Dial E for Ethical Enforcement: institutional VETO power as a governance primitive

Based on the paper "Dial E for Ethical Enforcement: Institutional Veto Power as a Governance Primitive" (accepted at ICLR 2026), here is a detailed technical summary.

1. Problem Statement

The paper identifies a critical governance paradox in AI: while there is widespread ethical awareness and risk knowledge regarding the militarization of Large Reasoning Models (LRMs) and dual-use technologies, there is a near-total incapacity to prevent harmful downstream uses.

• The Core Issue: Current governance mechanisms (e.g., model cards, responsible AI statements, impact assessments) function as reputational signals rather than decision-making architectures. They generate knowledge ($K$) but lack the institutional authority ($A$) to block deployment.
• The Consequence: This leads to "organized irresponsibility," where actors can disclaim control while harms proceed. The paper argues that without formal authority to refuse transfers or deployments, ethical knowledge cannot translate into enforceable safeguards.
• The Gap: Existing literature treats militarization as a downstream consequence to be mitigated by ethics, rather than an architectural problem requiring the redistribution of authority. There is a missing "governance primitive": a formal Institutional Veto Power.

2. Methodology and Theoretical Framework

The paper employs a multi-faceted methodological approach combining theoretical modeling, comparative institutional analysis, and operational design:

• Formal Logic Modeling: The authors introduce a logical framework to demonstrate the failure of current systems.
  • Current State: $(K \land \neg A) \implies M$ (Knowledge without Authority leads to Militarization).
  • Proposed State: $(K \land V \land R) \implies \neg M$ (Knowledge + Veto + Risk Identification leads to No Militarization).
  • Here, $V$ represents a legally protected veto, and $R$ is the identification of credible risk. The model posits that $P(M | V) \ll P(M | \neg V)$.
• Comparative Governance Analysis: The authors analyze historical precedents in high-risk domains to validate the feasibility of veto power:
  • Nuclear Nonproliferation: The IAEA's mandatory inspection and technology transfer restrictions.
  • Biomedical Ethics: Institutional Review Boards (IRBs) that possess binding authority to halt research before implementation.
  • Export Controls: Mechanisms (e.g., ITAR, EAR) that restrict the transfer of dual-use technologies based on end-use commitments.
• Toy Model Simulation (Appendix A): A mathematical model simulating governance chains to demonstrate three failure modes of "ethics without veto":
  1. Responsibility Diffusion: As the number of actors ($n$) increases, individual enforceability ($\rho$) approaches zero.
  2. Adverse Selection: In a voluntary restraint environment, competitors who ignore ethics gain a funding/advantage ($\beta$), driving the equilibrium toward militarization.
  3. Ethics Laundering: Documentation ($D$) inflates perceived responsibility without actual intervention ($I$), creating a gap between perceived and actual harm prevention.
• Operational Design: The paper proposes concrete implementation pathways (University contracts, Conference policies, Funding conditions, Repository governance, and Professional norms) rather than abstract policy recommendations.

3. Key Contributions

A. Institutional Veto Power as a Governance Primitive

The paper defines Institutional Veto Power not as arbitrary refusal, but as a formal, procedurally protected authority to halt the transfer, licensing, or deployment of research when specific, credible risks (e.g., weaponization, surveillance of marginalized groups) are identified. Crucially, this veto applies to downstream use, not the inquiry or publication itself.

B. Centering Marginalized Community Leadership

A primary contribution is the argument that veto authority must be held by affected communities, not just elite academic or policy institutions.

• Epistemic Justice: Communities subjected to surveillance or military AI possess "situated knowledge" about harm salience that external experts lack.
• Structural Reform: Veto bodies must include representatives from affected communities with equal decision-making power to technical experts and ethicists. This shifts governance from "symbolic inclusion" to "binding community control."

C. Operationalization Pathways

The paper moves beyond theory to provide five concrete mechanisms for embedding veto power into existing infrastructure:

1. University Contract Modifications: Embedding use-restriction clauses and audit rights in technology transfer licenses.
2. Conference-Level Governance: Requiring dual-use disclosure forms and enabling machine-readable use conditions on published papers.
3. Funding Agency Leverage: Conditioning grants on the establishment of institutional veto committees and compliance with their decisions.
4. Repository-Level Enforcement: Implementing machine-readable metadata for use restrictions and automated delisting for violations.
5. Professional Solidarity: Establishing binding codes of conduct that protect researchers who refuse military funding.

D. Design Principles for Legitimacy

To prevent the veto from becoming a tool for suppression or political capture, the authors propose strict design constraints:

• Scope Limitation: Veto applies only to specific downstream uses, not research inquiry.
• Predefined Thresholds: Triggers must be specific (e.g., "autonomous weapons lacking human control") rather than speculative.
• Collective Decision-Making: Decisions must be made by pluralistic committees, not individuals.
• Appeal and Revisability: Mechanisms for appeal and sunset provisions to prevent permanent censorship.
• Anti-Retaliation Protections: Legal and career safeguards for those exercising the veto.

4. Results and Findings

• Theoretical Validation: The toy model demonstrates that without a binding veto point ($V$), the equilibrium of research trajectories inevitably shifts toward militarization due to adverse selection and the diffusion of responsibility.
• Comparative Success: Analysis of nuclear, biomedical, and export control regimes confirms that binding authority at decision points (not just guidelines) is the only mechanism that has successfully constrained dangerous technologies in the past.
• Failure of Current Ethics: The paper concludes that "ethics laundering" (documentation without intervention) is a structural feature of current systems, not a bug. Only converting symbolic safeguards into enforceable rules can stop the accumulation of harm.
• Feasibility: The proposed mechanisms do not require new treaties or governments; they can be implemented by repurposing existing university TTOs, conference committees, and repository platforms.

5. Significance

• Paradigm Shift: The paper shifts the discourse from "how to encourage restraint" (which relies on voluntary compliance and fails) to "how to make restraint structurally rational" (by altering the incentive architecture).
• From Ethics to Law/Policy: It argues that AI safety cannot be solved by technical alignment alone; it requires a governance primitive that treats the prevention of militarization as a legal and procedural necessity.
• Justice-Oriented Governance: By mandating that affected communities hold veto power, the paper offers a pathway to decouple AI development from the logic of militarization and surveillance that disproportionately harms marginalized populations.
• Actionable Roadmap: Unlike many theoretical papers, this work provides a "how-to" manual for researchers, institutions, and funders to immediately begin implementing enforceable governance structures.

In summary, "Dial E for Ethical Enforcement" argues that the militarization of AI is a governance failure, not a technical inevitability. It proposes Institutional Veto Power, led by affected communities and embedded in existing institutional contracts and workflows, as the necessary primitive to convert ethical knowledge into enforceable responsibility.