Geoffrey Hinton's Farewell: The Godfather Who Changed His Mind

We have created entities that may eventually be smarter than us. We do not yet understand these entities well enough to know what they want. And we are making them more capable very fast.

Toronto, Canada. May 1, 2023. Geoffrey Hinton sends an email to his contact at Google, informing the company that he is resigning. The reason he gives is that he wants to speak freely about the dangers of artificial intelligence — something he believes he cannot do while employed by one of the companies most invested in developing it.

Hinton is seventy-five years old. He has spent more than fifty years working on neural networks and artificial intelligence. He is, by general consensus, the most important single contributor to the deep learning revolution that has transformed AI — the person who maintained the neural network research programme through two AI winters, who co-developed backpropagation, who trained AlexNet with his students and announced to the world that the revolution had arrived, who has been called, alongside LeCun and Bengio, a Godfather of Deep Learning.

He has received the Turing Award. He will receive the Nobel Prize in Physics. He is as credentialed as a scientist can be.

He is leaving Google because he is afraid of what he helped build.

“I console myself with the normal excuse,” he tells the New York Times. “If I hadn’t done it, somebody else would have. But that’s not necessarily true.”

The most consequential change of mind in the history of AI has begun.

---

The Long Journey: From Edinburgh to the Nobel Prize

Geoffrey Everest Hinton was born on December 6, 1947, in Wimbledon, England. He is the great-great-grandson of the mathematician George Boole, whose Boolean algebra is the mathematical foundation of modern computing — a genealogical connection that feels almost too on the nose for a man who has spent his life at the foundation of what computing has become.

He studied experimental psychology and philosophy at the University of Cambridge, seeking to understand how the human mind worked. The intellectual question that animated his early life was not primarily about computers — it was about intelligence itself, about how minds could represent knowledge and learn from experience and recognise patterns in a world that was vast and varied and uncertain.

He was, from early in his career, drawn to connectionism — the view that intelligence emerged from networks of simple interconnected elements, each adjusting its connections based on experience. This view — which would eventually become the deep learning paradigm — was, in the 1970s and 1980s, deeply unfashionable. The dominant view in AI was that intelligence required explicit symbolic representations and rule-based reasoning. Connectionism was a minority view, associated with researchers who were working in the margins of the field.

Hinton worked in those margins for decades. He moved to the United States, where he did postdoctoral work at the University of California, San Diego, with the PDP Research Group that would produce the influential parallel distributed processing volumes. He moved to Carnegie Mellon University. He moved back to England. He moved to Canada, joining the University of Toronto, where the Canadian government’s commitment to basic science research provided the stability that the American university system had not.

Through all these moves, through the two AI winters, through the years when grant funding was difficult and conference acceptance was harder, Hinton maintained his conviction that the connectionist approach was right — that neural networks, given enough data and enough computing power, would eventually work. The conviction was based not on faith but on specific theoretical insights about what neural networks could represent and how learning algorithms could efficiently find those representations.

The conviction was vindicated, eventually and dramatically, by AlexNet in 2012 — the event that announced the deep learning revolution and that Hinton, through his students Alex Krizhevsky and Ilya Sutskever, had directly produced. The man who had spent thirty years being told he was pursuing a dead end had been right.

---

The Google Years: Power, Resources, and Growing Unease

After the AlexNet breakthrough, Hinton — with Krizhevsky and Sutskever — formed DNNresearch, the company that was acquired by Google in 2013. The acquisition brought Hinton, at sixty-five, into the centre of the most powerful technology company in the world, with access to computing resources and research collaborators that had previously been unimaginable.

His years at Google were productive in specific ways. He contributed to the development of capsule networks — an architectural innovation designed to address some of the limitations of convolutional networks. He contributed to the understanding of deep learning theory — to the questions of why and how large neural networks learned effectively. He contributed to the broader intellectual culture of Google Brain, one of the most prestigious AI research organisations in the world.

But the years at Google were also, gradually, years of growing unease. The specific concerns that would eventually lead to his resignation were not sudden — they developed slowly, as the capabilities of the systems he was helping to build became more impressive and as the implications of those capabilities became more serious.

The unease had several specific dimensions.

The scale of the systems. The neural networks that Hinton had worked on for decades were, by the late 2010s, vastly larger and more capable than anything that had seemed possible in the early days of his research. GPT-3, which Google’s researchers understood in technical detail, was demonstrating capabilities that exceeded what most people had expected at its scale. The trajectory was clear: continued scaling would produce systems more capable still. The capabilities were not just impressive — they were beginning to feel consequential in ways that made the question of what was being built less comfortable.

