The public release of diffusion-based text-to-image models has produced a new aesthetic regime: outputs whose perceptual quality can be near-indistinguishable from human-made artifacts, generated at industrial scale. This convergence yields what we call a collapse of the criterion: customary markers that support robust claims of authorship, value, and responsibility (skill, effort, medium constraints, provenance) become epistemically fragile. Against two unsatisfying extremes—(i) humanist essentialism that treats AI outputs as categorically non-art, and (ii) aesthetic flattening that treats all outputs as equivalent—we propose a third path: strong aesthetic value depends less on perceptual features than on the traceability of situated intention and attributable responsibility. We formalize this claim through the Intention Traceability Metric (ITM), a five-dimensional forensic heuristic that scores decision-making, material constraint, exposure to error, ontological cost, and attributable responsibility. A pilot coding study on sixteen cases suggests substantial inter-rater agreement and discriminative capacity between minimal prompting, intermediate human–AI workflows, traditional works, and historically decisive conceptual gestures. We then read contemporary “forensic” technologies—Glaze, Nightshade, and Content Credentials (C2PA)—as socio-technical evidence that markets and institutions are rebuilding provenance under generative conditions. Finally, we treat Duchamp’s readymade as a crucial test case: it secures strong authorship despite low material labor because its performative gesture is traceable, risky, and responsibility-bearing—features that generative models can imitate in appearance but not instantiate as an act. We close by outlining implications for intellectual property, art markets, and aesthetic theory.