Human Responsibility Under Artificial Intelligence Mediation: Evaluative Capacity and the Conditions of Independent Judgment

Artificial intelligence governance increasingly assigns humans responsibility for consequential decisions while AI systems increasingly mediate the information through which those decisions must be judged. This creates a responsibility-capacity tension: formal human oversight can remain intact even as the practical conditions for independent assessment weaken. This paper examines the normative consequences of that tension through evaluability, understood as the preservation of sufficient informational conditions for meaningful independent reconstruction and assessment. The paper argues that dependence upon evaluative capacity does not by itself establish an obligation to preserve it. Normative reasons emerge when institutions assign responsibilities requiring independent judgment, materially control conditions necessary for that judgment, and can reasonably foresee how their choices may preserve or degrade the relevant capacity. These responsibilities are distributed and differentiated according to control, institutional role, authority, foreseeability, and consequence. They are also proportional rather than absolute. Evaluability need not be maximized or override privacy, security, confidentiality, or other legitimate obligations. The relevant standard is sufficient evaluability for the responsibility and stakes involved. Responsible AI governance must therefore consider not only who remains responsible, but whether those responsible actors retain a realistic capacity to exercise the judgment assigned to them.