eLife · 23 October 2026 · From issue 57
Large language models poorly capture how word order shapes representations of sentence meaning in the human brain
AI models fail to copy how brains track who did what to whom in sentences.
James Fodor, Carsten Murawski et al.
AI language models still get word order wrong when matching brain activity. Scientists scanned the brains of 30 people as they read 108 sentences built to test whether meaning depended on who did what to whom. Eight AI models that track word order matched brain patterns better than simple models that ignore order. But all eight still failed badly on sentences where knowing the doer and the receiver of an action mattered most.
The scale of the brain scan study used to test AI language models against human brain activity.
One brain scan study, 30 people, lab setting only.
Integrity screen: passed (3 checks) Checked 26 October 2026. Retraction record: none. DOI resolves at doi.org. Metadata record found (eLife). Read the source
This screen checks process signals. It does not tell you the findings are true.