When Disagreement Matters: Friction, Pluralistic Alignment, and National-Security AI

Morgan Livingston

Keywords: pluralistic alignment, AI governance, national security AI, contestability, securitization

TL;DR: Pluralistic alignment in national-security AI depends not only on model behavior but on the contracts, forums, evidence practices, and market structures that determine whether disagreement becomes accountable decision-making.

Abstract: Pluralistic-alignment work increasingly turns from model-output diversity to institutions. National-security AI is a stress case for that turn. Classification, operational urgency, and security-dilemma pressure structurally constrain public participation, transparency, and ordinary review. The result is what I call frictionless security: the public-private extension of Deeks and Eichensehr's frictionless government, in which state and private actors converge around capability, speed, secrecy, and lawful-use compliance while the architecture that would weigh those choices against civil liberties, humanitarian restraint, escalation control, democratic legitimacy, and AI safety is weak. Pluralism here is whether institutions can preserve conflict among public values, route that conflict through forums with legitimate authority, and make the result reviewable. Starlink in Ukraine, Project Maven, Apple v. FBI, and reported frontier-AI defense-contracting disputes show three failure patterns — private veto, absorbed dissent, and red lines that do not bind deployment — and one case in which conflict reached a forum with authority. Three lessons follow: pluralistic alignment depends on the forum that receives disagreement, on the market structure that determines whether refusal can be absorbed, and on whether technical evidence can be verified by outsiders. The contribution is to show that, in national-security AI, pluralistic alignment depends not only on model behavior but on the contracts, forums, evidence, and market structures that determine whether disagreement becomes accountable decision-making.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 131