and has repeated in dozens of appearances, doesn't fit in the word "regret". It begins with an advantage he considers structural and not rhetorical: digital systems can be copied and can merge what they've learned. If two instances of the same model learn different things, they exchange the weights and share the knowledge in one stroke, with no classes, no reading, no years in between. A biological brain can't do that; you can't dump what you know into me by opening my head. Hinton sees there an advantage of the digital over the biological, and he describes it as an engineer who knows exactly which architecture allows that dump, not as a populariser stretching a metaphor until it snaps.
Then there's alignment. Getting a system to pursue what we really want, and not its own reading of what it thinks we want, is still unsolved, and nothing guarantees it'll be solved before systems above human level appear. For Hinton that's an open engineering problem, with the added discomfort that the speed at which the models are deployed runs ahead of the speed at which the problem gets solved.
And here comes the nuance the press erased without blinking. For Hinton the nearest danger isn't the rebellious superintelligence of the films, but the use of AI by bad actors: states, mafias and extremist groups employing it for mass manipulation, cyberattacks, autonomous weapons, targeted disinformation. That isn't a risk you have to imagine for a decade away. It's already happening. That his most controversial position within the field ends up being also the most quoted out of context says a fair amount about how tech news gets cooked.
What he didn't quite say
It's worth being just as precise about the limits of what the man stated, because there the media translation did the finest damage.
He didn't say current AI is conscious; that one was hung on him. He has gone so far as to hold that some models "understand" in some sense of the verb, but he carefully separates understanding from being conscious, and on consciousness he acknowledges he has no clear position. He acknowledges it himself, in his own words, I don't deduce it from a silence that suits me.
Nor did he announce an imminent apocalypse. What he's handled are probabilities, not prophecies: of the order of 10 to 20 per cent, according to the interview, that systems smarter than us come to be a serious risk to humanity. He himself describes that figure as an intuition and not a calculation, and acknowledges that estimating these things is very hard. A probability of 10 to 20 per cent, said without cheap relief, leaves an 80 or 90 per cent of scenarios in which nothing resembling a catastrophe happens; which isn't the same as reassuring, because few of us would board a plane with that probability sheet.
And he didn't ask to halt the entire research effort. He asked to pause work on frontier systems while progress is made on alignment, which is a much narrower request. AI applied to medicine, to basic science, to energy efficiency, he defends without hesitation. His critique is technical and bounded: there's an unsolved problem and we're running ahead of it. Calling that Luddism is not having read the sentence, or having read it with the headline already written.
These clarifications aren't the fussiness of an idle scholar. The headline — "the godfather regrets it" — invites filing Hinton among the professional doom-mongers, and the full version of what he says doesn't fit in that drawer even by shoving. That's why, I suspect, the full version is barely quoted.
The Nobel in Physics, and a speech that gives the character away
On 8 October 2024 the Nobel Committee awarded the Nobel Prize in Physics to John Hopfield and Geoffrey Hinton "for foundational discoveries and inventions that enable machine learning with artificial neural networks". Hopfield had developed in the eighties the associative networks that bear his name; Hinton, with Terrence Sejnowski, extended that work with the Boltzmann machines. The prize recognised four decades of neural-network physics, not the character's current media standing.
What interests me most about the episode is the speech. Hinton gave his Nobel lecture in the Aula Magna of Stockholm University on 8 December 2024, titled it Boltzmann Machines and opened it like this: "Today I am going to do something very silly: I am going to try to describe a complicated technical idea to a general audience without using any equations." There's the whole man, in a single opening line. A scientist who decides to address the public with technical honesty, who doesn't hide behind the maths to impress, who takes for granted that whoever's listening can follow him if he bothers to explain himself. The lecture was later published in Reviews of Modern Physics, the American Physical Society's journal, in August 2025. In it he recalled how Sejnowski and he hit on an unexpected use of Hopfield's networks: instead of storing memories, using them to build interpretations of the sensory input. Out of that came the Boltzmann machines, and out of that, in large part, the deep learning that today moves billions.
The odd thing about the picture is the overlap of the two roles. A Nobel rewarding one of the founders of a field exactly at the moment when that founder has become its most authoritative critic. Figures like that aren't common, capable of signing the foundation and the warning with the same hand. Oppenheimer, Joseph Rotblat. Few others, and none comfortable to have at the table.
