The knowledge norm of assertion

The knowledge norm of assertion is the thesis that one must assert that p only if one knows that p, and that this norm is constitutive of assertion as a speech act. Williamson’s case for it draws on ordinary practice: we challenge assertions by asking “How do you know?”; “p, but I don’t know that p” is paradoxical; and asserting that a lottery ticket has lost seems improper even when the odds make it almost certain. Rivals propose norms of truth or of justified or reasonable belief, or, like Kelp, a function-first account on which assertion’s job is to generate knowledge in hearers. Lackey’s selfless assertions are a standard objection.

On Williamson’s view, a hearer’s testimonial gain runs through the assertion’s being regulated by the norm. Because knowledge is factive and modally robust, production so regulated inherits a connection to truth; it is one way of satisfying the Tracking Minimum originally.

Applied to language models, the question is not whether they have attitudes but what regulates their production. On the standard description it is conditional likelihood under a learned distribution, adjusted by preference optimisation against human ratings, which rewards what raters approve: the appearance of conformity to the norm. A process optimised to seem norm-conforming is not thereby regulated by the norm, for the same reason a student optimised to please graders is not thereby aiming at truth. Sycophancy is what this looks like in behaviour.

That description is now incomplete, and I concede it rather than finesse it. Training with verifiable rewards replaces the reward model with a deterministic verifier, such as a proof checker or a unit test, so the signal is correctness rather than approval. In such domains production is regulated by a procedure whose function is to constrain output to truth, which looks like, and I think is, original satisfaction. Whether it transfers to domains where nothing verifies is open.

Function-first accounts relocate the truth-connection rather than removing it. An assertion fulfils a knowledge-generating function only if the hearer’s belief is non-accidentally true, and that cannot be secured by a token whose production is unconnected to the truth of its content.