Inferences in Machines? Inferential Hallucinations in Large Language Models from the Perspective of Metacognitive Feelings

Large Language Models (LLMs) have recently gained considerable attention due to their ability to draw inferences from information provided by users. These neural networks appear to generate logical inferences on a probabilistic basis, often emulating patterns that resemble classical rules of inference. Traditionally, inference has been understood as a cognitive process through which conclusions are derived from premises. However, the nature of inference itself, and the possibility of extending it beyond the human domain, has long been debated. In particular, a central question arises: can the inferential capacity of LLMs be comparable to that of human beings? This article argues that the difference between human inference and the kind of inference produced by LLMs lies in the phenomenological, or affective, dimension of reasoning. More specifically, the distinction concerns the absence of metacognitive feelings that guide, monitor, and evaluate thought processes, such as reasoning or inferring. Although LLMs are capable of simulating novel inferences, their tendency to generate inferential hallucinations reveals the lack of the self-regulatory dimension that characterizes human reasoning. In contrast to the affective component present in human condition, inferential failures in LLMs can be interpreted in terms of three aspects: the reversal curse, inferential hallucinations resulting from decoding strategies, and overconfidence. Accordingly, it is concluded that one of the causes of inferential hallucinations is the absence of metacognitive feelings, including the lack of a feeling of correctness, the absence of the balance provided by a sense of reliability, and the lack of doubt.