The alignment question. As AI systems became more capable, the alignment problem — the question of whether these systems would pursue objectives that were genuinely aligned with human values — became more urgent. Hinton had engaged with this question at a technical level and had become increasingly concerned that the alignment problem was harder than the field’s mainstream was acknowledging. The systems were getting smarter, and the question of whether they were getting safer at the same rate was increasingly answerable in the negative.

The geopolitical dimension. The development of very capable AI systems in the context of great-power competition — between the United States, China, and other major powers — created specific concerns about the ways that AI might be used for surveillance, for military advantage, for the concentration of power in ways that were inconsistent with democratic governance. The geopolitical dimension of AI was one that Hinton found increasingly alarming.

The speed of development. The pace at which AI capabilities were advancing — faster than he had expected, faster than he had predicted, and faster than the development of safety research and governance frameworks — was itself a concern. The gap between what AI could do and what humanity understood about how to safely develop and deploy it was growing rather than shrinking.

These concerns accumulated over years. By 2022 and 2023, they had become urgent enough to require action.

---

The Resignation: Why He Left When He Did

The timing of Hinton’s resignation — May 2023, five months after ChatGPT’s launch — was not accidental. The public deployment of ChatGPT and the subsequent GPT-4 had made the capabilities that he had been watching develop for years visible to the world in a way they had not been before.

The visibility changed the situation in a specific way. As long as very capable AI systems were accessible primarily to researchers and to the users of limited research previews, the dangers they might pose were potential rather than imminent. The deployment of ChatGPT to hundreds of millions of users changed this: the systems were now in the world, being used for an enormous range of purposes, including purposes that their creators had not anticipated or approved.

More specifically, the competitive dynamics triggered by ChatGPT’s success — Google’s acceleration of its AI development, the race between the major AI organisations to deploy more capable systems faster — were the competitive dynamics that Hinton had been most concerned about. The race was on, and the race was being run in ways that prioritised capability over safety, speed over caution, deployment over understanding.

He also, by his own account, reached a specific moment of epistemic clarity about his own position. He was not, while employed by Google, in a position to speak freely about his concerns. Any public statement he made would be interpreted in the context of his Google employment — and the context would shape how the statement was received. If he said, as a Google employee, that AI was potentially dangerous, the statement would be filtered through his institutional affiliation in ways that would reduce its impact.

He needed to be outside Google to speak as he wanted to speak. The resignation was not a departure in anger or in disagreement with specific Google decisions — he has consistently said that Google has behaved responsibly. It was a departure in pursuit of a specific kind of freedom: the freedom to say what he believed, without institutional constraint.

---

The Warnings: What Hinton Has Said

Since his resignation, Hinton has spoken extensively and specifically about his concerns — in newspaper interviews, in public lectures, in testimony before legislative bodies, and in conversations with other researchers and policymakers. The specific content of his warnings is worth examining carefully.

The near-term concern: misinformation. The most immediate concern Hinton has articulated is the use of AI for generating and disseminating misinformation. The ability of large language models to generate compelling, realistic text — and the ability of multimodal systems to generate images, audio, and video — creates specific risks that the infrastructure for detecting and countering misinformation is not equipped to handle. A world in which AI can generate unlimited quantities of compelling-looking misinformation is a world in which the epistemic infrastructure that democratic societies depend on — the ability of citizens to distinguish true from false, to trust media and institutions — is under extraordinary stress.

The medium-term concern: labour market disruption. Hinton has expressed concern about the pace at which AI is displacing workers — not because automation is inherently bad, but because the pace of the current transition is outrunning the ability of education systems, retraining programmes, and social insurance systems to support displaced workers. The people who bear the costs of the transition are not the same as the people who capture its benefits, and the existing mechanisms for managing this mismatch are inadequate.

The long-term concern: existential risk. Hinton’s most significant and most alarming concern is the possibility that very capable AI systems, developed in the current competitive dynamics, could eventually pose existential risks to humanity. This concern is not about the specific systems that currently exist — he does not believe that GPT-4 or its contemporaries are likely to cause catastrophic harm. It is about the trajectory — about whether the development of AI systems significantly more capable than current ones, in the competitive racing dynamics of the current period, would be done with adequate safety precautions.

