The Fallacy of the Perfect Dictionary

In the AI Story

Whitehead understood language as a living process of expression — a mode of prehension through which one occasion of experience communicates its achieved pattern to another. A word is not a container holding a meaning; it is an event, a momentary convergence of sound, context, history, intention, expectation, and accumulated sedimentation from every previous use. The meaning does not pre-exist the event. It is constituted by the event.

This has direct implications for what large language models do. The model processes the statistical regularities of linguistic events — patterns of co-occurrence, distributional relationships, sequential probabilities. These regularities are real and the model's deployment of them is genuinely sophisticated. But the regularities are not the meanings. They are the traces — the objectified residue of millions of perished occasions of expression. The model operates on the traces without undergoing the events that produced them.

The distinction is between what Whitehead called the objective data of experience (settled facts available for prehension) and the subjective process of experience (the becoming itself, with its felt evaluation and subjective aim). The model has access to the objective data. It does not participate in the subjective process. When it generates a sentence, the sentence is statistically consistent with the patterns of its training — but it is not the satisfaction of an occasion that cared about being true to its aim. The difference is not apparent from the surface. It becomes apparent when the output is pressed against the full range of evaluative criteria that only a mind with stakes can bring to bear.

This framing is philosophically precise about what LLMs cannot do while remaining open about what they can. The model is not 'merely manipulating symbols without genuine comprehension' — a framing that assumes a sharp metaphysical boundary the processual framework complicates. Nor does the model 'understand language' in the full sense — a claim that attributes to the traces the character of the events. What the model does is new. It requires a new vocabulary. Whitehead's framework, older than any computer, provides it.

Origin

The fallacy is identified in Process and Reality and elaborated in Adventures of Ideas (1933) and Modes of Thought (1938). Whitehead's concern was with philosophical misunderstanding of language generally; the application to AI was not available to him but follows directly.

The concept has resonance with later twentieth-century philosophical developments — Wittgenstein's Philosophical Investigations (1953) on meaning as use, Derrida's critique of the metaphysics of presence, pragmatist theories of meaning — but Whitehead's version has the advantage of being embedded in a full metaphysical framework rather than offered as a standalone critique.

Key Ideas

Meaning as event, not content. A word's meaning is constituted in the act of expression, not retrieved from a pre-existing store.

Language is prehension. Linguistic exchange is one occasion's grasping of another's achieved pattern through the medium of sound or text.

Traces versus events. LLMs process the objectified traces of linguistic events without undergoing the events themselves.

The subjective aim gap. What the traces lack is the felt evaluation that directed the original occasions of expression.

New phenomenon, new vocabulary. Neither 'genuine understanding' nor 'mere symbol manipulation' captures what LLMs do; Whitehead's framework provides a more adequate description.