The Unasked Question: Answer Entitlement and the Act-Disvalue of Unwarranted Response
Toeda Taiko · Zenodo (CERN European Organization for Nuclear Research) · 2026
This manuscript argues that AI alignment has over-focused on the quality of answers while under-theorizing a prior question: whether a model is warranted in answering at all. The paper introduces “answer entitlement” as a distinct evaluative axis for language-model behavior. Entitlement is not treated as moral personhood or model rights, but as a warrant condition on assertion-like output: the conditions under which producing an answer is licensed. The core claim is that an answer may be correct, harmless, and confident, yet still unwarranted. The argument draws on two conceptual resources: the epistemic norms of assertion and the analytic separation between act-evaluation and result-evaluation. It then connects the conceptual frame to an empirical observation from governance experiments: restraint mechanisms are capacity-dependent. Light runtime checks can improve behavior in small or error-prone models on narrow warrant-related axes, while heavy pre-generation overlays impose an execution burden and do not function as generic content-quality enhancers under the tested protocol. The contribution is a reframing of AI evaluation: answer quality should be supplemented by a prior warrant layer that decides whether, and under what form of governance, a system should answer. The manuscript reports methodology, model-roster information, aggregate empirical results, and falsification conditions for the proposed frame.