Large language models (LLMs) like GPT-4 increasingly mimic human language, but a new study reveals they only partially capture the shared semantic representations that synchronize human brains during storytelling.
The Research
Researchers led by Chen Hong at the Chinese University of Hong Kong used magnetoencephalography (MEG) to record brain activity from speakers and listeners during natural storytelling (pseudo-hyperscanning). They rated content words along ten semantic dimensions—perception, motor, space, time, socialness, animacy, emotion, attention, causality, and drive—both with human raters and five recent LLMs. The team then tested how well these dimensions explained speaker-listener neural synchronization (NS) beyond acoustic and phonological features.
The study, posted on arXiv in June 2026, found that both human- and LLM-derived semantic spaces predicted NS, but the shared semantics are multidimensional, not a single global signal. Critically, these patterns also predicted individual differences in story comprehension. Yet, the overall prediction masked systematic differences: larger LLMs aligned more closely with human neural structure, but even the best models showed incomplete and dimension-dependent alignment. The largest divergences occurred for dimensions tied to agency, affect, and social experience.
Why It Matters
This means that while AI can model many aspects of how we construct meaning, it still struggles with the socially and emotionally grounded dimensions that are crucial for human communication. For everyday cognition, this highlights that our shared understanding is deeply rooted in social and affective experience—dimensions that AI has not fully mastered. For those interested in brain training, it underscores the importance of social interaction and emotional nuance in effective communication.
What You Can Do
To enhance your own semantic processing, engage in active listening and storytelling with varied emotional and social content. Practice explaining complex concepts to others and pay attention to the subtle cues that convey agency and affect. This can help strengthen the neural networks that underpin shared understanding.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.