Why Humans Reread Sentences That AI Finds Easy

Some sentences seem to glide through the mind, while others bring reading to a sudden halt. The eyes pause, retreat, and scan the words again as the brain struggles to determine what a sentence actually means. New research from New York University and the University of Massachusetts Amherst suggests that artificial intelligence can explain part of this process—but not the moment when language becomes structurally confusing.

The study, published in the Proceedings of the National Academy of Sciences, compared human eye movements with predictions generated by more than 400 artificial intelligence language models. The researchers found that humans and AI behave similarly during the earliest stage of reading, when readers identify a word from its sequence of letters. Both systems are strongly influenced by what word is likely to appear next. But when readers must integrate that word into the evolving structure and meaning of a sentence, the similarities break down.

Large language models are trained primarily through next-word prediction. During training, they process enormous quantities of text and learn statistical patterns that allow them to estimate which word is most likely to follow a given sequence. The more predictable a word is in context, the less processing effort it generally requires from both a language model and a human reader. This apparent overlap has made AI models useful tools for testing theories about how people understand language.

“Language models develop their remarkable language understanding capabilities by being trained to predict the next word in a sentence,” said William Timkey, a linguistics doctoral student at NYU and lead author of the study. “That led us to ask whether the same predictive processes that drive these AI systems could also explain how humans comprehend sentences.” The answer, the researchers found, is only partly yes. Prediction helps explain how long people look at words while their eyes move forward smoothly, but it does not explain why readers sometimes become stuck or reread an earlier part of a sentence.

To investigate the difference, the team tracked the eye movements of 368 adults as they read carefully constructed sentences. The materials included garden-path sentences, which are grammatically valid but initially encourage readers to adopt an incorrect interpretation. One famous example is “The old man the boat.” At first, “old man” appears to be the subject, leading readers to expect a verb. In fact, “man” functions as a verb, meaning that old people operate the boat. Once the initial interpretation fails, the reader must reorganize the sentence’s structure.

Eye tracking provides a detailed window into this process. Researchers can measure how long a reader’s eyes remain on a word, whether the reader returns to an earlier word, and how often a sentence triggers regressions. These backward movements are not random mistakes. They frequently reflect an attempt to repair a mental interpretation that no longer fits the incoming information. According to the researchers, approximately 20 percent of eye movements during reading are backward, yet standard AI language models offer no direct explanation for when or why readers decide to look back.

The comparison showed that AI predictions successfully captured the first stage of processing: recognizing a word as it appears in a sequence of letters. If a word is highly expected, people tend to move through it more quickly. If it is surprising, they usually spend longer looking at it. This relationship was visible across the language models tested, suggesting that next-word probability provides a meaningful account of early lexical processing—the rapid stage in which visual information is converted into a recognizable word.

The models performed far less well when the task required integrating a word into the larger meaning and grammatical structure of a sentence. In garden-path constructions, readers often spend substantially more time processing a word than its statistical predictability would suggest. They may pause, reread, or revise their interpretation because the sentence’s structure has become incompatible with the meaning they initially assigned. “The predictability of a word really doesn’t even come close to explaining just how much time we spend on difficult words and garden-path sentences,” Timkey said. “LLMs were drastically underpredicting the type of difficulty that we experience when reading.”

This distinction points to a gap between predicting language and understanding how language is organized. Human readers build a mental representation of a sentence as it unfolds. They use that representation to anticipate upcoming words, then update it when new information contradicts their expectations. The process involves syntax, meaning, memory, and decisions about whether an earlier interpretation should be abandoned. A model that predicts the next word may reproduce some of the statistical consequences of this process without implementing the same kind of structural repair.

Brian Dillon, a professor of linguistics at UMass Amherst and senior author of the study, said the findings identify an important boundary between current AI systems and human cognition. Tal Linzen, an associate professor of linguistics and data science at NYU, added that AI remains highly valuable for cognitive science, even when it fails to explain every aspect of reading. By showing where model predictions match human behavior and where they diverge, the research may help scientists develop computational systems that more closely reflect the mind. Such models could eventually improve language-learning technologies and contribute to a better understanding of reading difficulties, although the mechanisms behind rereading and structural confusion remain unresolved.

Subject of Research: People

Article Title: Eye movements reveal a dissociation between prediction and structural processing difficulty in language comprehension

News Publication Date: 7-Aug-2026

Web References: https://doi.org/10.1073/pnas.2532230123

References: Proceedings of the National Academy of Sciences; DOI: 10.1073/pnas.2532230123

Keywords

Human reading, eye movements, language comprehension, artificial intelligence, large language models, next-word prediction, garden-path sentences, cognitive psychology, syntax, rereading, structural processing difficulty

Tags: AI-driven language modelscognitive mechanisms in readingdifferences between human and AI language processingeye movement analysis during readinghuman reading comprehensionimpact of sentence structure on reading difficultyinfluence of word predictability on reading effortinsights into reading behavior from AI analysislarge language models and natural language understandingpredictive text modelingrole of context in reading comprehensionsentence complexity and structural confusion