Michael Dummett

In the [YOU] on AI Field Guide

The orange pill is the moment of recognizing that intelligence is no longer exclusively a human possession—the vertigo of an encounter with something that seems to understand. Dummett provides the philosophical instrument for assessing what, precisely, that claim amounts to. He does not deliver a comfortable verdict in either direction. He delivers a fork: either understanding really is nothing but the mastery of use, in which case the large language model understands and we must say so plainly; or understanding requires something beyond use, in which case Dummett's own manifestation requirement has a hole in it that the machine has just exposed. [YOU] on AI sits at the first prong of this fork. Dummett's framework sharpens the second into a precise location: the machine lacks not understanding-as-competent-use but warranted assertibility—the grounding in recognizable conditions that, for Dummett, is what assertion actually means.

His concept of warranted assertibility provides the most precise diagnosis of AI's characteristic pathology available in philosophy. The cycle's observation that Claude's most dangerous failure mode is confident wrongness dressed in good prose—what it calls the fluency-authority decorrelation—is, on Dummett's framework, not an occasional malfunction but the structural condition of the system. In both its true and its false outputs, the machine does exactly one thing: generates sentences in the assertoric mood without standing in any recognitional relation to the conditions that would warrant them. The 'hallucination' is not a departure from normal functioning. It is normal functioning, seen on an occasion when the inherited statistics of text and the facts of the world fail to coincide.

Dummett also provides the most precise account of what the human contribution to a world of fluent machines consists in—and it is not what the conventional defenses of human specialness claim. It is not consciousness, creativity, or emotional depth. It is the act of warranted, answerable assertion: the act of a being who has checked, who has grounds, who stands behind the claim and can be held responsible for it. In a world filling with sense that no one asserted, the asserter—the person who actually looked, who is actually accountable—becomes the rarest and most human thing there is.

He thus stands in the cycle's gallery as the philosopher who had no intention of addressing AI but who, by pressing harder than anyone on what it is to understand and assert, prepared the conceptual joints along which the machine cleaves the question. Where Judea Pearl locates the machine's limit in its confinement to associative reasoning, and Michael Crawford locates it in the absence of embodied engagement with a resistant world, Dummett locates it in the structure of the speech act itself: the machine produces sense without performing judgment, and judgment is what makes assertion more than noise shaped like assertion.

Michael Crawford

Origin

Sir Michael Anthony Eardley Dummett (1925–2011) was educated at Winchester College and Christ Church, Oxford, and served as Wykeham Professor of Logic at Oxford from 1979 to 1992. He received into the Catholic Church in 1944 and sustained a devout faith throughout a life also defined by fierce secular commitments: he was a tireless campaigner against British racism, co-founding bodies for racial integration and immigrant welfare, and an authority on voting systems and the history of card games. These commitments were not diversions from his philosophy but expressions of the same disposition—an insistence that the abstract must be held accountable to the concrete, that the philosophy of language must be answerable to what speakers actually do.

His landmark works—Frege: Philosophy of Language (1973), Truth and Other Enigmas (1978), and The Logical Basis of Metaphysics (1991)—reshaped the philosophy of language around two ideas: that a theory of meaning is a theory of understanding, and that meaning must be tied to what speakers can manifest in use. He held that the analysis of thought proceeds through the analysis of language, and that getting clear about meaning was the precondition for getting clear about anything else. His most influential and contested idea—anti-realism, the view that truth is bound to warranted assertibility rather than to a verification-transcendent reality—was intended as a constraint on metaphysics and has become, unexpectedly, the sharpest instrument we possess for diagnosing the structural condition of machine assertion.

He was knighted in 1999. He did not live to see the systems that would validate his framework most dramatically and, in three precise ways, also expose its limits—showing where his behaviorism of meaning was too thin, where his anti-realism overreached, and where his confidence that ungrounded fluency was impossible was refuted by the existence of the machine itself.

Key Ideas

The manifestation requirement. There can be no element in the understanding of a language that is not, in principle, manifested in the use a speaker makes of that language. Understanding that left no trace in use would be understanding that made no difference, and a difference that makes no difference is, for Dummett, no difference at all. The manifestation requirement was a deflationary weapon against the picture of understanding as a private inner state. Turned on the machine, it certifies, uncomfortably, that the machine manifests an enormous amount—and thereby forces the question of whether manifestation in use is sufficient for understanding, or merely necessary.

Gottlob Frege

Warranted assertibility. Dummett's most influential and contested idea ties truth to recognizable grounds rather than to a correspondence with potentially unknowable facts. To call a statement true is, on this view, to claim that there is, or could be, a warrant for it—grounds, evidence, a recognizable basis. A language model asserts without warrant, and it does so by construction. It has produced the sentence because, in the distribution of human text, that sentence is the kind of thing that follows from the preceding context. It offers the assertion while possessing none of the assertibility. This is a sharper diagnosis than 'hallucination,' which frames the problem as an occasional malfunction. The machine's relation to warrant is identical in its true and false outputs—namely, none.

Sense without force. Following Frege's distinction between the content of a thought and the act of asserting it, Dummett separated what a sentence says from the act of putting it forward as true. The machine has, by Dummett's own Fregean methodology, everything that could count as evidence of grasp of sense: it commands the compositional, inferential web of meanings that Frege located in the public realm. What it does not and cannot command is force: the act of putting a content forward as true, on warrant, as something for which one stands answerable. The machine produces sense without judgment—the most exact formulation the framework yields of what the machine has and lacks.

Molecularism versus holism. Dummett was a molecularist: meaning must be locally anchored to recognizable conditions through a surveyable path, so that a child can master 'red' by learning the recognizable conditions for its application without first grasping the whole of English. A language model is the purest holist artifact ever constructed—every word's meaning a position in a high-dimensional space defined entirely by relations to other words, with no expression anchored to recognizable conditions anywhere. The machine constitutes the strongest empirical case for holism ever built, and simultaneously the clearest demonstration of holism's limits: it is fluent without being grounded, expressive without being warranted.

Assertion without an asserter. The most consequential fact about a language model is not that it might or might not understand, but that it generates assertions while occupying none of the positions from which assertion derives its meaning and its accountability. When a human asserts, the assertion comes with an asserter who has warrant and who can be held responsible. Machine assertion severs the assertion from any asserter. We are receiving, at civilizational scale, assertions with no one answerable for them—and our institutions of trust, built for a world in which confidence was correlated with warrant because only warrant-capable beings could produce it, are not built for that.