Detection Is Not Provenance: Why an AI detector score is evidence about production, not a record of the production process

Institutions treat AI text detector output as a finding about how a document was produced. This paper argues that a detector estimate and a provenance record are different kinds of evidence, and that improving a detector does not convert one into the other. The argument does not depend on detectors being inaccurate. Drawing on published evaluations, including research co-authored by researchers affiliated with one of the detectors evaluated, the paper documents that the evaluated binary detectors performed substantially better on fully AI-generated text than on AI-edited text, the mixed category that policies permitting limited assistance must distinguish. The paper further argues that a detector label does not by itself establish a policy violation, since the permission lives in policy while the detector was trained on text. It distinguishes watermarking, where part of the evidence is created at generation, from estimation performed entirely after the fact, and examines contemporaneous alternatives including revision histories and C2PA Content Credentials together with their published limits. Deposit contents Detection Is Not Provenance, v1.1, the principal analytical paper. Sources and Verification Method, a companion document recording the sources consulted, verification method, evidence grades, two positions considered and rejected, published editorial choices, and the limits of the record. Disclosure: FERZ, Inc. builds deterministic governance infrastructure for AI systems. This paper concerns the evidentiary status of AI text detection and does not describe a FERZ product.