A new study published in Nature Neuroscience challenges the common assumption that the human brain predicts words in the same manner as large language models (LLMs). While LLMs function by calculating the statistical probability of the next word based solely on immediate context, researchers at New York University and collaborating institutions have found that the human brain employs a significantly more complex process. The research indicates that humans do not merely predict the next word in isolation. Instead, the brain analyzes larger linguistic structures known as grammatical constituents, which are groups of words functioning as a single unit. This approach is similar to solving a puzzle by considering the surrounding pieces and the overall image, rather than just guessing the single piece that fits in the next slot. Co-author David Poeppel from New York University explained that while LLMs treat every word's predictive context equally, the human brain first organizes words into phrases and then determines predictions within that grammatical framework. To reach these conclusions, the study team, including researchers from the Ernst Struengmann Institute and Zhejiang University, conducted a series of experiments. Participants were exposed to Mandarin Chinese sentences while their brain activity was measured using magnetoencephalography (MEG). The team also utilized behavioral Cloze tests, where specific words were removed from sentences for participants to fill in, to assess linguistic prediction. Additionally, the researchers analyzed data from patients exposed to English to confirm that these findings apply across different languages. The study quantified word predictability using two metrics derived from LLMs: entropy and surprisal. High entropy implies that the context allows for many possible following words, reducing predictability, while high surprisal indicates that a word was unexpected given the context. If the human brain operated exactly like an LLM, brain activity responses would correlate uniformly with LLM predictions. However, the results revealed significant variance, showing that brain responses were sensitive to the linguistic structural position of words. The findings demonstrate that while the brain can utilize next-word prediction capabilities similar to AI, it heavily modulates this process by considering grammatical organization. The brain actively groups words into constituents before making a prediction, a step that current LLM architectures do not require or replicate. This distinction highlights a fundamental difference between artificial intelligence and human cognition. LLMs generate predictions based on sequential data patterns, whereas the human brain integrates structural grammar to anticipate language. Ultimately, the study concludes that next-word prediction in humans is not a simple linear calculation but a nuanced process balanced by an understanding of sentence structure. This research suggests that while AI systems mimic certain aspects of human language processing, they lack the deep structural sensitivity that characterizes human linguistic prediction.