A new study from researchers at The University of Osaka investigated whether large language models (LLMs) conceptualize words in ways similar to human cognition by comparing AI ratings of word attributes with human psychological ratings. The team selected 695 early‑acquired English words — words typically learned by children — and asked LLMs to rate them on dimensions like concreteness, socialness, and arousal. These AI‑generated ratings were then compared with existing human norms from prior psychological studies.
The results showed that for some attributes, such as concreteness, imageability and body‑object interaction, there was a relatively strong correlation between how LLMs and humans rated the words. This suggests that even without direct interaction with the physical world, LLMs may encode some aspects of human‑like conceptual knowledge because patterns of human experiences are embedded in the language they were trained on.
However, the study also found significant differences in how AI and humans perceive certain linguistic features. For example, evaluations of iconicity — the degree to which a word’s sound resembles its meaning — diverged markedly between the two. Function words like prepositions and conjunctions also revealed systematic discrepancies: whereas human ratings varied widely, LLMs tended to assign consistently low values for concreteness, indicating they do not “experience” these word types the way humans do.
When examining how these attributes predicted age of word acquisition, the researchers noted that LLMs sometimes exaggerated patterns that are more subtle in human learning — for instance overestimating the relation between certain features and when children learn words. The team believes these insights could help build LLMs that better mirror human cognition or serve as complementary tools in psychology, shedding light on how humans learn and process language.
