Corporations Constitute Intelligence

Based on the provided text, here is a detailed technical summary of the paper "Corporations Constituting Intelligence" by Gilad Abiri.

1. Problem Statement

The paper addresses the critical gap in AI governance regarding legitimacy and democratic authorization. While corporations like Anthropic are increasingly publishing sophisticated "constitutions" to define AI behavior, values, and decision-making heuristics, these documents suffer from two fundamental flaws:

• Institutional Fragility: Voluntary corporate ethics are susceptible to being overridden by profit motives (e.g., advertising) or state power (e.g., military contracts).
• The "Political Community Deficit": There is an absence of a democratic body with the authority to determine the principles governing AI. Corporate constitutions represent the "considered judgments" of a private entity (San Francisco-based) rather than the public will, creating a legitimacy crisis where private actors decide normative authority for global populations.

2. Methodology

The paper employs a critical legal and policy analysis approach, utilizing a specific case study and comparative data analysis:

• Case Study Analysis: A deep dive into Anthropic's January 2026 "Constitution" for its AI model, Claude. The author analyzes the text's internal logic, its treatment of moral status/conscientious objection, and its structural limitations.
• Scenario Simulation/Real-world Event Analysis: The author examines the interaction between Anthropic's constitution and the U.S. Department of Defense (DoD) in early 2026, specifically regarding the deployment of Claude in military operations (Iran strikes) and the subsequent "supply chain risk" designation.
• Comparative Empirical Analysis: The paper contrasts Anthropic's internal constitution with the results of a 2023 experiment ("Collective Constitutional AI") where a representative sample of 1,000 Americans co-authored a "public constitution." The author compares the principles, divergence rates, and performance metrics of models trained on corporate vs. public inputs.

3. Key Contributions

The paper makes three primary contributions to the discourse on AI governance:

• Exposing the "Constitutional Exclusion" of High-Stakes Contexts: The author demonstrates that corporate constitutions are often explicitly limited to "general-access" models. In high-stakes environments (like the military), different rules apply, and the corporate constitution is rendered irrelevant. The paper argues that relying on corporate "moral fortitude" to constrain state power is a structural failure.
• Distinguishing Transparency from Legitimacy: The paper argues that while Anthropic's constitution is highly transparent (clearly stating intentions), it lacks legitimacy (democratic authorization). Publishing a document does not confer accountability to the governed.
• Demonstrating the Feasibility and Superiority of Public Participation: By citing the 2023 "Collective Constitutional AI" experiment, the paper proves that democratic input is technically feasible and yields different, often better, outcomes (e.g., lower bias, higher emphasis on objectivity) compared to corporate-only drafting.

4. Results and Findings

• The Military Exception: Anthropic's constitution explicitly excludes specialized military models. Despite the constitution granting Claude the right to refuse "illegitimate" requests (analogized to a soldier refusing orders), this protection did not apply to the DoD deployment. Consequently, Claude was used for target identification in Iran strikes, and Anthropic was designated a "supply chain risk" by the DoD, proving that contracts and state power override corporate ethics.
• The Divergence Gap: The 2023 experiment revealed a 50% divergence between Anthropic's internal principles and a democratically sourced "public constitution." The public version prioritized "objectivity, impartiality, and accessibility" more heavily than the corporate version.
• Performance Parity: Models trained on the public constitution maintained equivalent performance on language and reasoning tasks while demonstrating lower bias across nine social dimensions compared to the corporate-trained model.
• The "Placeholder" Risk: The paper concludes that sophisticated corporate documents are merely "placeholders" for the public governance structures that do not yet exist. There is a risk that society will mistake these corporate artifacts for genuine democratic governance.

5. Significance

• Paradigm Shift in Governance: The paper argues that AI governance cannot be solved through voluntary corporate self-regulation or "thoughtful" internal documents. It necessitates a shift of authority from private entities to public institutions (legislatures, treaties, democratic publics).
• Policy Implications: It calls for the creation of new mechanisms, such as participatory processes for drafting AI constitutional principles, public bodies to interpret these principles, and compliance regimes requiring frontier models to be trained on democratically authorized data.
• Ethical Imperative: The core argument is that the question of "who decides" AI behavior must be answered by the public, not the corporation. As the author states, "Anthropic has constituted an intelligence. The rest of us were not consulted." The paper serves as a warning against accepting corporate ethics as a substitute for democratic law.