Why language models predict brain activity across all languages

Neural encoding models built from large language models reliably predict brain activity during story listening across three typologically different languages, with spatial alignment patterns that remain consistent whether comparing English to Mandarin or examining deeper model layers. The surprise: this alignment doesn't come from the two leading computational candidates—predictive uncertainty (how surprising each word is) or information compression (how efficiently the model represents meaning). Instead, the alignment appears to arise from simpler lexical-semantic correspondences that transfer across languages, suggesting LLM-brain alignment is more about shared vocabulary-level features than mirrored hierarchical computation.