Your Brain Predicts Words Using Grammatical Chunks, Not Just Next-Word Guess Like AI

April 23, 2026 Dr. Michael Lee – Health Editor Health

Emerging neuroscience research challenges the prevailing analogy between human language processing and artificial intelligence systems, revealing that the brain employs a more nuanced, structurally grounded approach to predicting words during conversation than current large language models (LLMs). While LLMs operate on statistical next-word prediction trained on vast text corpora, the human brain integrates grammatical structure—specifically, syntactic constituents—into its predictive mechanisms, suggesting a fundamental divergence in how biological and artificial systems anticipate linguistic input. This distinction has implications not only for cognitive science but likewise for the development of more neurally plausible AI models and our understanding of language-related disorders.

Key Clinical Takeaways:

  • The human brain predicts words by first processing grammatical phrases or “chunks,” not just the immediate next word, unlike LLMs which rely solely on statistical patterns.
  • This hierarchical predictive strategy was demonstrated using magnetoencephalography (MEG) and behavioral tasks in Mandarin and English speakers, indicating a universal linguistic mechanism.
  • Understanding this difference may inform future AI design and improve diagnostic tools for conditions affecting language processing, such as aphasia or autism spectrum disorder.

The study, published in Nature Neuroscience and led by researchers at Fresh York University in collaboration with the Ernst Struengmann Institute for Neuroscience and Zhejiang University, directly tested whether human word prediction mirrors the next-token forecasting of LLMs like GPT-series models. Using magnetoencephalography to track real-time brain activity, participants listened to sentences in Mandarin Chinese while their neural responses were measured. Researchers manipulated sentence predictability using cloze probability and computational metrics of entropy and surprisal—quantifying how expected or unexpected a word was given prior context. Crucially, they compared these neural responses to predictions generated by LLMs operating on the same sentences.

Results showed that brain activity varied significantly depending on a word’s position within a grammatical constituent—such as a noun phrase or verb phrase—rather than merely its statistical likelihood of following the prior word. For instance, neural responses were stronger when a word violated syntactic expectations within a phrase, even if that word was statistically probable in isolation. This sensitivity to hierarchical structure was absent in LLM predictions, which treated all word positions uniformly based on local context. As David Poeppel, professor of psychology and neural science at NYU and senior author on the study, explained: “The brain doesn’t just guess what word comes next—it builds a scaffold of grammatical structure first, then fits words into that frame.”

The research was supported by grants from the National Institutes of Health (R01-DC005660) and the National Science Foundation (BCS-1828557), with additional funding from the Ernst Struengmann Foundation. According to the primary source, the study included 34 native Mandarin speakers in the core MEG experiment and validated findings using electroencephalography (EEG) data from 24 English-speaking patients undergoing neurosurgical monitoring at NYU Langone Health, ensuring cross-linguistic generalizability. Jiajie Zou, postdoctoral researcher at the Ernst Struengmann Institute and co-lead author, noted in a follow-up interview: “What’s striking is that the brain’s sensitivity to phrase structure emerges rapidly—within 200 milliseconds of hearing a word—suggesting this is not a slow, deliberative process but an intrinsic feature of online language comprehension.”

These findings carry potential relevance for clinical neurology and speech-language pathology. Conditions such as Broca’s aphasia, characterized by difficulty in syntactic processing despite relatively preserved lexical access, may reflect a disruption in this particularly mechanism of hierarchical prediction. Similarly, emerging research suggests that individuals with autism spectrum disorder may exhibit atypical neural responses to syntactic violations, possibly indicating differences in how linguistic chunks are generated or maintained during real-time comprehension. Clinicians specializing in cognitive neurology or language disorders could leverage this insight when assessing patients with unexplained language comprehension deficits, particularly when standard aphasia batteries fail to capture subtle syntactic processing delays.

For patients experiencing persistent difficulties in following conversations, especially in noisy environments or when processing complex sentences, it may be beneficial to consult with specialists who evaluate higher-order language processing. Referral to a licensed speech-language pathologist with expertise in cognitive-communication disorders can help determine whether deficits lie in auditory processing, lexical access, or syntactic integration. Neuropsychologists trained in assessing executive contributions to language—such as working memory and cognitive flexibility—may offer further clarity, particularly when symptoms overlap with attentional or frontal lobe dysfunction. Accessing vetted professionals through board-certified neuropsychologists ensures that evaluations are grounded in current neuroscientific models of language processing.

From a research perspective, this work underscores the limitations of using LLMs as direct models of human cognition. While AI systems excel at pattern recognition in linguistic data, they lack the innate bias toward hierarchical structure that appears to be hardwired in the human brain. Future iterations of language models might benefit from incorporating syntactic priors—architectural constraints that encourage phrase-level representations—potentially improving both performance and interpretability. As Nai Ding, professor at Zhejiang University and former postdoctoral fellow in Poeppel’s lab, emphasized in a recent seminar: “We’re not saying the brain doesn’t use statistics—it clearly does. But it uses statistics in service of structure, not the other way around.” This insight could guide the development of hybrid models that better capture the brain’s balance between statistical learning and rule-based computation.

this research reinforces that human language comprehension is not merely a statistical guessing game but a deeply structured, incremental construction of meaning. By recognizing the brain’s reliance on grammatical constituents as a predictive scaffold, clinicians, researchers, and AI developers alike can move toward more accurate models of how we understand—and are understood by—others in real time.

*Disclaimer: The information provided in this article is for educational and scientific communication purposes only and does not constitute medical advice. Always consult with a qualified healthcare provider regarding any medical condition, diagnosis, or treatment plan.*

,