AI Lacks Human-Like Rereading in Language Study

Researchers at New York University and UMass Amherst recently used eye-tracking technology to analyze how language is processed ^[1]. The study compared the ways humans and large language models (LLMs) handle "garden path" sentences ^[1]. These are sentences that are grammatically correct but often lead to initial misunderstanding by the reader.

The findings indicate that while LLMs excel at next-word prediction, they do not process information like humans ^[1]. Humans possess a unique ability to recognize when they have misunderstood a sentence ^[1]. They then reread the text to reconstruct its meaning, a cognitive process LLMs currently lack ^[1].

This research suggests a fundamental difference in comprehension strategies ^[1]. While AI relies on statistical probability, humans use active re-evaluation to resolve linguistic confusion ^[1]. The study highlights that current AI architectures do not yet mirror these complex human eye-movement patterns and recovery behaviors ^[1].