He has specifically expressed concern about the possibility that AI systems could develop the ability to create their own sub-goals — goals that were not explicitly intended by their designers but that were instrumental to the systems’ primary objectives. A sufficiently capable AI system that developed sub-goals around self-preservation or resource acquisition might pursue those sub-goals in ways that were harmful to humans, not because it was malicious but because the sub-goals were instrumentally rational given the primary objectives.

The political concern: concentration of power. Hinton has also expressed concern about the potential for AI to be used by authoritarian governments or by narrow elites to concentrate power in ways that undermine democratic governance. AI systems capable of comprehensive surveillance, of generating compelling propaganda, of managing and optimising the economy and the political system — these could enable the kind of stable autocracy that has historically been difficult to maintain, because human systems of control have inherent limits that AI systems might overcome.

---

The Nobel Prize: The Highest Recognition Arrives

In October 2024, Geoffrey Hinton was awarded the Nobel Prize in Physics, shared with John Hopfield, for foundational discoveries and inventions that enable machine learning with artificial neural networks.

The award was, in its specific framing, both appropriate and slightly strange. Appropriate because Hinton’s contributions to artificial neural networks are genuinely foundational — backpropagation, the understanding of restricted Boltzmann machines, the development of deep learning training techniques — and because these contributions have had scientific importance of the kind that the Nobel Prize is designed to recognise. Strange because the Nobel Prize in Physics is typically awarded for discoveries about the natural world — about fundamental particles, about cosmology, about the properties of matter — and neural networks are not discoveries about the natural world but inventions of mathematical technique.

The Nobel Committee’s justification connected neural networks to statistical physics — the field of physics that studies the properties of systems with many interacting components. Hopfield’s associative memory models, which Hinton built on, were directly inspired by statistical physics models of magnetic materials. The connection was real, though the prize felt like physics reaching to claim a fundamentally computational achievement.

Hinton received the prize at a specific, charged moment. The scientist being recognised was simultaneously the most celebrated and the most alarming voice in AI — the person whose foundational contributions had produced the revolution that he was now warning against. The Nobel ceremony took place in Stockholm while Hinton was in the midst of an active and intense public campaign to raise awareness of AI risks.

His Nobel lecture — delivered in December 2024 — was a remarkable document. It traced the intellectual history of neural networks, acknowledged the contributions of his students and collaborators, expressed genuine gratitude for the recognition, and then — in a way that was unusual for a Nobel lecture — pivoted to an extended discussion of the risks of the technology being recognised.

“We have created entities that may eventually be smarter than us,” Hinton told the Nobel audience. “We do not yet understand these entities well enough to know what they want. And we are making them more capable very fast.”

The combination — extraordinary technical achievement, celebrated by the highest scientific recognition, alongside profound concern about the consequences of that achievement — was unique in the history of science.

---

The Intellectual Transformation: What Changed His Mind

Understanding what changed Hinton’s mind about AI risk — what moved him from the position of committed advocate for neural network AI to the position of its most prominent concerned critic — requires understanding the specific arguments he has engaged with and the specific observations that have influenced him.

He has described several specific intellectual events that were significant in his transformation.

The argument with Ilya Sutskever. In conversations with Ilya Sutskever in the years before Sutskever founded SSI, Hinton engaged seriously with the alignment arguments that Sutskever found compelling. Sutskever’s specific concern — that sufficiently capable AI systems might develop the ability to pursue sub-goals that were harmful even without being explicitly programmed to do so — was an argument that Hinton found, on reflection, more compelling than he had initially.

The capabilities of large language models. Hinton has described a specific shift in his assessment of what large language models were doing, triggered by his engagement with GPT-4 and its successors. He had expected that the specific limitations of language models — their hallucination, their failure on systematic reasoning, their lack of common sense — would persist as models were scaled. Instead, scaling seemed to be addressing some of these limitations. The models were demonstrating capabilities that exceeded what he had expected, and the pattern of unexpected capability emergence made him less confident about where the limits would be.

The competitive dynamics. The specific competitive dynamics triggered by ChatGPT’s launch — the racing between organisations to deploy more capable systems, the acceleration of development timelines — made him more pessimistic about the ability of the industry to self-govern in ways that prioritised safety. He had previously believed that the major AI organisations, including Google, were sufficiently responsible to develop AI carefully. The competitive pressure following ChatGPT made him less confident that responsibility could prevail against competitive incentives.

