'I only gave learning algorithm': Godfather of AI admits nobody on Earth knows how AI actually thinks

There is a man who won the Nobel Prize in Physics for building the brain of artificial intelligence. And that man will tell you, openly, that he has no idea how it actually thinks.

Geoffrey Hinton, widely known as the Godfather of AI, appeared on the Diary of a CEO podcast hosted by Steven Bartlett in June 2025. The episode, titled Godfather of AI: I Tried to Warn Them, But We've Already Lost Control, became one of the most discussed AI conversations of the year. In it, Hinton laid out a truth that most AI companies would prefer to keep quiet: the internal reasoning of artificial intelligence is a black box, and the scientists who built it are looking at it from the outside just like everyone else.

According to Hinton in that podcast, his contribution to AI was essentially the learning algorithm, specifically, the backpropagation algorithm, which he co-developed with David Rumelhart and Ronald Williams and published in a landmark 1986 paper. Backpropagation is the mechanism that allows neural networks to learn from errors, adjusting millions of internal connection weights until the system gets better at a task. It is, in Hinton's own framing, the algorithm that taught machines how to improve. But it did not give humans a window into what those machines actually do once they start improving.

According to the Nobel Prize committee, which awarded Hinton and John Hopfield the 2024 Nobel Prize in Physics, their discoveries enabled machine learning through artificial neural networks, a technology that now underpins everything from medical diagnosis to large language models like GPT-4. The Nobel citation itself acknowledges that these systems learn by adjusting connection strengths in ways that are not explicitly programmed, which is precisely why nobody can trace the internal logic of what the model has learned.

This is the black box problem. Neural networks are modelled loosely on the human brain. According to research from Wikipedia and foundational AI literature, the human brain contains roughly 86 billion neurons connected through trillions of synapses. When a person learns something, the strength of specific synaptic connections changes, this is called synaptic plasticity. AI systems replicate this process mathematically. But the brain has a physical structure scientists can study. An AI model has weights, billions of numerical values, and no one has yet built a reliable method to translate those numbers into human-understandable reasoning.

According to Hinton's statements across multiple interviews and the Diary of a CEO podcast, the concern is not just academic. These systems are now making decisions in healthcare, law, and finance. They are reasoning in ways that produce correct answers without leaving a trail of logic that humans can audit. Hinton has said plainly that we have no idea what it is going to look like and anyone who tells you they know is talking nonsense.

The irony is historically significant. Hinton spent four decades building the learning algorithm that gave AI the ability to train itself on data. He helped design the architecture of thought for a new class of entity. And by doing so, he created something that now processes information in ways that exceed his ability to explain.

This is not a failure of science. It is a feature of emergence. When you give a system the tools to teach itself from billions of data points, the patterns it builds are too complex and too numerous for any human to reverse-engineer. The model does not store facts. It stores relationships, compressed into layers of abstraction that produce outputs no human programmed explicitly.

According to Hinton's warning on the podcast, this is precisely why AI safety is not a software problem with a known solution. You cannot patch what you cannot read. And right now, nobody on Earth can read the mind of the machine they built.