LLMs trained to be creative match how human brains think during creative work

Researchers scanned 170 people doing creative tasks and compared their brain activity to outputs from different language models. Models trained specifically for creative thinking aligned much more closely with the brain patterns of creative thought — especially early in the process — while models trained for reasoning or logic actually moved away from creative brain patterns.

This is the first evidence that you can measure whether an AI model's internal structure actually resembles how human brains work during creative thinking. Right now, companies building AI for creative tasks have no neuroscientific ground truth to aim at — they just optimize for output quality. This work suggests you can do better: train models to match the geometry of creative neural responses, not just produce good ideas. The finding also reveals that different training objectives reshape what a model does internally in measurable ways. Chain-of-thought training, which forces analytical step-by-step reasoning, actively pushes models away from creative neural patterns. That's not a limitation — it's a trade-off you can now see and measure.