There is no shortage of definitions of AGI, but Google DeepMind CEO Demis Hassabis has come up with one of the most interesting ones yet.
Hassabis, who has spent two to three decades working on AI and holds a PhD in neuroscience, recently laid out his definition of AGI — and why he believes today’s systems, impressive as they are, still fall well short of it. His framing is rooted in biology, benchmarked against the human brain, and capped with a thought experiment that raises the bar considerably for what true general intelligence would look like.
“My definition of AI has never changed,” Hassabis said at the India AI Summit. “We’ve always defined it as a system that can exhibit all the cognitive capabilities humans can. The brain is the only existence proof we have — that we know of — of a general intelligence. That’s also partly why I studied neuroscience, because I wanted to understand the only data point that we have that this is possible.”
He was quick to acknowledge that this is a high bar. “If you wanted to test a system against that, it would have to be capable of all the things humans can do with this brain architecture, which is incredibly flexible. It’s clear today’s systems, although they’re very impressive and they’re improving, don’t do a lot of those things. True creativity, continual learning, long-term planning — they’re not good at those things.”
Hassabis also pointed to what he called a lack of general consistency as a key limitation. “What is missing is general consistency across the board in capabilities. In some circumstances they can get gold medals in international Maths Olympiad questions, like we did last summer with our systems. But they can still fall over on relatively simple maths problems if you pose it in a certain way. That shouldn’t happen with a true general intelligence. It shouldn’t be a jagged intelligence like that.”
He then offered a striking test for what genuine AGI might look like in practice. “The kind of test I would be looking for is training an AI system with a knowledge cutoff of, say, 1911, and then seeing if it could come up with general relativity, like Einstein did in 1915. That’s the kind of test I think is a true test of whether we have a full AGI system. I think we’re still a few years away from that, but I think it’s going to be possible eventually. It’s clear today’s systems couldn’t do that.”
The framing is significant. Hassabis is essentially arguing that AGI is not about doing known tasks well, but about the capacity to generate genuinely new knowledge — to reason from existing information to discoveries that have never been made before. There are strides being made in this regard — AI systems are taking their first steps towards solving Erdos problems, and helping researchers prove novel theorems. This distinction matters because it shifts the conversation away from benchmark performance, where AI has made dramatic progress, toward something far harder to manufacture: scientific imagination. It is also a self-aware standard from someone whose own company, Google DeepMind, made headlines with AlphaFold’s breakthrough in protein structure prediction and with AlphaGeometry’s strong performance on Olympiad-level geometry problems. Those achievements, remarkable as they are, Hassabis seems to be saying, are still not enough. The AGI threshold, by his measure, remains ahead of us — and getting there will require systems that do not just recall and recombine, but genuinely discover.