The Duck Criterion in Artificial Intelligence: Evaluative Structure and the Attribution of Agency, Authority, and Responsibility

Read full text Summarise with VIRAT
Authors: Peter Kahl
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Contemporary discourse on artificial intelligence routinely attributes agency, authority, and responsibility to systems on the basis of behavioural performance. This article argues that such attributions are systematically misdirected. Behavioural sophistication does not determine epistemic or normative status; what matters is evaluative structure — defined as ℰ = (Ω(ℒ), ℛ(ℒ)), where Ω(ℒ) specifies the source of evaluative standards and ℛ(ℒ) specifies the conditions under which those standards may be revised. The article introduces a framework that distinguishes between epistemic maintenance systems, which optimise within inherited evaluative standards, and generative systems, which can revise those standards under conditions of evaluative breakdown. The Duck Criterion formalises this distinction as a constraint on inference: a system that exhibits adaptive, context-sensitive behaviour but cannot revise its evaluative standards is structurally a maintenance system, regardless of behavioural complexity. The article further argues that even where structural access to evaluative revision exists, resource constraints compress the effective space of evaluative alternatives — a phenomenon analysed as second-order rational inattention, extending classical bounded rationality to the level of evaluative selection rather than action selection. On this basis, three claims are defended. First, apparent intelligence does not entail epistemic agency, as optimisation can proceed without evaluative authorship. Second, the “responsibility gap” is largely a misdiagnosis: responsibility tracks control over evaluative structure and therefore remains with those who define and govern it. Third, epistemic authority requires participation in shaping evaluative standards, a condition not satisfied by current architectures. The article concludes that AI governance must shift upstream, from regulating outputs to specifying, controlling, and auditing evaluative structure as the locus of normative accountability.

Fields of Study & Subjects

Ethics and Social Impacts of AI Cybernetics and Technology in Society Philosophical Ethics and Theory

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
22 March 2026
Language
EN
Document type
Preprint