Centaur knew the answers but not the questions: study undercuts AI "thinking" claim

The Centaur model announced a breakthrough: AI had supposedly learned to mimic human thinking across 160 cognitive tasks at once, from memory to decision-making. A new study rejects that claim: the model did not think, it memorized patterns.

An old debate, a new player

The nature of the human mind is one of the longest-running questions in psychology. Is thinking a single system with a shared "engine", or a set of specialized modules, each responsible for its own domain: memory, attention, language, spatial navigation, decision-making? This is not academic wordplay. The answer determines how to build tests of cognitive ability, how to treat disorders such as dementia or dyslexia, and how to design AI systems that claim to "understand".

After decades of debate, no consensus has emerged.

Centaur arrived as a possible answer from an unexpected direction. The idea: train an AI on data about human behavior in cognitive tests, then see whether the model reproduces the same behavior. If it does, it has "captured" the structure of thought. The authors claimed Centaur handles 160 different cognitive tasks with accuracy comparable to humans. The scale of the claim was rare: no earlier model had claimed such cognitive coverage.

Where the argument breaks down

Researchers who independently checked Centaur's work found a fundamental problem: its strong results come from memorization, not "understanding". The specific weaknesses in the methodology:

• Data overlap: the training set overlaps with the test set, so the model effectively "saw" the right answers before being tested
• No transfer: on unfamiliar tasks with the same structure, the model did not deliver stable results
• Pattern without mechanism: a high match with human answers does not mean the model reproduces cognitive processes, only their outward form
• Weak verification: the authors did not test on deliberately altered tasks where memorized answers would not work

According to critics, this is a classic case: the model knew what to answer without understanding the question.

Why it matters

Centaur's claim was not merely scientific; it was instrumental. If AI really models cognitive processes, it opens a new method for studying the mind: instead of expensive neuropsychological experiments, theories could be tested directly on an AI model. That would speed up cognitive science dramatically.

But the line between "memorization" and "thinking" in AI evaluation remains undrawn. Language models regularly post high benchmark scores, and on close inspection it turns out they are reproducing patterns from training data rather than generalizing.

"The model knew what to answer, but not why it should be that way" is roughly how critics sum up their objection to

Centaur.

The story is not new. Each time an AI earns a high score for "understanding", a detailed review follows and finds that the test measured something else. Centaur fits this pattern.

What it means

A convincing benchmark and real understanding are different things. Until standards emerge that fundamentally separate pattern memorization from cognitive generalization, claims about an "AI that thinks like a human" will remain in question. For cognitive science, the next step is obvious: design tests that, by definition, cannot be passed through rote memorization.