Zhou, Y., Liu, E., et al. (2024).
Advances in neural information
processing systems, 37, 137999–138031.
Abstract
Do machines and humans process language in similar ways? Recent research has hinted at the affirmative, showing that human neural activity can be effectively predicted using the internal representations of language models (LMs). Although such results are thought to reflect shared computational principles between LMs and human brains, there are also clear differences in how LMs and humans represent and use language. In this work, we systematically explore the divergences between human and machine language processing by examining the differences between LM representations and human brain responses to language as measured by Magnetoencephalography (MEG) across two datasets in which subjects read and listened to narrative stories. Using an LLM-based data-driven approach, we identify two domains that LMs do not capture well: social/emotional intelligence and physical commonsense. We validate these findings with human behavioral experiments and hypothesize that the gap is due to insufficient representations of social/emotional and physical knowledge in LMs. Our results show that fine-tuning LMs on these domains can improve their alignment with human brain responses.
Here are some thoughts:
This study looked at where AI language models and human brains think differently about language. The researchers had people read Harry Potter and listen to storytelling while measuring their brain activity, then checked how well an AI model could predict those brain signals word by word. The words the AI got wrong tended to fall into two buckets: social and emotional content (feelings, relationships, tone) and physical commonsense (how objects and bodies work in the real world). When they trained the AI on these specific topics, it lined up better with human brains. The likely reason is simple: humans learn emotions and physical knowledge by living in the world and interacting with people, while AI only reads text, so those are exactly the areas where it falls short.
This is a clever and intuitive study, but I would hold the conclusions loosely. The brain data is thin (only narrative stories, and the listening portion came from just one person), so it is hard to know how far the findings generalize. The core idea that AI struggles with emotional and physical knowledge is believable and matches other research, but "the AI predicts brain signals better after fine-tuning" is a modest result, not proof that the model now understands feelings the way people do. Predicting brain activity and actually thinking like a brain are not the same thing, and the authors are honest about this limitation. So treat it as a promising hint about why AI and humans differ, not a settled answer.
