AI hallucinations are, according to Geoffrey Hinton, not hallucinations at all.
The Nobel Prize-winning computer scientist and “Godfather of AI” has offered one of the most illuminating explanations yet for why large language models generate false information — and his argument is striking precisely because it implicates human cognition just as much as machine intelligence. In a recent discussion, Hinton reframed the entire debate around a single, better word: confabulation. His remarks cut to the heart of something researchers and businesses have wrestled with since the rise of generative AI: why do these systems confidently say things that are simply untrue?
“They shouldn’t be called hallucinations. They should be called confabulation, if it’s with language models. Confabulation, better known as lies. Lies. Psychologists have been studying them in people since at least the 1930s, and people confabulate all the time — at least I think they do. I just made that up.”
That last line — delivered with characteristic wit — is itself a demonstration of confabulation in action. Hinton’s point is that human memory is not a recording device. It does not retrieve stored files the way a computer retrieves data from a hard drive. Instead, it reconstructs.
“If you remember something that happened recently, it’s not that there’s a file stored somewhere in your brain — like in a filing cabinet or in a computer memory. What’s happened is recent events changed your connection strengths. And now you can construct something, using those connection strengths, that’s pretty like what happened a few hours ago or a few days ago.”
This is precisely the mechanism that underlies neural networks, including the transformer-based models powering today’s AI systems. They do not store facts; they encode patterns of connection across billions of parameters. When asked a question, they generate a plausible response by drawing on those patterns — the same way a person draws on neural connection strengths shaped by experience.
The uncomfortable implication, Hinton argues, becomes clearest when memory is tested over longer timescales.
“If I ask you to remember something from a few years ago, you’ll construct something that seems very plausible to you, and some of the details will be right and some will be wrong — and you may not be any more confident about the details that are right than about the ones that are wrong. Now, it’s often hard to see that, because you don’t know the ground truth.”
That final observation is the crux of the problem for both humans and AI systems. Confidence is not a reliable signal of accuracy. A language model, like a person recounting a memory from years ago, produces its output with the same generative process regardless of whether the details are correct. The model does not know what it does not know.
This framing has significant implications for how businesses and developers think about AI reliability. The instinct has been to treat hallucination as a flaw to be patched — a bug in the system. Hinton’s argument suggests it may be something more fundamental: an emergent property of any system that learns by adjusting connection strengths rather than by storing and retrieving discrete facts. Cognitive scientist Elan Barenholtz has made a similar case, arguing that memory is generation, not retrieval — a principle that applies to both biological and artificial neural networks.
Not everyone views this as purely a problem. Demis Hassabis has suggested that hallucinations can be valuable, and that the same generative tendency driving errors also drives creativity — a trade-off that may be inherent to how these systems work. At the same time, the practical stakes remain high. A recent study found over 100 hallucinated citations in papers submitted to NeurIPS 2025, one of the field’s most prestigious conferences — a sign that confabulation is far from a solved problem even in rigorous academic settings. On the benchmarking front, some models are performing meaningfully better than others, suggesting that while confabulation may be structural, its frequency can still be reduced.
Hinton’s contribution to this conversation is to give the phenomenon its proper name — and in doing so, to remind us that the problem predates artificial intelligence by decades. Confabulation is not a machine failing. It is what happens when any system generates rather than retrieves. The question for the AI industry is not how to eliminate a bug, but how to build systems that know, as reliably as possible, the difference between what they know and what they are constructing.