Geoffrey Hinton

Geoffrey Hinton (born 6 December 1947) is a British-Canadian cognitive psychologist and computer scientist whose work on neural networks helped create the field now called deep learning. He shared the 2024 Nobel Prize in Physics with John Hopfield and received the Turing Award in 2018 with Yoshua Bengio and Yann LeCun 1.

His most influential contribution is the popularisation of backpropagation, the algorithm that trains multilayer networks by propagating errors backwards through the layers. The idea was not his alone — it had been discovered earlier and independently by several researchers — but the 1986 paper he wrote with David Rumelhart and Ronald Williams demonstrated it clearly and convinced a generation of researchers that learning internal representations was possible 2.

Since leaving Google in 2023 he has become one of the most prominent voices warning that artificial intelligence could become uncontrollable. That shift, from building the technology to warning about it, is the second act of a career that spans the field's entire modern history 3. This entry covers his education, his research contributions, the deep-learning revival, his years at Google, his risk warnings, and the criticism those warnings have attracted.

Hinton was born in London and studied experimental psychology at the University of Cambridge, graduating in 1970. He has said that he wanted to understand how the brain works and that psychology, as practised at the time, seemed unlikely to provide the answer; he turned to computation as a way of building models of mental processes 4.

He completed a PhD in artificial intelligence at the University of Edinburgh in 1978, working on neural networks at a time when the approach was deeply unfashionable. The dominant paradigm in AI was symbolic: intelligence as manipulation of explicit rules and representations. Neural networks, which learned from data, were regarded by many senior researchers as a dead end 5.

He held postdoctoral positions in the United States and Canada, including work at the University of California, San Diego, where he encountered the parallel distributed processing movement. That group argued that cognition could be modelled as patterns of activation across networks of simple units, an idea that became the foundation of connectionism 4.

He joined the University of Toronto in 1987 and has remained there, becoming a professor emeritus and a central figure in the Canadian AI community. Toronto became a training ground for a generation of researchers, including some who went on to found or lead major AI companies 5.

The 1986 Nature paper described how a network with hidden layers could be trained by computing the gradient of an error function with respect to each weight, using the chain rule to propagate errors backwards. The authors showed that the hidden units learned representations that were not specified in advance, which was the key claim: the network discovered its own features 2.

The algorithm was not new. Versions had been described in the 1960s and 1970s, and Paul Werbos had applied it in his doctoral thesis. What the 1986 paper did was make the idea legible and demonstrate it on problems that mattered, and it is the paper that most researchers cite when they describe the origin of modern deep learning 6.

Backpropagation still underpins the training of every large language model. The details have changed — better activation functions, adaptive optimisers, enormous parallel hardware — but the core procedure of computing gradients through a computational graph and updating weights is the same one described in that paper 6.

Hinton also developed Boltzmann machines and deep belief networks, generative models that could learn representations without labelled data. Those methods were partly superseded, but they established the idea that unsupervised pre-training could make deep networks trainable, an idea that influenced the later revival of the field 5.

The ImageNet competition of 2012 is usually treated as the moment deep learning became impossible to ignore. AlexNet, a convolutional network built by Alex Krizhevsky and Ilya Sutskever under Hinton's supervision, won the classification task by a margin of more than ten percentage points over the best classical system 7.

The result was not just a better score. AlexNet was trained on graphics processors, which made the computation practical, and it used techniques — dropout, rectified linear units, data augmentation — that became standard. The combination of a large dataset, a deep network, and GPU training defined the recipe that the field followed for the next decade 7.

Hinton's group had been working on the problem for years, and the win was less a sudden insight than the convergence of several trends: enough data, enough compute, and an algorithm that could exploit both. The paper is one of the most cited in computer science, and it marked the point at which industry began investing heavily in deep learning 6.

After 2012 the field moved quickly. Hinton's students founded and joined companies, and in 2013 Google acquired DNNresearch, a small company formed around Hinton and two of his students, and Hinton joined Google as a part-time researcher 4.

Hinton worked at Google from 2013 to 2023, dividing his time between the company and the University of Toronto. At Google he contributed to the development of neural network methods and to the research culture that produced the Transformer and other advances, though his direct role in large-scale product work was limited 4.

He remained an academic at heart, supervising students and publishing on topics such as knowledge distillation, a technique for compressing a large model's behaviour into a smaller one. That method is now standard in the deployment of models on devices, and it illustrates his continued interest in making neural networks practical as well as theoretically interesting 5.

He also became a public explainer of the field. Interviews and lectures presented neural networks as models of cognition rather than as engineering tricks, and he was often described as one of the "godfathers of AI" alongside Bengio and LeCun. The label was journalistic, but it reflected a real role: he had kept the approach alive when it was unfashionable and trained many of the people who built it 6.

