Why Two AI Brains Can Look Alike Yet Think Differently

When scientists want to know whether an AI 'thinks' like a human brain, they often peek inside both and compare the patterns of activity. If the patterns line up, the assumption goes, the two systems must be doing the same job. A team of researchers publishing at ICML, a major AI conference, tested that assumption with mathematics — using simple networks they could fully solve — and found it doesn't hold up.

In their experiments, internal patterns and actual behaviour came apart. Two networks could produce almost identical internal activity while solving a problem in genuinely different ways, or solve the problem the same way while their insides looked nothing alike. Doing well on tests didn't help either: networks that scored highly, or that shrugged off messy, noisy inputs, were not pushed into using task-specific internal patterns.

Only one thing locked networks into a shared, task-specific style: robustness to damage in their own connections. If you knock out or scramble some of the links inside a network and it keeps working, it has to develop a specific, meaningful internal language. Networks that merely tolerate noisy inputs can stay flexible inside. That is a real, practical distinction for anyone building or testing AI systems.

The catch: this is theory, done on deliberately simple networks, not the giant models behind ChatGPT. So it doesn't prove real products behave this way — it points at where to look next. The takeaway for everyone else is healthy scepticism. When a company or a paper claims an AI's inner patterns 'match the human brain', that resemblance alone doesn't prove it works like a brain. Similarity is a clue, not proof.