(Poster) Asymmetric Lexical Representations of Affect: On Emotion Understanding in LLMs

(This poster summarizes a paper to be published in the ACM AI Leadership Summit 2026 proceedings) This paper proposes that emotion understanding in large language models (LLMs) is asymmetric to human affective cognition and therefore faces an inevitable performance ceiling when applied to areas which require sustained interpersonal relating, such as mental health care. While LLMs share in the correlative, grammatical and inductive aspects of emotion understanding, they differ in that they do not share in the same non-linguistic, interoceptive sources of information that humans do. This makes the lexical representations they build asymmetric and incomplete when compared with humans. Therefore, any apparent emotion understanding in LLMs is the result of the successful elicitation of emotion understanding in human interlocutors across a dynamic representational medium. The knowledge itself rests in the source and destination, rather than the intermediate representation generated by LLMs. This fundamental asymmetry is not solved by scaling toward multi-modal variants of third personal data. Research directions which prioritize emotion self-understanding in humans, along with the facilitation of human-human interaction, will advance Emotion AI further than multi-modal integration alone.