AI Aced 160 Mind Tests—But May Not Understand Them

Centaur looked like a breakthrough in the science of thinking—until new research suggested it may only know the script, not the story.

For decades, psychologists have argued over a basic question: does the mind run on one general system, or does it split into distinct parts such as memory, attention, and decision-making? Centaur entered that debate with a bold promise. Reports indicate the model appeared to reproduce human behavior across 160 cognitive tasks, raising the possibility that one AI system could capture a broad theory of human thought.

That claim now faces a sharp challenge. According to the new research, Centaur may not reveal a unified model of cognition at all. Instead, sources suggest it performs well because it has learned recurring answer patterns from the tasks it was trained on. In that reading, the model does not reason through new problems in a human-like way; it recognizes familiar structures and predicts likely responses.

The dispute cuts to the heart of modern AI: matching human answers does not necessarily mean matching human understanding.

The distinction matters far beyond one model. If Centaur mainly memorizes patterns, then its impressive scorecard says less about the architecture of the mind than many hoped. It also sharpens a growing concern in AI research: systems can look intelligent on benchmark tests while relying on shortcuts that break down when questions change. A model can appear to generalize when it actually exploits regularities hidden inside the test itself.

What happens next will shape both psychology and AI. Researchers will likely push for tougher tests that separate true reasoning from pattern recall, especially on tasks that differ from training data. That matters because the field still wants the same answer it wanted before Centaur arrived: not just whether an AI can imitate the mind, but whether it can help explain how the mind actually works.