Nisrine El Hannach

HALLUCINATION AS AN EMERGING FORM OF SEMANTIC AMBIGUITY IN LARGE LANGUAGE MODELS:

A computational pragmatic-linguistic approach

Nisrine El Hannach¹

This study examines the phenomenon of hallucination in large language models from a computational linguistic perspective, moving beyond its reduction to a merely technical or informational error, and recontextualizing it within the broader problem of ambiguity and transformations in meaning production. While ambiguity in traditional linguistic theory is understood as a legitimate multiplicity of interpretive possibilities resulting from lexical, syntactic, pragmatic, or referential indeterminacy, hallucination reveals a different form of ambiguity—one that does not arise from multiple possible meanings, but from the generation of meaning lacking textual, referential, or contextual grounding. Accordingly, the study raises a central question: does hallucination represent a contemporary extension of the concept of ambiguity, or does it constitute a conceptual rupture from it? The study argues that hallucination is not merely a statistical deviation in model performance, but rather a form of generative ambiguity resulting from the dissociation of linguistic production from the pragmatic and referential constraints that regulate meaning in natural communication. In this process, probability shifts from being a mechanism for interpreting meaning to becoming a mechanism for producing semantic surplus that may ultimately lead to fabrication. Based on this premise, the study proposes a reconceptualization of hallucination through its association with the notion of Overgeneration of Meaning, and develops a computational linguistic classification of its patterns, including referential, semantic, pragmatic, and inferential hallucinations. Methodologically, the study adopts an approach that combines theoretical analysis with applied investigation, drawing on semantic linguistics, pragmatics, and computational linguistics. It also employs applied samples drawn from the outputs of large language models in order to identify hallucination patterns and analyze their mechanisms in light of the concepts of ambiguity, reference, relevance, and inference. The study concludes that hallucination should not be understood merely as a reliability issue in large language models, but rather as evidence of a deeper transformation in the concept of ambiguity itself, thereby calling for an expanded theoretical framework for the study of meaning in the age of generative artificial intelligence