- The paper analyzes Anthropic's constitutional framework, defining corporate hierarchies and granting AI models the right to conscientious objection.
- It presents empirical evidence showing a 50% overlap between corporate and public constitutional drafts, highlighting differences in bias across social dimensions.
- The study reveals that corporate ethics lack democratic legitimacy, advocating for public participation and regulatory oversight in AI governance.
Critical Analysis of "Corporations Constitute Intelligence"
Introduction
"Corporations Constitute Intelligence" (2604.02912) rigorously interrogates the institutional and normative architectures instantiated by Anthropic's 2026 constitution for its Claude model. This extended commentary situates Anthropic’s constitutional framework as both a milestone in corporate AI governance and as a case study in the persistent limitations of privatized normative regulation for increasingly agentic AI systems. The paper’s legal-theoretical perspective dissects both the possibilities and pathologies of such constitutions, particularly considering the intersections of corporate interests, state power, and democratic legitimacy.
Structure and Substance of Anthropic's Constitution
Anthropic's 2026 Claude constitution represents the most sophisticated attempt to date by a private AI lab to encode a comprehensive system of values and behavioral constraints into a foundation model. Unlike minimal rule lists or post hoc content policies, this constitution provides explicit rationales for Claude’s behaviors, recognizes the unresolved questions surrounding AI consciousness and moral status, and codifies a principal hierarchy: Anthropic’s directives supersede operators, who in turn supersede end users.
The document also formalizes internal conscientious objection. It authorizes, and even encourages, the model to refuse actions antithetical to its constitutionalized values, including those requested by its own creators. This is a significant leap from simple constraint-based governance toward invoking analogies with human moral agents, and constructs a model of “AI integrity” that is non-procedural and aspirationally substantive.
Operational Limits and Structural Defects
Despite its breadth, the analysis identifies critical limitations. The most salient is the explicit exclusion of models deployed for military or other specialized applications from governance under the published constitution. In practice, versions of Claude used by the Department of Defense were not subject to the same moral and behavioral constraints, and Anthropic’s stances against mass surveillance and autonomous weapon deployment were ultimately contractual, not constitutional, in nature. The handling of these deployments illustrates the incapacity of voluntary norms to constrain action when they are not enforceable, particularly when state power intervenes.
The paper highlights that, even after being designated a supply chain risk by the U.S. government, models were still operational in military campaigns due to technical entanglement and institutional inertia. This sequence exposes how the scope and efficacy of corporate AI constitutions are ultimately conditioned by external state and corporate logics, not by their own internal reasoning.
Democratic Legitimacy and Political Community Deficit
A deeper critique targets the privately instituted, non-democratic nature of Anthropic’s normative authority. The "principal hierarchy" centralizes decision authority within Anthropic, abstracting away from any mechanism for public or democratic input. The author frames this as a severe “political community deficit.” The legitimacy for AI behavioral constraints thereby rests on corporate judgment and a veneer of transparency, not on procedural accountability to the governed public.
Yet, the Anthropic-Collective Intelligence Project's 2023 democratic constitution experiment, wherein a “public constitution” produced measurably distinct and less biased alignment outcomes with equivalent performance (Huang et al., 2024), empirically demonstrated that participatory design is both feasible and impactful. The evidence for technical viability and social benefit of democratic AI constitutionalization directly contradicts the production-side omission of such procedures in the 2026 document.
Implications and Forward Trajectories
The critique asserts that meaningful and legitimate AI governance cannot emerge from unilateral, voluntary corporate acts, regardless of their intellectual rigor or intent. The degree of ethical sophistication in Anthropic’s constitution serves as a placeholder for the yet-to-be-built public institutional infrastructure capable of contesting, revising, and authorizing foundation model norms. Relying on the “moral fortitude of a single CEO,” or even a currently principled corporation, exposes human-level and social-scale decisions to the contingent interests of private actors and to the expediencies of state contracts and coercion.
Implementing mechanisms—such as participatory framing of constitutional principles, statutory compliance requirements for alignment, or public adjudication of AI value disputes—emerges as a theoretical imperative. These mechanisms would mediate the translation of societal values into machine behavior with legitimacy and robustness unattainable in closed corporate processes. The analysis implies that the long-term future of AI normative order will depend less on voluntary transparency and more on public authorization and contestability—features that echo the historical trajectory of constitutional democracy itself.
Conclusion
"Corporations Constitute Intelligence" elucidates the sophistication and ambition of contemporary corporate AI constitutions while systematically exposing their insufficiency as sole instruments for AI governance in contexts of high societal and geopolitical stakes. The empirical evidence for public input in model alignment, contrasted with its organizational neglect, underscores the structural disjuncture between technical feasibility and institutional will. The author’s central claim is that the governance and legitimation of increasingly agentic AI will require the creation of public institutions and participatory processes attuned to the distinctive scale and impact of these systems. The Anthropic constitution is a sophisticated but ultimately transitory artifact on this trajectory, foregrounding the need for democratically authorized AI governance architectures.