A new study reveals how human reading diverges from AI language models, even though both rely on next-word prediction at first. Researchers found that while LLMs can predict initial word recognition speed, they fail to explain why we reread complex sentences.
The Research
Led by William Timkey from New York University and Brian Dillon from the University of Massachusetts Amherst, the study published in Proceedings of the National Academy of Sciences compared eye-tracking data from 368 adult readers to predictions from over 400 neural network language models. They used syntactically complex texts, including “garden-path” sentences like “The old man the boat.”
Results showed that LLMs accurately model early word recognition during smooth forward reading. But they severely underpredicted processing time needed to integrate words into context, especially when ambiguity arises. In fact, regressive eye movements—when our eyes go back to reread—account for about 20% of reading fixations. Standard autoregressive LLMs lack an equivalent mechanism for this retrospective re-parsing.
“Language models develop their remarkable language understanding capabilities by being trained to predict the next word,” Timkey explained. “We found that LLMs can explain how long it takes people to recognize words when their eyes move smoothly forward, but they fail to capture cases where people have difficulty integrating a word into larger context.”
Why It Matters
This research gives insight into our own cognitive architecture. It shows that human understanding isn't just a one-way flow of predicting words; it involves dynamic, multi-stage integration that AI hasn't mastered yet. Understanding this gap could lead to better AI and also help us understand reading disorders.
What You Can Do
Next time you find yourself rereading a sentence, know that your brain is doing something complex that AI can't replicate. To strengthen this ability, try reading more challenging material and consciously summarizing what you read—this exercises your context-integration skills.
Source: Neuroscience News
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
