Geoffrey Hinton: The godfather of deep learning who walked out to warn us

Geoffrey Hinton was born in London in 1947, into a family of scientists stretching back to the mathematician George Boole. He read experimental psychology at Cambridge, graduating in 1970, worked briefly as a carpenter, and in 1972 began a PhD at the University of Edinburgh under Christopher Longuet-Higgins, then one of the few places in Britain studying artificial intelligence. He chose to work on neural networks at almost exactly the moment the field's establishment, exemplified by Marvin Minsky, had declared them a dead end.

He stayed with the idea anyway. After postdoctoral work in the United States and a stint at Carnegie Mellon, he moved to the University of Toronto in 1987, where he has remained ever since, later helping to found the Vector Institute and serving as its chief scientific adviser. With Terrence Sejnowski and David Ackley he invented the Boltzmann machine in the mid 1980s, a learning method the Nobel committee would cite four decades later. In 1986, with David Rumelhart and Ronald Williams, he published the Nature paper that demonstrated backpropagation could train multi-layer networks to discover useful internal representations, the result on which the modern field rests.

For twenty years the backpropagation result was respected but unfashionable, because the hardware and data did not yet exist to make it work at scale. Hinton's contribution in the 2000s was to keep the research program alive until they did. His 2006 work on deep belief networks, with Simon Osindero and Yee-Whye Teh, showed that deep networks could be trained layer by layer, and the phrase deep learning entered the vocabulary.

The vindication arrived in 2012, when his graduate students Alex Krizhevsky and Ilya Sutskever, working with him, entered the ImageNet competition with a deep convolutional network, now known as AlexNet, that cut the error rate on image classification roughly in half relative to the best conventional systems. Google acquired the trio's startup, DNNresearch, in 2013, and Hinton spent a decade at Google as a vice president and engineering fellow. His former students and collaborators seeded much of the industry: Sutskever went on to co-found OpenAI. In 2019 Hinton received the 2018 Turing Award alongside Yann LeCun and Yoshua Bengio, and in October 2024 he shared the Nobel Prize in Physics with John Hopfield for the foundational discoveries behind machine learning.

In May 2023 Hinton resigned from Google, telling the New York Times he wanted to speak freely about the dangers of AI without having to consider how his words reflected on his employer. He was careful to say Google had acted responsibly, and he was equally careful about what he was claiming: not that catastrophe was certain, but that the pace of progress had outrun his own expectations and that a part of him now regretted his life's work.

His reasoning is technical, not mystical. Biological brains are tied to mortal, power-hungry hardware, so knowledge dies with the organism. Digital networks can be copied exactly, and thousands of copies can share what each one learns by averaging their weight updates, which means digital intelligence can accumulate knowledge at a rate no biological system can match. Chatbots trained on human text convinced him that this advantage had begun to tell. By late 2024 he was putting a number on the downside, estimating a 10 to 20 percent chance that AI could lead to human extinction within the coming decades, an estimate he describes frankly as intuition rather than measurement.

It is easy to misfile Hinton alongside the field's internal critics, but his position is nearly the opposite of Gary Marcus's or Yann LeCun's. He thinks large language models genuinely understand what they say, that the old linguistic argument about statistical parrots misunderstands how prediction forces learning, and that scaling works. His alarm is built on capability, not incapacity: he fears these systems because he believes they work, and will work better.

Since leaving Google he has directed his influence toward mitigation. He signed the 2023 statement that mitigating extinction risk from AI should be a global priority, has argued that a fraction of AI research budgets should go to safety, and in 2025 began advocating for what he calls maternal AI, building systems with something analogous to a mother's drive to protect her offspring, on the grounds that the only known example of a more intelligent entity caring for a less intelligent one is a mother and her child.

Hinton has been right about the two biggest questions of his career. He was right that neural networks were the correct path when nearly the entire field believed otherwise, and he was right that scale would unlock capabilities researchers did not anticipate. Both judgments looked wrong for years before they looked obvious, which is worth remembering when his current warnings are dismissed as an old man's anxiety.

The contested ground is real. His extinction probability is an intuition, and he says so. His timeline estimates have moved around, and he has been wrong about specific forecasts before, including a 2016 claim that radiologists would soon be obsolete. The fair reading is that his scientific track record entitles his warnings to be treated as evidence rather than noise, while his policy numbers deserve the same skepticism as anyone else's. That combination, the authority and the humility, is precisely what makes him the risk debate's most credible witness.