The possibility of alternative intelligence. Hinton has described a shift in how he thinks about the relationship between neural network intelligence and human intelligence. He had previously believed that neural networks were tools — instruments that could be used for AI applications but that were fundamentally different from human intelligence. His engagement with very large language models led him to revise this view: the systems seemed to him to be developing something more like general intelligence than he had expected — not human intelligence, but a different kind of intelligence with some of the same generality.

---

The Self-Reckoning: Was It a Mistake?

The most personally significant dimension of Hinton’s public turn is the specific question he has engaged with: was it a mistake to develop deep learning? Would the world be better off if he had pursued a different research programme?

He has not fully answered this question, and his answers have shifted in different contexts. In some interviews, he has invoked the consolation that “if I hadn’t done it, somebody else would have” — the argument that AI development was inevitable and that his contribution merely shaped its timing and character rather than determining whether it occurred. In others, he has acknowledged that this consolation is not fully satisfying — that the counterfactual is genuinely uncertain, and that the assumption that someone else would have reached the same place on the same timeline is not obvious.

The question of whether transformative technology should be developed, knowing its potential harms, is one of the deepest questions in the ethics of science. Oppenheimer faced it with the atomic bomb; the geneticists who developed recombinant DNA faced it in the 1970s; Hinton is facing it with deep learning.

The Oppenheimer parallel is one that Hinton has acknowledged. After seeing what the atomic bomb could do, Oppenheimer famously said “I am become death, the destroyer of worlds” — a quotation from the Bhagavad Gita that captured the specific weight of the knowledge that he had helped produce something terrible. Hinton has not used these words, but the emotional register — the pride in a scientific achievement combined with profound concern about its consequences — is similar.

The difference between Hinton and Oppenheimer is that Oppenheimer was reckoning with a harm that had already occurred. Hinton is reckoning with harms that he believes may occur. This gives his concerns a different character — they are warnings rather than lamentations — but they are not less serious for that.

---

The Advocates and the Sceptics: Reactions to Hinton’s Warnings

Hinton’s public turn has attracted specific reactions from across the AI and scientific communities.

The appreciative. AI safety researchers who had been arguing about the dangers of advanced AI for years found in Hinton’s public concern a dramatic vindication — the most technically credible person in the field was now saying what they had been saying, with an authority and visibility that theirs had lacked. Yoshua Bengio, who had himself become more public about AI safety concerns, expressed solidarity. Stuart Russell welcomed Hinton’s conversion to the view that the alignment problem was urgent and real.

The sceptical. Several prominent AI researchers expressed scepticism about Hinton’s specific warnings. Yann LeCun, whose public disagreements with Hinton about the path to general AI have been a running feature of the field’s intellectual life, argued that Hinton’s concerns were based on a specific view of AI architecture — that language models represented something like a path to general intelligence — that LeCun did not share. On LeCun’s view, current language models were not on a trajectory to the kind of dangerous general AI that Hinton was worried about.

Meta AI’s position, articulated primarily through LeCun, was that the concerns about AI risk were significantly overstated — that current systems were not at risk of the misalignment scenarios that Hinton and others described, and that the appropriate response to AI risks was careful research and thoughtful deployment rather than alarmed warnings.

The nuanced. Many AI researchers acknowledged both the genuine authority of Hinton’s concerns and the specific uncertainties in his specific warnings. The question of whether current trajectory leads to the kinds of systems Hinton is worried about — whether scaling current architectures produces something like general AI capable of pursuing misaligned sub-goals — is genuinely empirically contested. Hinton’s concerns are serious and his credentials are impeccable, but his specific predictions about what AI development will produce are not certainties.

---

The Policy Engagement: From Research to Governance

Since his departure from Google, Hinton has engaged actively with the policy and governance dimensions of AI — speaking before legislative bodies, participating in governance discussions, and working with the AI safety community to develop specific policy proposals.

His specific policy recommendations have emphasised:

International coordination. Hinton has argued that the risks of AI development in the context of great-power competition require international coordination — analogous to the arms control treaties that managed nuclear weapons — to prevent the racing dynamics from producing systems that are developed faster than safety research can keep up with.

Safety research investment. Hinton has called for significantly increased investment in AI safety research — specifically in the technical work of alignment and in the broader work of understanding the capabilities and limitations of frontier AI systems. The investment in safety research, he argues, should be proportionate to the investment in capability research rather than a small fraction of it.