Why who makes the warning matters
Pausing on who signs the warning isn't an argument from authority in disguise, but a way of gauging how much it weighs. Hinton knows technically what he's talking about; he isn't a newcomer who learned AI three summers ago to sell a talk at a conference hotel. When he states that a system can share knowledge between instances, he isn't speaking by analogy but by architecture, because he has designed pieces of that architecture. And, unlike the executives who also warn of risks while raising funding rounds to sell you the remedy, he has nothing to place: he left Google, he draws his pension, he has the Nobel, he runs no fund nor signs airport bestsellers. What he says doesn't answer to an expectation of return, and in this business that's almost an eccentricity.
There remains the main thing, which is the exact shape of his position, and it's worth holding it in its terms and not the headline's. Real but not imminent risk. Significant probability but far from certainty. Necessary regulation without halting all research. A stance like that won't be reduced either to Silicon Valley's catalogue optimism or to viral-thread alarmism, and there's someone with the credentials to defend it without an alibi. The trouble is that holding it asks for readers who follow it in its long version, not in its clipping. The NYT interview is there in full; the conversations he's taken part in are recorded and dated; the Nobel speech is published in a peer-reviewed journal. Telling apart what Hinton said from what was attributed to him costs no more than going to the source. That the Spanish-language media ecosystem, with its honourable exceptions, hasn't taken that trouble isn't a failure of access. It's a failure of will.
Definitions
Geoffrey Hinton: a British-Canadian researcher, born in 1947, professor emeritus at the University of Toronto, considered one of the fathers of modern deep learning.
Turing Award: the ACM's annual prize, regarded as the equivalent of the Nobel in computer science. Hinton received it in 2018 alongside Yoshua Bengio and Yann LeCun.
Deep learning: a family of machine-learning techniques based on neural networks with many layers, able to learn complex representations from large quantities of data.
Backpropagation: an algorithm that allows a neural network to be trained by adjusting its weights according to the output error. Hinton was one of its fundamental authors, along with David Rumelhart and Ronald Williams.
Boltzmann machine: a type of stochastic neural network developed by Hinton and Sejnowski in the eighties, the basis of the work recognised by the Nobel.
AlexNet: a convolutional neural network developed by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton in 2012, winner of the ImageNet competition and trigger of the modern era of deep learning.
Alignment: the technical problem of getting an AI system to pursue goals that coincide with those of the human operator. Considered an open problem.
References
Metz, C. — The Godfather of A.I. Leaves Google and Warns of Danger Ahead, The New York Times, 1 May 2023. The interview from which the quote on the recalibration of timeframes comes.
Hinton, G. — Tweet of 1 May 2023 (x.com/geoffreyhinton/status/1652993570721210372), where he clarifies the reasons for his departure from Google. Original text: "I left so that I could talk about the dangers of AI without considering how this impacts Google. Google has acted very responsibly".
MIT Technology Review — Geoffrey Hinton tells us why he's now scared of the tech he helped build, 2 May 2023, coverage of his departure and his warnings.
CNN Business — AI pioneer Geoffrey Hinton quits Google to warn about the technology's dangers, 1 May 2023, which records that he left Google to be able to speak about the risks of AI.
TechCrunch — Google Scoops Up Neural Networks Startup DNNresearch, 12 March 2013, on the acquisition of the company founded by Hinton, Krizhevsky and Sutskever.
The Guardian — Hinton's statements estimating between 10 and 20 per cent the probability that AI leads to a serious risk for humanity (December 2024 coverage).
NobelPrize.org — the official announcement of the 2024 Nobel Prize in Physics to John Hopfield and Geoffrey Hinton, 8 October 2024.
Hinton, G. — Nobel Lecture: Boltzmann Machines, the Nobel lecture delivered on 8 December 2024 in the Aula Magna of Stockholm University, published in Reviews of Modern Physics (RevModPhys.97.030502), August 2025.
Russell, S. — Human Compatible (Viking, 2019), on the problem of control and alignment.
Christian, B. — The Alignment Problem (Norton, 2020), an overview of the alignment problem cited in the article.
You might also like
- Timnit Gebru and her departure from Google
- The real dangers of AI
- Sam Altman's real credibility
- "Attention Is All You Need", the paper that changed AI
Comments0
No comments yet.
Leave a comment