AI and the World's Hard Problems · Linguistics · arXiv · 2026-10-19
The Score Is Not the Structure: Brain Alignment and Cross-Lingual Transfer
Similarity scores are often cited as evidence that a model shares structure with the brain or across languages, and the authors examine what these scores show once that structure is removed or the measuring instrument is faulty. Across seventeen languages, a grammaticality probe transferred worse between more distant languages, but the probe's own accuracy fell along the same axis and was at chance in four languages. In brain alignment, a training objective raised similarity to fMRI responses from 0.10 to 0.34, yet targets with the brain correspondence destroyed still scored 0.31, so the authors detected no distance-graded steering effect and no syntactic benefit from alignment.