When Silence Must Be Proven: The Burden of Abstention in Artificial Intelligence

Description

This work introduces accountable abstention as a missing and necessary primitive in the governance of artificial intelligence systems. Building on prior research on question-first cognition, cognitive permission, legitimacy gating, and governed silence, the paper argues that the ability of an AI system to justify and evidence its non-participation is as critical as its ability to reason correctly.

Contemporary AI governance frameworks increasingly recognize that illegitimate cognition—reasoning without sufficient epistemic grounding, authority, or consequence awareness—is a dominant risk. However, silence and refusal are typically treated as terminal safety outcomes rather than accountable epistemic acts. This work challenges that assumption.

We argue that systems which cannot demonstrate that they were capable of reasoning, evaluated legitimacy, and intentionally refrained from cognition under epistemic pressure are structurally unsafe, regardless of output correctness or downstream safety controls. Silence without accountability provides no epistemic assurance and cannot support regulatory, institutional, or audit requirements in high-stakes domains.

The paper formalizes the Burden of Abstention: the requirement that intelligent systems be able to prove that non-participation was possible, intentional, and structurally enforced prior to reasoning activation. It reframes abstention not as the absence of cognition, but as an active, governed epistemic act subject to accountability without reintroducing authority simulation.

This work is conceptual and architectural rather than algorithmic. It proposes no implementations or training methods. Its contribution lies in extending AI governance from legitimacy of reasoning to responsibility for restraint, establishing accountable non-action as a first-class constraint for intelligent systems operating in regulated and high-consequence environments.