Document Type
5 minute oral
Start Date
6-3-2026 2:00 PM
Description
Wittgenstein’s distinction between what can be said and what can only be shown frames a limit of propositional language. Affect and aesthetic are not primarily matters of correct description but of how expressions function within shared forms of life. Large language models (LLMs), however, increasingly produce fluent language that resembles such understanding by reproducing the patterns through which people ordinarily talk about emotion and perspective.
This paper argues that apparent Theory of Mind (ToM) in LLMs is understood as competence in the publicly shared patterns of mental-state language, rather than as grounded understanding. Using abstract artworks, we show that LLMs tend to interpret art-works in emotional terms, even when alternative interpretations are relevant. To make this contrast visible, we propose an evaluation framework that compares model interpretations with human responses. This framework examines whether interpretations remain open to revision in light of perspective and intention, a capacity closely related to affective ToM.
Included in
Beyond the Sayable: Wittgenstein, Theory of Mind and Affective Simulation of LLMs
Wittgenstein’s distinction between what can be said and what can only be shown frames a limit of propositional language. Affect and aesthetic are not primarily matters of correct description but of how expressions function within shared forms of life. Large language models (LLMs), however, increasingly produce fluent language that resembles such understanding by reproducing the patterns through which people ordinarily talk about emotion and perspective.
This paper argues that apparent Theory of Mind (ToM) in LLMs is understood as competence in the publicly shared patterns of mental-state language, rather than as grounded understanding. Using abstract artworks, we show that LLMs tend to interpret art-works in emotional terms, even when alternative interpretations are relevant. To make this contrast visible, we propose an evaluation framework that compares model interpretations with human responses. This framework examines whether interpretations remain open to revision in light of perspective and intention, a capacity closely related to affective ToM.