AI Models Self-Organize Like the Brain. That's a Design Clue, Not a Human Mind.

A researcher runs a sentence through a large language model and watches which internal neurons activate. Then a math problem. Then a social dilemma. Then a description of how a person would catch a falling glass. Across 46 such tasks, the activation map starts to look familiar: the model's specialized circuits cluster the same way the human brain does, with separate regions handling language, formal reasoning, social reasoning, and physical reasoning.

That convergence is the finding at the center of a new arXiv preprint from Pengrui Han and colleagues, "Modular Cognitive Architecture Emerges in Large Language Models". It is a single study, not yet peer-reviewed, and the authors do not claim AI is becoming human. They argue something narrower and arguably more interesting: that two completely different optimization regimes, biological evolution and gradient descent on a transformer, may have arrived at the same architectural answer because that answer is a load-bearing design principle of any intelligent system.

The team ran N=46 tasks across four cognitive domains and identified which neurons in the model were most active for each task. Tasks that draw on the same network in the human brain recruited overlapping neurons in the model. Tasks that draw on different networks recruited distinct neurons. The overlap pattern is the load-bearing detail. It is also the part the paper calls "mirrors the human brain," and that word choice is the authors'.

The result is that the model self-organizes without anyone telling it to. No one labeled the four cognitive domains during training. No one wired the sub-networks by hand. Gradient descent discovered the partition on its own, the same way evolution arrived at it in biological nervous systems. The authors call this "convergent emergence" and argue it suggests modularity "may be a fundamental property of intelligent systems." That is a strong interpretation. It is also the part of the paper most likely to be flattened by wire coverage into "AI thinks like us."

The study covers one model class and a modest task battery. There is no independent replication in the source set, and the "mirrors the human brain" claim is an overlap in task-recruiting neurons rather than a settled mapping between the two systems. The four-domain, 46-task battery is the credibility number: broad enough to test the claim, narrow enough to leave room for competing explanations, and small enough that the result should be read as suggestive rather than proven.

The author's project page and code release make the circuit-analysis setup reproducible. Anyone with the right tooling can rerun the activation mapping and check whether the same overlap pattern shows up in a different model class, with a different task set, or under a different training objective. Until that happens, the right reading of the result is the design-principle one: modularity may be what any system that has to do many kinds of cognition efficiently is forced into, not a sign that AI is on a path toward human-like cognition.

The next milestone is replication. If the overlap pattern holds across model classes, training objectives, and a wider task battery, the case for modularity as a general design principle gets stronger. If it does not, "convergent emergence" becomes a claim about a specific training regime rather than a general principle. The team's setup is reproducible enough for that second result to land within a year.