Large Language Models (LLMs) exhibit impressive linguistic performance, yet their cognitive and communicative status remains controversial. This paper argues that behavioural fluency alone cannot justify cognitive ascription unless assessed within a designsensitive framework that clarifies what counts as a minimally adequate cognitive architecture. We distinguish structurally constrained functionalism from unconstrained functionalism and propose a minimal cognitive core – environmentally coupled perception, inference, and action – as a baseline for cognition. Against this baseline, current LLMs fall short: as standalone text-based systems, they lack sensorimotor coupling, integrated inferential responsiveness, and autonomous agency, and thus cannot sustain the forms of world-directed engagement characteristic of even basic cognitive systems. We examine sycophancy and the absence of spontaneity as diagnostic consequences of architectures optimised for user alignment rather than epistemic commitment, showing how these limitations systematically undermine claims of cognitive or communicative competence. Finally, we develop the deprivedinterface argument: without embodied or interactive access to an environment, LLMs cannot revise or ground their representations, resulting in a derivative form of linguistic idealism – the view according to which an LLM’s world consists of nothing but training data. While future multimodal or embodied architectures may indeed mitigate aspects of interface deprivation, such advances would require structural changes enabling genuine worlddirected engagement rather than simply expanding input modalities. Clarifying these distinctions is essential for avoiding category mistakes in AI evaluation and for grounding cognitive ascription in principled criteria rather than surfacelevel behavioural success. We conclude that LLMs simulate cognition and communication without instantiating the structural conditions required for either, and we outline the implications for AI evaluation and cognitive ascription.
Chatting with LLMs: Neither Cognition nor Communication
Salis Pietro
;Da Pelo Matteo
2026-01-01
Abstract
Large Language Models (LLMs) exhibit impressive linguistic performance, yet their cognitive and communicative status remains controversial. This paper argues that behavioural fluency alone cannot justify cognitive ascription unless assessed within a designsensitive framework that clarifies what counts as a minimally adequate cognitive architecture. We distinguish structurally constrained functionalism from unconstrained functionalism and propose a minimal cognitive core – environmentally coupled perception, inference, and action – as a baseline for cognition. Against this baseline, current LLMs fall short: as standalone text-based systems, they lack sensorimotor coupling, integrated inferential responsiveness, and autonomous agency, and thus cannot sustain the forms of world-directed engagement characteristic of even basic cognitive systems. We examine sycophancy and the absence of spontaneity as diagnostic consequences of architectures optimised for user alignment rather than epistemic commitment, showing how these limitations systematically undermine claims of cognitive or communicative competence. Finally, we develop the deprivedinterface argument: without embodied or interactive access to an environment, LLMs cannot revise or ground their representations, resulting in a derivative form of linguistic idealism – the view according to which an LLM’s world consists of nothing but training data. While future multimodal or embodied architectures may indeed mitigate aspects of interface deprivation, such advances would require structural changes enabling genuine worlddirected engagement rather than simply expanding input modalities. Clarifying these distinctions is essential for avoiding category mistakes in AI evaluation and for grounding cognitive ascription in principled criteria rather than surfacelevel behavioural success. We conclude that LLMs simulate cognition and communication without instantiating the structural conditions required for either, and we outline the implications for AI evaluation and cognitive ascription.
I metadati presenti in IRIS UNICA sono rilasciati con licenza Creative Commons CC0 1.0 Universal, mentre i file delle pubblicazioni sono protetti da diritto d'autore, salvo diversa indicazione.
