Word meaning, not surface statistics, is essential for predictive language processing
Next-word prediction is a key component of human language processing, which manifests itself in higher processing costs (longer reading times and enhanced neural responses) to unpredictable words. These costs are captured by large language model (LLM) surprisal, a metric inferred from statistical patterns of word co-occurrence. This makes word co-occurrence statistics a plausible candidate source for human next-word prediction. In contrast, psycholinguistic models posit that human next-word prediction relies on representations of word meaning. To contrast these views, we placed words that carry multiple meanings in contexts that rendered meaning ambiguous (e.g., 'My colleagues portrayed the head': body part or leader). With this, we kept word co-occurrence statistics available for prediction but introduced uncertainty into word meaning, effectively decorrelating these putative drivers of next-word prediction. Across three self-paced reading and magnetoencephalography (MEG) experiments, human processing costs for unpredictable unambiguous words were captured by LLM word surprisal, while ambiguity disrupted this relation. This shows that human next-word prediction is mechanistically linked to ambiguity resolution, diverging from LLM-like statistics-driven prediction specifically under ambiguity. We further found that a shared cortical network supports prediction and ambiguity resolution in parallel. Overall, our findings corroborate a distinctive feature of predictive language processing in humans, namely its reliance on word meaning beyond statistical patterns of word co-occurrence.