In May 2023 he announced that he had left Google so that he could speak freely about the risks of the technology. He said he had not left to criticise the company, and that Google had behaved responsibly; the reason was that he wanted to discuss danger without it being attributed to his employer 3.

In interviews after his departure, Hinton said that his estimate of how quickly AI would become powerful had changed. He had expected human-level reasoning to be decades away, and he concluded that the systems he had helped create might soon surpass people in general tasks. He described the risk as serious enough to justify public warning rather than private concern 3.

His central argument is about control. Intelligence that is digital can be copied, run at high speed and combined across many instances, which gives it advantages over biological intelligence. If such systems are given goals and the ability to act, humans may find it difficult to constrain them, especially if the systems are better than people at planning and persuasion. He has also argued that the speed of digital replication makes oversight harder, because a model can be copied and run thousands of times while institutions change slowly 4.

He has also warned about nearer-term harms: disinformation, election interference, the automation of jobs, and the concentration of power in the companies that control the largest models. He has said that these problems are already visible, and that the longer-term risk of loss of control is the one he considers most important 3.

He has supported safety research and has argued that companies should be required to demonstrate that their systems are safe before deployment. He has been careful to say that he does not know how to solve the control problem, and that his purpose is to raise it rather than to claim a solution 1.

Hinton's stated position is that the development of superhuman AI is likely within his lifetime or shortly after, and that the probability of a catastrophic outcome is not negligible. He has given estimates in the range of one in ten to one in five, while saying that such numbers are guesses rather than calculations 3.

He distinguishes between the technology's potential and its trajectory. AI could improve medicine, education and scientific research, and he has said that the benefits are real. His concern is that the competitive dynamics between companies and states create pressure to deploy quickly, and that safety work is underfunded relative to capability work 4.

On policy he has favoured regulation, licensing and international coordination, and he has been sceptical of voluntary commitments. He has also argued that open-source release of powerful models is dangerous, a position that puts him at odds with parts of the research community that value openness 3.

His views on machine consciousness are distinctive. He has suggested that large language models may already have some form of subjective experience, an argument that most researchers reject, and he has used it to argue that the moral stakes of building such systems are higher than usually acknowledged. He has said that he does not expect a single moment of takeover, but rather a gradual transfer of decisions to systems that people no longer fully understand 4.

His risk warnings have been criticised as speculative. Yann LeCun, a fellow Turing laureate, has argued that the scenario of an AI system seizing power is implausible and distracts from concrete harms; Andrew Ng has said that warnings of extinction resemble a regulatory strategy for incumbents. Hinton has responded that his critics underestimate how quickly capabilities are improving 4.

Some of the criticism is about evidence. Predictions about superintelligence rest on extrapolation from current trends, and the mechanisms by which a system might escape human control are described in general terms rather than demonstrated. Hinton acknowledges the uncertainty but argues that the stakes justify caution even at low probability 3.

The timing of his warnings has also attracted comment. He left Google in 2023, as public attention to AI risk was rising, and his profile increased sharply afterwards. Supporters note that he had been discussing these issues privately for years and that he gave up a comfortable position to speak; critics suggest that the attention reinforced his later claims 4.

His scientific reputation is secure regardless. The backpropagation paper, the Boltzmann machine work, AlexNet and the training of a generation of researchers would be enough for a major career even without the Nobel. The controversy is about his second act, not his first 1.

The disagreement about AI risk is often framed as optimism versus pessimism, but the substantive dispute is about timelines and control. Hinton believes that systems with general capabilities are close enough that the control problem must be solved soon; others believe that current models lack the properties — persistent goals, embodiment, autonomy — that would make the scenario possible. Both sides agree that the field is moving quickly and disagree about what that implies 4.

The Nobel Prize was awarded for foundational work on neural networks, and some commentators questioned whether physics was the right category for computer science. The committee framed the work as physics: Hopfield's model drew on statistical mechanics, and Hinton used ideas from physics to describe how networks of simple units can store and retrieve patterns 1.

His claim about machine consciousness is the most contested of his recent statements. Most cognitive scientists regard current models as lacking the architecture required for experience, and Hinton's suggestion is presented as a possibility rather than a finding. He has not published a technical argument for it, which makes it a provocation rather than a result 3.

What is not disputed is the historical record. Hinton kept neural networks alive during a period when they were unfashionable, produced the paper that made backpropagation standard, supervised the work that produced AlexNet, and trained many of the researchers who built the current industry. The argument about his warnings will be settled, if at all, by events; the argument about his contributions is already settled 6.