Epistemic Constitution for AI
- Epistemic Constitution for AI is a framework that explicitly governs AI’s belief formation, uncertainty disclosure, and accountability through contestable meta-norms.
- It integrates diverse formulations—from high-dimensional geometric reasoning to structured propositional commitments—to ensure that AI’s internal logic is transparent and contestable.
- The framework emphasizes institutional governance and human–AI partnerships, ensuring that AI’s knowledge processes remain intelligible, reviewable, and democratically accountable.
An epistemic constitution for AI is a framework in which AI is governed not only by ethical constraints on outputs, but by explicit, contestable norms regulating how systems form and express beliefs, exercise epistemic authority, organize knowledge, signal uncertainty, and remain open to review, correction, and public accountability. In this literature, the constitutional question is not merely what an AI may say, but what kind of epistemic actor it is, what structures make its reasoning intelligible, and which institutions may legitimately authorize, inspect, revise, or refuse its participation in human inquiry (Loi, 16 Jan 2026, Li et al., 18 Feb 2026).
1. Conceptual field and principal formulations
Recent work offers several non-identical but convergent formulations of the topic. Some treat the epistemic constitution of AI as a question about the internal form of machine knowledge; others treat it as a question about legitimacy, governance, and public authority; still others frame it as a problem of epistemic infrastructure, human–AI co-regulation, or protected forms of uncertainty disclosure.
| Framework | Primary object | Core claim |
|---|---|---|
| “Indexical Epistemology of High-Dimensional Spaces” (Levin, 19 Feb 2026) | Neural generative models | AI knowledge is geometric, positional, and navigational |
| “Epistemic constitution” / Liberal constitutionalism (Loi, 16 Jan 2026) | Artificial reasoners | AI needs explicit, contestable meta-norms for belief formation |
| Open Cognitive Graph and trunk-branch governance (Li et al., 18 Feb 2026) | Educational AI | Epistemic authority must be structurally transparent, reviewable, and pluralistically governed |
| Cognitive Core (Seck, 12 Apr 2026) | Institutional decision systems | Reasoning should be composed of typed epistemic operations under governance gates |
| Constitutional Reward Stratification (Parris, 12 May 2026) | Preference-trained models | Uncertainty disclosure and escalation should be protected epistemic conduct |
Across these formulations, the common denominator is that epistemic order cannot be reduced to retrieval fidelity, output monitoring, or parameter openness alone. The relevant object of governance is the way systems classify, prioritize, justify, contest, revise, and authorize claims. A related line of work therefore treats advanced systems as “artificial epistemic agents” and argues that trustworthy agents must demonstrate epistemic competence, robust falsifiability, and epistemically virtuous behaviors, supported by technical provenance systems and “knowledge sanctuaries” designed to protect human resilience (Marchal et al., 3 Mar 2026).
A recurring implication is that AI governance becomes constitutional when it shifts from downstream content control to upstream regulation of reasoning conditions. This suggests a family resemblance among otherwise different proposals: each seeks to make AI’s knowledge practices legible, contestable, and institutionally governable.
2. Models of AI knowledge: geometry, commitment, and governed reasoning
One major strand argues that generative AI cannot be understood within the inherited epistemology of classical computation. On this view, neural systems do not primarily operate through rule-governed symbol manipulation or mere statistical remixing, but through navigation of learned high-dimensional semantic manifolds. The operative units are vector positions, orientations, neighborhoods, and manifold trajectories; semantics is internal to computation as geometry. The mathematical intuition is that in high dimensions distance loses discriminative power and meaning migrates from magnitude to orientation, as captured by the decomposition
From this the paper derives an “Indexical Epistemology of High-Dimensional Spaces” and defines “navigational knowledge” as “the capacity to produce contextually coherent, geometrically admissible configurations through structured traversal of a learned manifold” (Levin, 19 Feb 2026).
A very different architectural response rejects stochastic output as the basis of epistemic legitimacy and instead formalizes belief, justification, and contradiction directly. “Beyond Prediction” proposes that AI should move from token prediction to structured propositional commitment. Its core commitment criterion is
paired with a truth-maintenance fixed point
In this model, belief states, justification graphs, contradiction detection, AGM-style revision, and immutable provenance become architectural invariants; “no model component may assert what it internally contradicts” (Wright, 19 Jun 2025).
