Contributory Expertise

In the AI Story

The distinction becomes clearest at the boundary between knowing-that and knowing-how — between being able to discuss a practice and being able to perform it. Gilbert Ryle's version of this distinction treats knowing-how as the manifestation of individual skill. Collins's version treats it as inherently social: one knows how to do physics by being a member of the community that does physics, not by possessing an individual set of skills that could be abstracted from community participation.

The Collins and Thorne 2026 paper provides the empirical case. Gravitational wave physicists, asked why they ignore a fringe science paper, can articulate reasoning that sounds like it comes from general methodological principles. But the reasoning is actually grounded in the community's accumulated social knowledge — reputation assessments, histories of prior claims, tacit agreements about what evidence would be required to take the claim seriously. A language model can produce text that mimics the reasoning. It cannot produce the reasoning, because the reasoning requires participation in the community that generates the tacit judgments on which it rests.

The implications for AI-assisted work are specific. Tasks that require only interactional expertise — summarizing literature, explaining concepts, producing text that follows a domain's conventions — are tasks machines can do. Tasks that require contributory expertise — making novel scientific judgments, setting community standards, exercising the situational judgment that distinguishes competent from excellent practice — remain human work. The difficulty is that the two kinds of tasks are often woven together in a single workflow, and users without contributory expertise themselves cannot reliably tell which kind they are delegating to the machine.

Origin

Collins introduced the term alongside interactional expertise in Rethinking Expertise (2007), as part of the Periodic Table of Expertises he developed with Robert Evans. The distinction crystallized insights from his decades of ethnographic work on gravitational wave physics and the sociology of scientific knowledge more broadly.

Key Ideas

Doing, not discussing. Contributory expertise is the capacity to advance a practice, not merely to participate in its conversation.

Socially constituted. Contributory expertise requires membership in the community of practice — participation in its ongoing work, not merely exposure to its textual output.

The PhD example. A PhD requires both interactional fluency (mastery of the literature) and contributory competence (original contribution), and LLMs are at PhD level on the first and not the second.

The evaluation problem. Whether an output reflects contributory expertise can only be judged by someone with contributory expertise — producing a structural asymmetry in AI evaluation.