Upholding Epistemic Agency: A Brouwerian Assertibility Constraint for Responsible AI
The paper “Upholding Epistemic Agency: A Brouwerian Assertibility Constraint for Responsible AI” is available as a preprint on arXiv. Its central thesis is that responsible AI should not merely produce answers in high-stakes domains; it should be entitled to answer. Where no publicly inspectable and contestable warrant is available, a system should not turn uncertainty into authority. It should return Undetermined.
Generative AI speaks fluently, often with a confidence that exceeds what is actually warranted. In high-stakes domains, this matters. A system can transform uncertainty into an authoritative-seeming verdict, shifting attention away from the reasons, evidence, and contestation on which democratic judgment depends.
Inspired by Brouwer, the paper proposes an assertibility constraint for AI systems. A system may assert or deny a claim only when it can provide a public certificate of entitlement: a warrant that can be inspected, challenged, and revised. Without such a certificate, it must not force a conclusion.
This leads to a three-status interface: Asserted, Denied, and Undetermined. These statuses do not describe the world directly. They describe the system’s standing to speak categorically. A claim may be true, false, or still unsettled in the world; the question is whether the system has earned the right to present it as settled.
The certificate forms the bridge between internal computation and public accountability. It connects technical evidence, such as margins, bounds, or other witnesses, to a form that users can examine and contest. Entitlement is therefore time-sensitive: stable when numerical estimates improve, but open to revision when the public record changes.
Operationally, the paper translates this idea into a decision layer. Confidence scores or argmax outputs do not become final answers on their own. They must first pass through an entitlement check. If an adequate witness is present, the system may assert or deny. If not, it abstains with a reason-coded Undetermined status.
A central result clarifies why this is not just a cautious design choice. Any binary interface that always answers, while claiming to be certificate-sound, already decides the predicate across its declared scope. Where no adequate witness exists, a forced answer simulates certainty without entitlement. Undetermined is therefore not a fallback option, but a necessary public status.
At its core, the paper argues for AI systems that remain answerable to reasons. Against the persuasive force of automated speech, responsible AI must preserve the space in which claims can be questioned, justified, and publicly held to account.