Abstract
This article investigates how large language models (LLMs) engage with the writings of Charles Sanders Peirce, whose semiotic theory defines meaning-making as triadic, recursive, and historically grounded. Analyzing conceptually dense passages from Peirce’s Collected Papers—particularly those on icons, indices, symbols, and interpretants—we examine the interpretive activity of four leading LLMs: GPT-4 (OpenAI), GPT-4o (OpenAI), Claude 3 Opus (Anthropic), and Gemini 1.5 (Google DeepMind). While these systems generate fluent and syntactically coherent paraphrases, their outputs largely remain within the level of paraphrastic and explanatory interpretation. Features associated with recursive interpretant formation, central to Peirce’s concept of semiosis, appear only intermittently and are not sustained across responses. In Peircean terms, this suggests a predominance of Secondness, or reactive patterning, with only partial approximation of Thirdness: the recursive mediation and abductive generalization associated with interpretive depth. Peirce’s framework is thus used as a diagnostic lens for distinguishing symbolic fluency from different forms of interpretive activity. The findings point to a gap between linguistic coherence and sustained semiotic mediation, highlighting the importance of evaluating machine-generated meaning within humanities frameworks that account for context, normativity, and the continuity of interpretation. In this sense, semiotics provides not only a critical vocabulary but also a methodological foundation for assessing the interpretive claims of generative models.
https://doi.org/10.1093/llc/fqag133
Authors:
Alon Friedman and Martin Thellefsen,
