Scientific study reveals AI processes language differently than the human brain

A study published in the journal eLife indicates that large language models, such as ChatGPT, Gemini, and Claude, do not process sentence meaning in the same way as the human brain. Researchers from the University of Melbourne and Hitotsubashi University used 7-tesla functional magnetic resonance imaging (fMRI) to monitor the brain activity of 30 participants as they read 108 specifically designed sentences.

The team compared these neural responses to the internal patterns of eight transformer-based models. While the AI models outperformed basic systems that ignore word order, they struggled to predict brain activity accurately when faced with sentences involving inverted semantic roles.

The researchers concluded that while these AI systems can identify syntax in isolation, they fail to automatically integrate grammatical structure with lexical meaning to form a unified representation of a sentence, a process fundamental to human cognitive function.