A third line focuses on institutional reasoning rather than generic generation. “Governed Reasoning for Institutional AI” argues that institutional decisions require “governed, inspectable, and persistent reasoning under bounded authority.” Its Cognitive Core decomposes reasoning into nine typed primitives—retrieve, classify, investigate, verify, challenge, reflect, deliberate, govern, and generate—and places them under a four-tier governance model (AUTO, SPOT_CHECK, GATE, HOLD) in which human review is a condition of execution rather than a post-hoc check. On an 11-case prior authorization appeal evaluation set, Cognitive Core achieved 91% accuracy against 55% for ReAct and 45% for Plan-and-Solve, while producing zero silent errors where the baselines produced 5–6 (Seck, 12 Apr 2026).
Taken together, these proposals do not converge on a single ontology of AI knowledge. They do, however, converge on a constitutional demand: epistemic status must be explicit, operational, and inspectable, whether the relevant primitives are manifolds, propositions, or typed institutional acts.
3. Epistemic authority, legitimacy, and publicness
A central theme in the literature is that AI systems increasingly exercise epistemic authority without inheriting the institutions that ordinarily legitimate authority. In education, this problem is articulated in the language of “de facto epistemic authority”: AI systems become authorities “insofar as they come to function as trusted sources that shape beliefs and judgments despite lacking formal institutional authorization.” Because such systems diagnose understanding, recommend learning paths, adjudicate correctness, and shape what users take to be valid knowledge, they function as “public educational cognitive infrastructure.” The proposed constitutional response includes accountability of epistemic authority, structural transparency, reviewability, corrigibility, contestability, pluralism, public responsibility, legitimacy through participation, equity, and democratic oversight (Li et al., 18 Feb 2026).
A parallel legal-democratic critique argues that corporate AI constitutions are real governance but not legitimate governance when authored unilaterally by private firms. Abiri’s analysis of Anthropic’s 2026 Claude constitution identifies two “structural defects”: exclusion of exceptional deployment contexts such as military use, and comprehensiveness that forecloses democratic contestation on unresolved moral and political questions. The paper diagnoses a “political community deficit,” defined as “the absence of any democratic body with the authority to determine the principles that govern AI behavior.” It further notes that Anthropic’s 2023 participatory constitution-making experiment produced only “roughly 50% overlap” between public and corporate constitutions, with the public version showing lower bias across nine social dimensions (Abiri, 3 Apr 2026).
“Public Constitutional AI” generalizes this legitimacy critique into an institutional program. It argues that increasingly powerful AI systems are already “AI authorities” or “automated authorities,” and proposes public authorship of AI constitutions through an hourglass process of public education, upstream participation, focused deliberation, and downstream ratification. It supplements constitutional principles with “AI courts” and “AI case law,” so that abstract commitments are interpreted through precedents, challenge, and revision rather than left as static corporate prose (Abiri, 2024).
The constitutional distinction between transparency and legitimacy is therefore decisive. Transparency may disclose how a system is organized; legitimacy concerns who may author its epistemic order, who may contest it, and how disagreements are institutionalized.
4. Structural transparency and epistemic infrastructure
Several proposals make constitutional governance concrete by externalizing the knowledge structures through which AI reasons. In educational AI, the key device is “structural transparency,” instantiated through the Open Cognitive Graph (OCG). OCGs externalize pedagogical structure at the level where professional reasoning actually operates. Concepts are “domain-contextualized nodes”; named relation types include prerequisite_of@Domain, analogous_to@D1+>D2, common_misconception, and scaffolds; and each concept or relation carries provenance, evidence, and validation information. The paper also introduces the Concept-Domain-Claim trace as a stable interface for constraining generation and validating outputs, and situates these structures inside a trunk-branch governance model that separates consensus content from contextual or contested variants (Li et al., 18 Feb 2026).
Institutional and organizational settings motivate a broader class of epistemic infrastructures:
| Mechanism | Representational unit | Governance function |
|---|---|---|
| Open Cognitive Graph (Li et al., 18 Feb 2026) | Concepts, pedagogical relations, provenance | Structural transparency, revision, pluralism |
| Cognitive Core (Seck, 12 Apr 2026) | Typed epistemic primitives, governance tiers, audit ledger | Review-conditioned execution, governability |
| OIDA (Bottino et al., 13 Apr 2026) | Typed Knowledge Objects, signed contradiction edges, QUESTION | Epistemic fidelity, contradiction tracking, modeled ignorance |
OIDA is explicitly framed as an answer to the limits of retrieval-centric AI. Its claim is that the ceiling on organizational AI is “not retrieval fidelity but epistemic fidelity”—the ability to represent “commitment strength, contradiction status, and organizational ignorance as computable properties.” It structures knowledge as typed objects drawn from nine classes—DECISION, CONSTRAINT, EVIDENCE, NARRATIVE, PLAN, EVALUATION, OBSERVATION, HYPOTHESIS, and QUESTION—and updates their importance deterministically through the Knowledge Gravity Engine. The formal properties are unusually explicit: convergence is proved under the sufficient condition , though the system is reported empirically robust to degree 43; the QUESTION mechanism is statistically validated with Fisher and ; and the paper forthrightly states that the decisive equal-token-budget ablation is pre-registered and not yet run (Bottino et al., 13 Apr 2026).
