New benchmark reveals AI vision models fooled by prediction accuracy alone

A new paper from Ken Nakamura and colleagues at Japanese institutions (arXiv:2605.20127) challenges the standard practice of measuring AI vision model performance against the human brain. The authors argue that prediction accuracy—how well a model's internal representations predict neural activity—tells only half the story. Their proposed framework, Target-Space Recovery Profiles, first identifies reproducible response dimensions in fMRI data from repeated measurements. Then, instead of a single accuracy number, it quantifies which specific brain-response dimensions are recovered by a model's predictions.

Applying the framework to the Natural Scenes Dataset (8 subjects viewing natural images), the team found that early-to-intermediate visual-cortex responses form a low-dimensional set of reproducible dimensions. Brain-to-brain comparisons provided a human reference baseline. Critically, pretrained and randomly initialized models often achieved similar prediction accuracy but showed starkly different recovery profiles across these dimensions—meaning one model might match face-processing regions while another matches scene-processing regions, yet both score identically on standard metrics. This diagnostic lens reveals when a model is mimicking brain activity for the wrong reasons, offering a more rigorous path for aligning artificial and biological vision.