Researchers in the US conducted an experiment with eye tracking and 400 language models, proving that AI cannot explain the process of rereading complex sentences in humans, reports infohub.kz.
Scientists from New York University and the University of Massachusetts Amherst conducted a study involving 368 adult readers and over 400 AI language models. The work, published in the journal PNAS, showed in which aspects the mechanisms of text perception in humans diverge from neural network algorithms.
Using eye-tracking technology, the researchers analyzed the reading process of grammatically complex sentences. It turned out that language models explain well the initial stage of recognizing a single word based on predicting the next element. However, AI cannot model the next stage - integrating the word into the overall semantic context of the phrase.
When encountering complex texts, people spend significantly more time and in 20% of cases look back to reread. The authors of the work, William Timkey, Brian Dillon, and Tal Linzen, noted that the findings will help create more accurate cognitive models of the human brain.