"Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions" is a foundational academic essay published by the Knight First Amendment Institute (Knight Columbia) (Knight Columbia) on March 3, 2026. Its authors are Gillian K. Hadfield (Johns Hopkins / Schwartz Reisman), Rakshit Trivedi (MIT), and Dylan Hadfield-Menell (MIT). The essay introduces a named framework — the "democratic matrix" and the concept of "normative competence" — and sets out a technical research agenda for building AI agents that can participate in evolving democratic norms rather than being aligned to fixed preferences or laws.
Central argument
The essay argues that current technical approaches to democratic alignment for AI — preference aggregation, rule encoding, Constitutional AI, "democratic inputs" fine-tuning, and "law-following AI" — implicitly assume that democratic values can be exhaustively elicited and encoded as static parameters. The authors contend these approaches fail to address the computational challenges of participating in evolving normative contexts.
In the authors' framing, democracy is a dynamic complex adaptive system — a "dancing landscape" (a term they attribute to Kauffman) — constituted by the concurrent adaptation of independent agents responding to and anticipating the behavioral patterns of others. Any effort to elicit and encode human values, preferences, laws, or norms in AI, they argue, will fail not only to be democratic but, more fundamentally, to protect democracy's stability.
The central claim is that AI agents must possess normative competence: the computational ability to detect social sanctions, attribute them to specific behaviors, and adjust future actions accordingly (a concept the authors attribute to Hadfield 2024). On this view, agents must operate within the democratic matrix — from the Latin matrix, womb, denoting the context and structure that nurtures life and development — rather than being aligned with preferences or laws as such.
The Hadfield-Weingast framework: normative social orders
The essay grounds its argument in the Hadfield-Weingast (2012, 2014) theory of normative social orders. A normative social order is described as an equilibrium state in which a group of independent actors pursuing ordinary self-interested utility — not innately motivated by pro-social preferences — are incentivized and coordinated to take costly efforts to punish members who engage in behaviors classified as punishable. That classification is produced by a classification institution. The result is a reliable, stable set of behaviors patterned on the shared classification scheme: consistent with self-interest, agents avoid predictably punished behaviors.
The classification institution must satisfy several "legal attributes" drawn from Fuller's rule-of-law criteria: stability, generality, clarity, neutrality, and impersonal reasoning. These closely track the rule of law, but in the authors' framework they are good attributes because they secure the stability of a normative social order coordinated on a shared classification scheme, not because they realize the prescriptive ideal of the rule of law as such.
The authors describe the framework's distinctive move as focusing not on getting agents to comply with norms but on third-party punishment, which they characterize as a uniquely human behavior. In their account, if the punishment problem is solved, compliance follows, because people avoid punished behaviors. This treats norms as the output of an interactive system, produced in equilibrium by the behaviors of agents, rather than as exogenous primitives.
The Adam Smith connection: the impartial spectator
The essay draws on Adam Smith's Theory of Moral Sentiments as a metaphor for the cognitive capacity of normative competence: an imaginary observer — "impartial and well-informed" — whom each moral person carries around inside themselves, judging conduct against the standards of the community. A moral person, in this account, looks not to avoid actual blame but to avoid blame-worthiness, that is, doing anything that "though it should be blamed by nobody, is, however, the natural and proper object of blame."
The authors characterize the impartial spectator as not a passive capacity: it represents the use of cognition to direct attention to the behavior and assessments of others, evaluated not as mere data but as inputs to reasoning that guides appropriate judgment in society. This, in their account, is normative competence. Smith's moral agents, they note, are motivated to engage in upholding the normative social order; they actively condition their conduct on the avoidance of punishment and are incentivized to participate in punishment guided by complex assessment of appropriateness. On this reading the agents are not merely normatively competent but also normatively compliant.
AI agents as actors in the democratic matrix
Section 4 of the essay argues that AI agents will become actors in the democratic matrix even when built to engage in ordinary economic activity. In support, the authors cite OpenAI's "five stages" of AI (chatbots, then reasoners, then agents — stage 3, described as AI systems that can spend several days taking actions on a user's behalf); Mark Zuckerberg's projection of "hundreds of millions, billions of different AI agents, eventually probably more AI agents than there are people"; and Mustafa Suleyman's "Modern Turing Test," framed as "Go make $1 million on a retail web platform in a few months with just a $100,000 investment" — a test the authors argue necessarily implicates democratic norms in its execution.
The technical research agenda
Section 5 sets out concrete technical research directions along two lines.
For normative competence in AI agents, the proposed directions are sanction-detection mechanisms (computational capacity to detect when behaviors are being treated as punishable), attribution mechanisms (linking sanctions to specific behaviors rather than background noise), and behavioral-adjustment mechanisms (updating future actions based on sanction signals). The authors position these as distinct from preference aggregation, rule encoding, and Constitutional AI.
For normative institutions legible to AI, the essay proposes building digital equivalents of the classification institutions that resolve ambiguity, fill gaps, and produce shared classification schemes. These would need to operate at the speed and scale of AI — machine-readable, real-time, and interpretable to agentic systems — and to support voluntary participation in the distributed enforcement of rules and norms.
Relation to other work
The essay introduces two framework concepts: normative competence, presented as Hadfield's distinctive concept and the computational primitive for democratic AI, and the democratic matrix, the dynamic-complex-adaptive-systems framing of democracy as a "dancing landscape."
It engages critically with existing alignment approaches. Toward AI Alignment (Constitutional AI, RLHF, preference-aggregation) and Law-Following AI (citing O'Keefe et al. 2023), the authors treat these methods as load-bearing but insufficient. The agents-as-economic-actors claim connects to Agentic AI and frames the question of agents in democracy.
The essay also relates to two companion papers from the same Knight Columbia 2026 symposium. Compared with "AI as Social Technology" (Farrell-Shalizi), which emphasizes the lossiness and power of AI coarse-grainings as a feature of social technologies, the present essay emphasizes the dynamic-normative-classification dimension; the authors present the two as complementary rather than contradictory. Compared with "A Conceptual Model to Guide AI Risk Governance Strategies" (Mulligan-Marda-Wang), both share a sociotechnical-system orientation, and Hadfield et al. offer the constructive technical research agenda that Mulligan et al. argue is missing.
Relationships
- supports: Normative Competence, The Democratic Matrix, Classification Institutions
- engages-with: AI Alignment, Constitutional AI, Law-Following AI — treated as load-bearing but insufficient
- engages-with: AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026), A Conceptual Model to Guide AI Risk Governance Strategies (Mulligan + Marda + Wang, Knight Columbia, March 16 2026) — companion essays
- related: Agentic AI, Agent Architecture Patterns, Agent Autonomy Spectrum (5 Levels), Principal-Agent Problem Applied to AI, AI and Democracy, Gillian K. Hadfield, Dylan Hadfield-Menell, Rakshit Trivedi, Herbert A. Simon, Lon Fuller, Adam Smith, Knight First Amendment Institute (Knight Columbia)
- part-of cluster: Knight Columbia 2026 AI-in-democratic-society symposium
Sources
Raw Sources/Building AI for the Democratic Matrix A Technical Research Agenda for Normative Competence and Normative Institutions.md- Published at Knight Columbia: knightcolumbia.org