The Logic of Attributing Understanding to Language Models

The article analyses the logical scheme by which understanding is attributed to contemporary language models on the basis of their success on benchmark tasks. It is shown that the transition is a second-figure syllogism with an undistributed middle and is therefore illegitimate. The error is masked by a suppressed premiss introduced by the canonical format, which lends the conclusion credibility before its grounds are examined. The masking mechanism is analysed within V. N. Bryushinkin’s systemic model of argumentation: the defect lies in the logical core; the suppressed premiss that a task class represents understanding belongs to the cognitive structure; the canonical format, as a communicative condition, renders that premiss acceptable without making it explicit. A comparison with S. S. Averintsev’s account of Byzantine rhetoric shows that the masking layer is not fixed: it follows whatever has the force of a canonical form of credibility — deduction in one case, the quantitative format in the other. In contemporary sociology and in the evaluation of artificial intelligence this defect is registered empirically; the article identifies its logical form. The distinction between the false and the illegitimate is proposed as an operational criterion: repeated failure on that distinction withdraws the attribution of understanding as holding a concept as a whole, whereas success does not certify it. Evidence may gain weight without becoming proof. In ordinary communication errors of this kind can hardly be anticipated systematically; in work with language models they can, provided the canonical format is not allowed to mask the suppressed premiss.