A recent study using magnetoencephalography (MEG) and behavioural “cloze” tests with Mandarin and English speakers compared brain responses to word predictability against LLM predictions (Poeppel et al. 2026). The core finding shows a clear divergence: while LLMs predict uniformly across all sequences, the human brain’s responses varied systematically depending on a word’s position within hierarchical grammatical constituents. Human listeners first group words into structured units and only then generate predictions within those units. This difference is not merely technical; LLMs predict continuously using statistical patterns without an intermediate structure-building stage. Humans are biologically constrained to evaluate statistical patterns only within a priori determined structures that are hierarchical in their relationship. The article notes that Geoffrey Hinton has claimed Natural Language is inherently associative and its structural constraints are entirely predictable through next-token statistical evaluation, a claim described as entirely unfounded based on decades of research in Generative Grammar, Biolinguistics and Evolutionary Biology. For a multilingual society, conflating LLM usefulness with genuine understanding carries real costs, particularly in education.