The cycle asks what machines can and cannot do, and Collins provides the most empirically grounded answer available from the sociology of knowledge. His distinction between mimeomorphic and polimorphic action—between actions whose correctness depends on copying surface behavior and actions whose correctness depends on understanding the social context—maps directly onto the split [YOU] on AI observes in the Trivandrum engineers. The eighty percent of their work that Claude Code absorbed was predominantly mimeomorphic: the syntax of Python does not change depending on who writes it or why, and a system trained on sufficient examples can reproduce it without understanding the social context in which it is used. The twenty percent that remained—the judgment, the architectural instinct, the taste that separates a feature users love from one they tolerate—was polimorphic. It depended on understanding the specific social context: the team, the users, the competitive landscape, the accumulated conventions of a particular codebase.
Collins would add a complication that [YOU] on AI acknowledges but does not fully resolve: the eighty percent was not merely filler. It was the apprenticeship through which the twenty percent was formed. The engineer who spent years debugging dependency conflicts and reading error messages was building, through that specific struggle, the collective tacit knowledge that now constitutes his irreplaceable value. The current generation of senior engineers possesses polimorphic expertise because they lived through the mimeomorphic work that built it. The question Collins poses for the next generation is genuinely open: will they develop equivalent judgment by directing AI rather than implementing code, or will the judgment thin with each generation as the apprenticeship that built it disappears?
His most pointed contribution to the cycle's central concern about AI fluency is the diagnosis he made in 2025 and confirmed empirically in 2026: large language models 'have proved to have no moral compass—they answer queries with fabrications with the same fluency as they provide facts.' This is not a moral judgment. It is a sociological observation. A moral compass is a product of collective tacit knowledge—of the community's evolving sense of what is acceptable and what is not, transmitted through socialization and maintained through ongoing social participation. A system that has not been socialized cannot possess it, regardless of the volume of text about ethics it has processed. It can mimic the surface features of moral reasoning. It cannot distinguish between a genuinely moral position and a rhetorically persuasive one, because the distinction is made by the community, not extractable from its published texts.
The specific test in Collins and Thorne's 2026 paper captures the gap precisely. They asked a large language model to reason about why gravitational wave physicists would dismiss a particular fringe science paper—a task that required not formal reasoning from published principles but social judgment about the credibility of the author and the community's implicit standards for what counts as a paper worth engaging with. The model produced plausible-sounding arguments. It could not produce the reasoning. The reasoning depended on tacit knowledge built in spoken discourse within closed groups of experts—precisely the knowledge that is not captured in published text and therefore not available to a system trained on published text.
Born in 1943, Collins studied sociology at the University of Essex and built his career at the University of Bath, where he co-founded the Centre for the Study of Knowledge, Expertise and Science. His long engagement with gravitational wave physics began in the early 1970s and has produced a series of books that are landmarks in the social study of scientific knowledge: Changing Order (1985), which introduced the experimenter's regress; Gravity's Shadow (2004) and Gravity's Ghost and Big Dog (2013), which followed the field through its early detection attempts and the controversies they produced; and Tacit and Explicit Knowledge (2010), which systematized his three-species taxonomy of tacit knowledge.
The taxonomy—relational, somatic, and collective tacit knowledge—is the tool the cycle uses most heavily. Relational tacit knowledge is tacit for contingent reasons and can in principle be articulated; this is the species AI handles best, absorbing vast quantities of knowledge that individuals possessed but had not formalized. Somatic tacit knowledge resides in the body; important in physical domains, less so in software. Collective tacit knowledge resides in the social group and can only be acquired through socialization into that group; this is the species AI cannot acquire by training on the group's textual output, however comprehensive that output.
Collins's engagement with AI began seriously with his 1990 book Artificial Experts, which argued against the strong AI project of the era on sociological grounds. His position has not changed in its essentials despite the massive capabilities improvements of the intervening decades: the machines have become spectacularly better at copying. They have not become better at understanding. The distinction between copying and understanding, in Collins's account, is not a technical limitation but a sociological fact about what knowledge is and how it is produced.
Mimeomorphic and polimorphic action. Collins and Kusch's distinction between actions whose correctness depends on copying surface behavior and actions whose correctness depends on understanding social context. AI systems excel at mimeomorphic competence; most of what matters in expert practice is polimorphic. The division explains why AI can reproduce the surface of expert output while failing at the social judgment that makes the output genuinely expert. See Mimeomorphic and Polimorphic Action.
Three species of tacit knowledge. Relational tacit knowledge is tacit for contingent reasons; somatic tacit knowledge resides in the body; collective tacit knowledge resides in the social group. AI training absorbs much of the first, cannot access the second, and is structurally excluded from the third because collective tacit knowledge is acquired through socialization, not text processing.
Interactional versus contributory expertise. Interactional expertise is the fluency acquired through sustained linguistic engagement with a community—genuine but limited. Contributory expertise is the capacity to advance the practice. LLMs possess interactional expertise at unprecedented scale; Collins's 2026 paper demonstrated they cannot reproduce the contributory expertise that requires collective tacit knowledge.
The Surrender. Collins's term for the cultural tendency to defer to computers because their outputs look competent and humans, confronted with confident-sounding output, tend to accept it rather than challenge it. The Surrender is not about machine power but human deference—the abdication of the evaluative authority that would distinguish interactional fluency from contributory expertise. [YOU] on AI identifies the same risk when Segal describes almost keeping Claude's passage on democratization because it sounded better than it thought.
Socialization vs. training. Collins's 2025 formulation: 'Learning from the internet is not the same as socialisation.' The internet contains the textual residue of social life. Socialization is the process of living inside a community—being corrected, praised, embarrassed, and gradually absorbing norms that no text specifies. A model that has processed all text ever published has undergone no primary socialization and cannot possess the collective tacit knowledge that socialization produces.