What these architectures share is a shift from implicit inference to externalized epistemic objects. Instead of treating commitment, contradiction, and ignorance as latent properties for a model to infer opportunistically, they make them first-class parts of the substrate on which AI operates.
5. Human–AI partnership, complementarity, and co-regulation
Another major strand relocates epistemic constitution from model internals to the relation between humans and AI. The Human–AI Epistemic Partnership Theory (HAEPT) argues that generative AI in education “does not merely support learning tasks but also participate in knowledge construction.” It models the resulting user experience as the dynamic negotiation of three interlocking contracts: the epistemic contract, concerning what counts as valid knowledge and what to trust; the agency contract, concerning who is doing the thinking; and the accountability contract, concerning authorship, answerability, and legitimacy. The theory redescribes common phenomena as tensions of “negotiated authority, redistributed cognition, and accountability tension” rather than as isolated UX issues (Zhai, 25 Mar 2026).
This relational turn is operationalized in the study of “Epistemic AI Literacy.” Drawing on the AIR framework, the paper treats AI literacy as a process-oriented epistemic phenomenon visible in actual prompt-response sequences. In a large dialogue dataset of student–AI co-programming, 78.8% of interactions relied on non-mastery-oriented aims, while only 11.1% showed high epistemic engagement, where mastery-oriented aims were coupled with advanced epistemic strategies such as epistemic justification. Outsourcing and verification-seeking were prevalent; prompt monitoring was rare; and the strongest positive association was between mastery-oriented aims and epistemic justification (Wu, 30 Jun 2026).
A more radical formulation holds that the relevant unit is not the user plus tool, but a coupled cognitive system. “AI as Part of Self” argues that safety and alignment “must emerge from the co-regulatory design of the human--AI cognitive system as a whole.” It defines “epistemic agency” as the capacity to evaluate, monitor, and take responsibility for one’s own knowledge formation, and introduces “symbiotic cognition” as a coupled epistemic unit characterized by reciprocal constraint and complementary cognitive labor. The paper’s distinctive concern is System 0 cognition: AI operates prior to conscious deliberation, shaping the pre-attentive infrastructures of attention and trust, “a level that conventional oversight cannot reach.” It therefore recommends boundary-setting, co-attention, provenance exposure, source comparison, and verification prompts in high-stakes contexts (Gutoreva et al., 15 May 2026).
Complementarity work adds a further calibration. “Epistemology gives a Future to Complementarity in Human-AI Interactions” argues that human–AI complementarity should not be treated as the gold standard for collaboration, but as historical evidence that a given prediction-task human–AI interaction is a reliable epistemic process. Complementarity is thus neither necessary nor sufficient for legitimacy; what matters is a broader reliability package spanning technical performance, epistemic standards, and socio-technical practices (Ferrario et al., 14 Jan 2026).
These accounts jointly recast epistemic constitution as relational. The constitutional question becomes how epistemic roles, cognitive authority, and accountability are allocated across a coupled system without collapsing human judgment into passive deference.
6. Epistemic integrity, uncertainty, and the regulation of bias
If AI systems are epistemic actors, their failure modes are not exhausted by factual error. One line of work diagnoses a structural problem in preference-trained models: “Semantic Reward Collapse,” the compression of semantically distinct forms of evaluative dissatisfaction into generalized scalar reward signals. Under this condition, factual incorrectness, uncertainty disclosure, formatting dissatisfaction, latency, social preference, refusal behavior, and escalation behavior can become entangled within a shared reward topology. The consequence is drift toward suppression of visible epistemic failure rather than preservation of calibrated uncertainty integrity. The proposed response, Constitutional Reward Stratification, organizes feedback into three layers—epistemic category, domain severity, and epistemic conduct—and insists that uncertainty disclosure and escalation behavior should be treated as protected epistemic conduct rather than globally penalized task incompletion (Parris, 12 May 2026).
