Merton's 1942 essay 'The Normative Structure of Science' identified four norms governing the scientific community's institutional behavior: universalism (knowledge claims evaluated by impersonal criteria), communalism (findings shared as communal property), disinterestedness (institutional arrangements minimizing personal interest in outcomes), and organized skepticism (no claim exempt from critical scrutiny). These norms, later given the acronym CUDOS, are not descriptions of how scientists actually behave but institutional ideals—standards against which conduct is evaluated and deviations sanctioned. The norms function by creating a gap between aspiration and practice that the community works to close. They are the dams Merton helped the scientific community build around knowledge production, ensuring that inquiry serves cumulative understanding rather than private advantage. The AI community faces the question of whether to adopt these norms, modify them, or replace them with alternatives suited to commercial development.
The tension between communalism and privatization is the most visible normative contest in contemporary AI. The open-source movement (Llama, Mistral, community models) embodies the communal ideal—knowledge should be shared. The proprietary frontier models (GPT-4, Claude Opus, Gemini) represent the commercial reality—models are expensive to train, organizations need revenue, and unrestricted access carries safety risks. Each position has genuine justifications; the conflict is structural, not resolvable through better arguments. What matters is whether the community settles on norms that balance these pressures or allows one to dominate entirely.
Disinterestedness faces its own crisis in AI development, where the organizations producing research are the organizations commercializing it, and the financial stakes are measured in trillions. The structural separation between knowledge production and knowledge exploitation—a separation that disinterestedness requires—has collapsed more thoroughly than in any previous knowledge domain. The researcher evaluating a model's safety is employed by the firm profiting from deployment; the peer reviewer is employed by a competitor. The mechanisms Merton identified as guarantors of disinterested evaluation—independence of the evaluator from consequences—are largely absent.
Organized skepticism is under pressure from competitive timelines that compress evaluation below the threshold at which thorough scrutiny is possible. The norm demands that no claim be accepted without critical testing; the market demands speed. The acceleration produces what scholars have identified as a crisis of credulity: claims about AI capability and safety are accepted at face value because the institutional incentives reward confidence and penalize caution. The inversion of organized skepticism is not intentional—no one advocates for credulity—but it emerges from the structural pressure of competition operating faster than the institutional capacity for skeptical evaluation.
Merton published 'Science and the Social Order' in 1938 and 'Science and Democratic Social Structure' in 1942, later combining them into the canonical statement of the normative structure. The timing was not incidental: Merton was writing as science's relationship to state power was being transformed by World War II, and the norms were partly an effort to articulate what made science different from propaganda, ideology, or politically directed inquiry. The norms were simultaneously descriptive (this is how the scientific community actually aspires to function) and prescriptive (this is how it must function to remain science rather than becoming something else).
John Ziman's 1970s coinage of the CUDOS acronym made the norms more memorable, but it also contributed to their misreading as a simple checklist rather than a complex institutional structure. Merton emphasized that the norms are in tension with each other—communalism conflicts with the priority system's reward of individual credit, universalism conflicts with the reality that institutional prestige affects how claims are evaluated—and that the community's actual practice involves continuous negotiation of these tensions rather than perfect adherence to abstract ideals.
Universalism. Knowledge claims evaluated by impersonal criteria, independent of the claimant's social characteristics—violated when institutional prestige or demographic identity shapes evaluation.
Communalism. Scientific findings as communal property to be shared and built upon—violated by secrecy, proprietary data, and the privatization of knowledge produced through public investment.
Disinterestedness. Institutional structures minimizing the influence of personal interest on knowledge evaluation—violated when the evaluator has financial or reputational stakes in the outcome.
Organized Skepticism. The institutionalized commitment to subjecting all claims to critical scrutiny—violated when competitive pressure compresses evaluation timelines below the threshold of adequate testing.
Norms as Aspirational Standards. The norms are not descriptions of actual behavior but ideals that create the standards against which deviations are identified, enabling communities to recognize and sanction conduct that undermines collective goals.