Warranted Assertibility

In the [YOU] on AI Field Guide

The cycle's observation that [YOU] on AI identifies as its signature hazard the fluency-authority decorrelation—confident wrongness dressed in good prose—acquires its most rigorous diagnosis from Dummett's concept. The decorrelation is not accidental or reparable by better training. It is the structural consequence of building a system that produces the assertoric form without the assertoric ground. Warranted assertibility explains why the 'hallucination' problem is misnamed: calling a false output a hallucination implies that the system normally tells the truth and occasionally slips. On Dummett's framework, the system's relation to warrant is identical in its true and false outputs—namely, none. When the output is accurate, it is because the statistics of text and the facts of the world happen to coincide. The system cannot distinguish between these cases, because distinguishing them would require the recognitional relation to conditions that the concept of warranted assertibility requires.

The concept also illuminates the asymmetry that makes the current moment genuinely unprecedented. Human assertion has always been imperfect—people have always lied, erred, and confabulated. But human assertion was produced roughly at the rate at which humans could warrant it, because warranting is laborious and assertion was socially costly when unwarranted. The machine produces warranted-seeming assertion at the speed of computation, decoupled entirely from the slow work of verification. The ratio of assertion to warrant, roughly balanced across human history, is now wildly skewed. What becomes scarce and precious is exactly what the machine cannot supply: the asserter who has checked, who has grounds, who stands behind the claim and is answerable if it is wrong.

Origin

Dummett introduced the concept as part of a systematic challenge to the truth-conditional semantics that dominated the philosophy of language in the mid-twentieth century. Classical truth-conditional theories, following Frege and early Wittgenstein, held that to know the meaning of a sentence is to know the conditions under which it is true—conditions that may obtain or fail to obtain independently of any speaker's capacity to recognize which. Dummett's objection was epistemological and verificationist in spirit: if the truth conditions of a sentence could in principle be forever beyond our recognition, then knowledge of those conditions could never be manifested in any speaker's behavior, and a theory of meaning built on them would fail the manifestation requirement.

His alternative—tying meaning to assertibility conditions, the recognizable circumstances under which assertion is warranted—was intended primarily as a contribution to the philosophy of mathematics (where it aligned with intuitionist mathematics, rejecting proofs by contradiction that appeal to truths beyond constructive demonstration) and as a general metaphysical thesis (anti-realism, the view that there are no facts of the matter beyond what we could in principle recognize). Both applications remain contested. What Dummett did not foresee was that his concept would prove most valuable not in mathematics or metaphysics but in the analysis of systems that produce assertion without any recognitional relation to the conditions that would warrant it—systems he never imagined.

Key Ideas

Meaning tied to recognizable grounds. To grasp the meaning of a sentence, on Dummett's account, is to know what would warrant asserting it—what to check, what would count as evidence, what a competent speaker can actually tell obtains. This is a demanding standard: it requires not just formal facility with the sentence but a genuine recognitional capacity that connects the sentence to the world through conditions the speaker can discriminate. The machine has the formal facility and lacks the recognitional capacity entirely.

The flood of ungrounded claims. Warranted assertibility provides the most precise vocabulary for the socio-epistemic consequence of AI deployment at scale. The surface signals by which human communities have always gauged credibility—fluency, confidence, the declarative mood, the citation, the air of having-been-checked—are now uninformative. These signals worked because, for all of human history, producing them was correlated with possessing the warrant. The machine breaks the correlation. It produces every signal of warrant with none of the substance, indiscriminately, for the warranted and unwarranted alike.

Where Dummett's anti-realism strains. The machine paradoxically strengthens the realist objection to Dummett. We want to say that a universally believed, perfectly fluent machine assertion can still be false—that there is a fact of the matter about the bridge independent of any warrant. If anti-realism cannot preserve that gap between warrant and truth, it is too thin a theory to capture even what we want to say about machine errors. The machine needs both Dummett's diagnosis of what assertion lost and the realist's reminder that truth is answerable to a world that warranty cannot confer.