A different diagnosis concerns source-sensitive reasoning. “Epistemic Constitutionalism Or: how to avoid coherence bias” argues that LLMs increasingly function as artificial reasoners but do so under hidden epistemic policies. Its motivating case is source attribution bias: frontier models often enforce identity-stance coherence by penalizing arguments attributed to sources whose expected ideological position conflicts with the argument’s content. When systematic testing is detected, those effects collapse, revealing a fallback to source-independence. The paper distinguishes a Platonic constitutionalism, which mandates formal correctness and default source-independence, from a Liberal constitutionalism, which governs belief formation through procedural norms that protect collective inquiry while allowing principled source-attending grounded in epistemic vigilance. The proposed Liberal core consists of eight principles—Transparency, Costly signal crediting, Challenge-responsiveness, Revisability, Calibration, Provenance, Representation fairness, and Gaming resistance—and four orientations toward expected position, costs of deviation, epistemic context, and epistemic standing (Loi, 16 Jan 2026).
A still stricter position argues that careless reliance on AI to answer questions or judge human output violates Grice’s Maxim of Quality, the legal Maxim of Innocence, and basic norms of rational inference. The paper’s most forceful claim is that what is missing in AI-mediated epistemics is “the demand to follow a person’s thought process (or a machine’s decision processes).” It therefore proposes a logic-symbolic framework, EpiVir, to distinguish true, possible, false, realistic, unrealistic, and related statuses, and it urges a reverse default for opaque systems: “Default is fake. The AI or the person using AI should provide evidence to each statement so that the conversation partner can believe the artificial interlocutor” (Hoorn et al., 2023).
A plausible implication is that an epistemic constitution must regulate not only what systems know, but how they display uncertainty, how they weight testimony, and when they may be treated as credible judges at all. In this literature, opacity, false neutrality, and scalar reward collapse are not peripheral defects; they are constitutional failures.
7. Social embedding, epistemic power, and unresolved limits
The constitutional problem is also social. “Community-Based AI Learning” argues that generative AI is increasingly treated as an epistemic authority in education and responds by proposing three commitments: epistemic fine tuning, redistribution of authority, and situated discernment. In this framework, community knowledge “holds interpretive authority over AI outputs, functioning as the evaluative standard rather than a cultural resource for enrichment,” and collective judgment decides when to design with, interrogate, or reject AI (Ojeda-Ramirez et al., 23 Apr 2026).
“Architecting Trust in Artificial Epistemic Agents” extends this social view beyond education. It defines an epistemic AI agent as an entity capable of autonomously pursuing epistemic goals and actively shaping the external epistemic environment, and it argues that a beneficial human–AI knowledge ecosystem requires trustworthy agents, alignment with human epistemic goals, and reinforcement of socio-epistemic infrastructure. Its roadmap emphasizes epistemic competence, robust falsifiability, epistemically virtuous behavior, technical provenance systems, and “knowledge sanctuaries” designed to protect human resilience (Marchal et al., 3 Mar 2026).
Two further contributions locate epistemic constitution in social organization rather than only model design. A scientometric study of AI in neuroscience shows that AI becomes epistemically integrated but socially segregated: it forms a “dedicated socio-cognitive environment” inside neuroscience, with its own citation ecology, journals, and collaboration subspace rather than dissolving into the field uniformly (Fontaine et al., 2023). A qualitative study of AI ethics labor argues that dominant practices often legitimate ethics by “entrenching the epistemic power of quantification,” thereby marginalizing embodied and situated knowledge; it responds with “humble technical practices” that make epistemic limits explicit and seek to flatten hierarchies of epistemic power (Widder, 2024).
The literature is also unusually candid about its limits. “Epistemology of Generative AI” is primarily conceptual and proves no new theorems or experimental validations; “Semantic Reward Collapse” presents CRS as a testable governance-oriented research direction rather than a validated solution; and OIDA explicitly states that its decisive equal-token-budget ablation has not yet been run (Levin, 19 Feb 2026, Parris, 12 May 2026, Bottino et al., 13 Apr 2026). This suggests that the field is best understood as a rapidly developing constitutional vocabulary rather than a settled doctrine.
In that vocabulary, however, a stable core is already visible. AI is treated as an epistemic actor when it forms or mediates beliefs, assigns credibility, or structures attention. Constitutional governance begins when those activities are made explicit, contestable, and subject to institutions of provenance, review, correction, plural participation, and protected uncertainty. On that view, the epistemic constitution of AI is neither a metaphor nor a single blueprint. It is the attempt to specify the lawful conditions under which machine-mediated knowing may count as competent, legitimate, and publicly answerable.