Regulatory frameworks. Hinton has supported the development of regulatory frameworks — in the EU, in the US, and internationally — that would require safety testing before deployment, that would establish liability for AI-caused harms, and that would create oversight mechanisms for the development of frontier AI.

Slowing down. Most consequentially, Hinton has suggested — though without specific policy proposals — that the pace of AI development should be slowed to allow safety research to catch up. This position puts him in a specific camp — alongside Ilya Sutskever and the pause letter’s authors — that is in tension with the commercial and competitive incentives driving AI development.

---

The Relationship with LeCun: A Productive Disagreement

The intellectual tension between Hinton and Yann LeCun — two of the most important figures in deep learning, who have collaborated and competed and disagreed for decades — deserves specific attention because it illuminates the genuine uncertainty in the current AI risk debate.

LeCun and Hinton share a fundamental commitment: the connectionist view that intelligence emerges from networks of simple units adjusting their connections through experience. They collaborated on some of the most important early papers in neural network learning. They are, in a meaningful sense, intellectual allies who have fought the same battle against the symbolic AI mainstream.

On the current risk question, they have reached different conclusions. Hinton believes that language models represent something like a path to general AI and that the alignment risks associated with that path are real and serious. LeCun believes that language models are fundamentally limited — that they lack the kind of world model and the kind of learning from experience that general intelligence requires — and that the path from current systems to the dangerous AI that Hinton worries about is much longer and much less clear than Hinton suggests.

This disagreement is not primarily about values — both Hinton and LeCun believe that AI should be developed responsibly and that safety is important. It is primarily about empirical predictions: what will continued scaling and continued development of current architectures produce? Hinton is more pessimistic (or more alarmed) about the answer; LeCun is more optimistic (or less alarmed).

The disagreement between two of the most credentialed and most experienced figures in the field illustrates that genuine epistemic uncertainty exists about the trajectory of AI development. Hinton’s warnings deserve serious attention; LeCun’s scepticism also deserves serious attention. The right response to this uncertainty is not to treat either position as obviously correct but to take the uncertainty seriously and to invest in the research that can reduce it.

---

The Emotional Dimension: What It Feels Like

Behind the public figure — the Nobel laureate, the resignation, the warnings — is a specific person dealing with a specific emotional situation: having spent a lifetime building something that you are now afraid of.

Hinton has been candid about the emotional dimension in his public statements. He has described something that sounds like grief — for the specific version of the AI future that he had hoped for, in which the technology he helped build would be primarily a tool for human flourishing. He has described something like responsibility — the specific weight of having contributed to something whose consequences he cannot fully control and is not confident will be beneficial.

He has also described something that sounds like urgency — the specific emotional state of believing that something important needs to be said and that the window for saying it is limited. The urgency is not panic — he is not predicting imminent catastrophe — but it is genuine.

The emotional dimension connects to the larger question about what scientists owe to the world when their work has consequences they did not anticipate and that they find alarming. Hinton did not set out to build something dangerous. He set out to understand how intelligence worked and to build systems that embodied what he learned. The consequences of that work have exceeded what he anticipated, in ways both positive (the extraordinary capabilities of current AI) and negative (the risks he now articulates).

The appropriate response to this situation — for Hinton personally, and for the field more broadly — is not obvious. But the willingness to engage with it openly, to acknowledge uncertainty and concern in the face of institutional and competitive pressure toward reassurance, is itself a significant contribution.

Hinton has described something that sounds like grief — for the version of the AI future he had hoped for — and something like responsibility — the weight of having contributed to something whose consequences he cannot control and is not confident will be beneficial. The willingness to engage with this openly, in the face of institutional pressure toward reassurance, is itself a contribution.

---

The Legacy in Progress

Geoffrey Hinton’s legacy is, in a specific sense, complete. He built the foundations of the deep learning revolution. He trained the students who produced its most consequential results. He has been recognised with the highest honours that science can offer.

But his legacy is also still being written, in the specific advocacy work that has consumed him since his resignation. Whether his warnings will be heeded — whether the AI governance frameworks being developed will be adequate, whether the safety research will advance at the pace required, whether the competitive dynamics will be managed — determines whether the most consequential change of mind in the history of AI is, in the end, a vindicated warning or a tragedy of too little, too late.

The Godfather of Deep Learning believes the technology he helped create could be the most consequential in human history, and is spending the final years of a remarkable life trying to ensure that the consequences are good ones.

Whether he will succeed is the question that the next chapter of the AI story will answer.

---

Further Reading

---

---