The Parity Principle

In the AI Story

The principle operates as a rebuttal to what Clark calls bioprejudice — the assumption that cognition must occur in biological tissue to count as cognition. If the cognitive status of a process depended on its substrate, then the principle would fail trivially. But Clark argues that no philosophical account of what makes something cognitive, other than question-begging appeals to biology, actually requires substrate specificity. Cognition is defined by what it does: integrates information, guides behavior, supports reasoning. If an external process does these things, it is cognitive by the only criterion we actually have.

The principle's modesty is strategic. It does not claim that every external process is cognitive — only that external processes are not disqualified from being cognitive by virtue of their externality. Further conditions are required: reliability, availability, automatic endorsement, historical trust. These conditions narrow the scope of genuine cognitive extension considerably. A random scrap of paper is not part of someone's mind. Otto's notebook, which meets all the conditions, is.

Large language models present the sharpest test of the principle. When an AI performs cross-domain association, holds context in working memory, integrates concepts across fields, and generates novel combinations, it is performing precisely the functions that philosophers have traditionally regarded as the hallmarks of sophisticated cognition. The parity principle's verdict is that these functions are cognitive regardless of whether they occur in biological or computational substrate.

Critics argue that the principle proves too much. If AI processing is cognitive by parity, then by the same logic a calculator's processing is cognitive, a thermostat's processing is cognitive, and the concept loses its meaning. Clark's response is that the additional conditions — reliability, availability, trust, endorsement — do the work of narrowing the scope. The calculator does not meet all the conditions because we do not automatically endorse its outputs as our own. The AI system, when coupled with a trusting user, does.

Origin

The principle appears in the 1998 paper in a single sentence that has been quoted thousands of times since: "If, as we confront some task, a part of the world functions as a process which, were it to go on in the head, we would have no hesitation in accepting as part of the cognitive process, then that part of the world is (for that time) part of the cognitive process." Clark has acknowledged that the phrasing is more careful than it first appears — the counterfactual framing is essential, and the temporal qualifier matters.

The principle has been refined substantially over twenty-seven years, with Clark responding to critics by articulating the conditions more precisely. The current formulation requires not merely functional equivalence but glue-and-trust conditions — the external resource must be reliably present, reliably endorsed, and embedded in a history of trust.

Key Ideas

Location-neutrality. A cognitive process does not become non-cognitive by moving outside the skull; the substrate is incidental to the cognitive status.

Counterfactual test. Ask whether, if the process occurred inside the head, we would call it cognition. If yes, externalization does not change the answer.

Functional role, not mechanism. What matters is what the process does in the cognitive economy of the agent, not how it does it.

Conditions still apply. Not every external process passes — reliability, availability, and automatic endorsement remain necessary conditions.

Scales with the tool. The principle's stakes grow as external components become more cognitively sophisticated; AI represents the most consequential application yet.