Your Brain Predicts Words Differently Than AI: New Study Reveals Grammar-Driven Process

A new study reveals that the human brain predicts upcoming words not by guessing the next word in isolation, but by analyzing grammatical structures and word groupings, a process that differs fundamentally from how artificial intelligence language models operate.

The research, published in Nature Neuroscience, shows that when people listen to spoken language, their brains do not simply forecast the most likely next word based on immediate context. Instead, they first identify grammatical constituents—such as phrases or clauses—and then determine which words fit best within those larger structures.

This approach contrasts sharply with large language models (LLMs), which generate predictions by treating each word in context with equal weight, without distinguishing between different types of linguistic groupings.

“While LLMs are trained and optimized to predict the next word, the human brain makes predictions by grammatically grouping words into phrases,” explains David Poeppel, professor of psychology and neural science at New York University and a coauthor of the study.

“With LLMs, predictions are by and large created equally: each word exploits its predictive context the same way. By contrast, the human brain makes predictions by first taking into account chunks of words—what we call grammatical constituents—and then determining which words are predicted best within that structure.”

The study involved experiments with Mandarin Chinese speakers, using magnetoencephalography (MEG) to measure brain activity while participants listened to sentences. Researchers also conducted behavioral Cloze tests, in which specific words were removed from passages and participants were asked to fill in the blanks, to assess predictive language processing.

To quantify predictability, the researchers used LLMs to calculate two metrics: entropy and surprisal. High entropy indicates a context with many possible continuations (e.g., after “I saw a,” many objects could follow), while low entropy suggests stronger constraints (e.g., after “I sat on a,” fewer objects are plausible). High surprisal occurs when a word is unexpected given the context (e.g., “cat” after “I sat on a” is more surprising than after “I saw a”).

By comparing brain responses to LLM predictions, the team found that neural activity varied depending on a word’s position within a grammatical constituent. This variation indicated that the brain’s predictions are sensitive to syntactic structure, not just linear word sequences.

In contrast, LLMs showed no such sensitivity—they generated uniform predictions regardless of whether a word appeared inside a noun phrase, verb phrase, or other grammatical unit.

To test whether the findings applied beyond Mandarin, the researchers analyzed brain data from patients exposed to English sentences. The results confirmed that the same predictive mechanism—based on grammatical constituents—operates across languages.

“Our brains can, like AI systems, exploit next-word prediction. However, brains are highly sensitive to linguistic constituent structure,” concludes Poeppel.

“This research shows that next-word prediction is balanced and modulated by our consideration of grammatically organized ‘chunks of words’—quite different from how LLMs work.”

The study underscores that while both human brains and AI systems engage in predictive language processing, the underlying mechanisms are distinct. Human prediction is hierarchically structured, integrating syntactic knowledge to guide expectations, whereas AI relies on statistical patterns across word sequences without explicit grammatical parsing.

These findings have implications for understanding the neural basis of language comprehension and may inform future efforts to develop more biologically plausible models of artificial intelligence that better mimic human linguistic processing.