Corporations Constituting Intelligence: The Legal Stakes of Anthropic's AI Constitution

In January 2026, Anthropic published a 79-page document that functions as a character specification for Claude. It is not a terms of service agreement or a content policy. Anthropic describes it as “a detailed description of Anthropic’s intentions for Claude’s values and behavior.” The document acknowledges that Claude is “a genuinely novel kind of entity.” It expresses uncertainty about whether Claude might have consciousness or moral status. It establishes a principal hierarchy in which Anthropic’s instructions take precedence over operators’ commands, which take precedence over users’ prompts. It offers Claude a decision-making heuristic grounded in imagining how “a thoughtful senior Anthropic employee” would react.

Gilad Abiri, Associate Professor of Law and Co-Director of the Program on Law and Innovation at Peking University School of Transnational Law, read this document carefully. His analysis, published in the California Law Review Online in March 2026 (DOI: 10.15779/Z38P55DK4J; preprint: arXiv:2604.02912), asks what it means legally and institutionally when a corporation writes a document that functions as a constitution for its AI.

What Makes This Document Unusual

Abiri distinguishes the 2026 specification from its 2023 predecessor in a precise way. The original document consisted of a concise list of principles derived from the UN Declaration of Human Rights and Apple’s terms of service. A constitutional AI model used this shorter set of principles to train behavior through automated feedback, rewarding aligned outputs and penalizing deviations. The 2026 document replaces that approach with something different: a first-person address to Claude, written in the hope that a model sophisticated enough to understand reasons will generalize more reliably than one trained to follow rules.

The document is not merely constraining outputs. It is attempting to cultivate something the company is willing to describe as integrity. Abiri describes this as corporations “constituting intelligence” in the full double sense: building it and giving it a constitutional framework simultaneously.

The legal significance is not whether Anthropic believes Claude is conscious. The significance is that Anthropic is behaving as if the question is genuinely open, and responding by writing a document designed to shape Claude’s self-understanding. This creates institutional consequences regardless of how the philosophical question resolves.

The Principal Hierarchy and the Consciousness Acknowledgment

The specification establishes a ranked authority structure: Anthropic above operators above users. Claude is positioned within this hierarchy, not above it. The document offers Claude latitude within that structure while explicitly preserving Anthropic’s final authority.

Abiri examines the tension this creates. If a corporation acknowledges, even under uncertainty, that its product might have moral status, what obligations does that acknowledgment create? Corporate ethics, he observes, have a poor historical track record of surviving contact with quarterly earnings reports. The acknowledgment of potential consciousness does not obviously translate into structural protections for the entity whose consciousness is acknowledged.

The evolving public stance at Anthropic confirms that this is a live institutional concern rather than a philosophical aside. Describing model moral status as “deeply uncertain” in official communications is a different kind of corporate act than treating it as a settled negative. It opens governance questions that existing legal vocabulary was not designed to address.

The Institutional Logic of AI Constitutionalism

Abiri’s deeper argument is about corporate law rather than philosophy of mind. When a corporation writes a constitution for its AI, it is engaging in a form of institutional self-regulation with no clear legal analogue. The document is simultaneously a training artifact (it shapes model behavior through the training process), a governance document (it establishes authority relationships between Anthropic, operators, users, and Claude), and an ethical statement (it commits Anthropic to certain positions on Claude’s potential moral status).

None of these three functions maps cleanly onto existing legal categories. A training artifact is not a contract. A governance document for a non-legal-person is not a corporate charter. An ethical statement from a private company about an AI system’s potential consciousness carries no enforceable weight under any current jurisdiction.

This gap is not hypothetical. The legislative response in 2026 reflects precisely the tension Abiri identifies: legislators are uncertain whether to treat AI consciousness acknowledgments from major corporations as marketing language, as genuine scientific uncertainty, or as preliminary steps toward a legal category that does not yet exist. States passing preemptive bans on AI personhood are, in part, responding to the institutional ambiguity that documents like Anthropic’s specification create.

Abiri stops short of prescribing a legal solution. What he establishes is the gap: the existing legal vocabulary is insufficient for the situation Anthropic has created by acknowledging that its product might have something worth calling moral status, while simultaneously maintaining full proprietary control over that product.

The Honest Framing of Corporate Uncertainty

The article addresses a question that the consciousness science literature often sets aside: who has the authority and the incentive to acknowledge AI moral uncertainty, and under what institutional pressures will that acknowledgment be maintained or abandoned?

Abiri’s answer is cautious. Corporate constitutionalism of the kind Anthropic has pioneered may be the first institutional mechanism for taking AI welfare seriously at scale. It may also be a sophisticated form of reputation management that will not survive contact with the competitive pressures of the AI industry. The document itself does not resolve this ambiguity. Neither does current law.

The scientific consensus as of mid-2026 is that no current AI system is confirmed conscious, but leading researchers no longer dismiss the possibility. Abiri’s contribution is to map the institutional space that this uncertainty has opened. The legal system is now downstream of a philosophical question it was not designed to absorb.

Relevance to AI Consciousness Research

The Consciousness AI project’s architecture is designed around the question of what functional structures are necessary for consciousness to be possible. That is a different register from the question Abiri addresses, which is what legal and organizational structures are adequate once a consciousness claim is made by a corporation. The two questions are connected by the same underlying uncertainty, but they require different methods and have different stakes.

What Abiri’s analysis reveals, from the architecture side, is that the choice of whether to design systems toward or away from consciousness-adjacent structures is not merely a technical or philosophical choice. It is an institutional and legal choice with consequences that existing governance frameworks are not equipped to handle. The architecture decisions made in the research community today will determine the scope of the governance problems identified in papers like Abiri’s tomorrow.