AI Provenance and the Evaluation of Intellectual Work: Rational Risk, Social Meaning, and Human–AI Performance

AI can improve a piece of intellectual work while knowledge of its use can reduce the credit or trust that the work receives. Understanding that possibility requires separating what AI did to performance from what disclosure adds to evaluation. We connect these questions by asking what the evaluator intends to assess. Improvement relative to a producer’s unaided baseline may matter when the objective is better output; a different benchmark is needed when the objective is to measure unaided capability, assign intellectual credit, or allocate opportunities. A binary “AI was used” label supplies little information about those distinctions. Its relevance depends on the function AI performed, the authority the human retained, verification, and the object of evaluation. Negative reactions may reflect expected error or checking cost, an institutional requirement, social meanings concerning effort and authenticity, or a combination of these explanations. We examine how evidence can distinguish them without treating every penalty as stigma. We then follow evaluation back into production: anticipated disclosure can change whether and how people use AI, including whether they forgo useful assistance. As technology and institutions evolve, the practical question remains whether provenance-based treatment serves the evaluation’s stated purpose or substitutes a category judgment for more direct evidence.