Beyond Observable Performance: Sanctionability and the Limits of Meaning in Large Language Models.

Sofia Rosa Tinto 

Università degli studi di Torino

This paper addresses the question of whether large language models (LLMs) genuinely possess semantic competence or merely simulate it. Drawing on Kripke’s rule-following consideration, I adopt the view that meaning is grounded not in inner mental facts but in publicly accessible normative practices. On this basis, some recent accounts take the sophisticated performances of LLMs to support attributions of language comprehension. I argue that such performance-based approaches mischaracterize the nature of semantic competence. Satisfying external criteria is not sufficient for attributing linguistic competence: it requires participation in normative practices of a linguistic community. From this account, sanctionability would constitute a necessary condition for the attribution of genuine semantic competence in artificial systems.

Chair: tba

Time:

Location: