Show Your Work: From Disclosing AI Use to Demonstrating Rigour. A Worked Case and a Contributor-Role Standard for AI-Assisted Scholarship

Background. Publishing has settled on one answer to generative AI: an AI cannot be an author, and its use must be disclosed. Problem. Disclosure asks the wrong question. A statement that AI was used cannot be falsified in either direction, and a large recent study suggests it is widely ignored anyway. Argument. The question worth asking is whether the AI was used well, in a way a reader can check. We set out three demands: coverage (a stated, reproducible search), verification that runs both ways and admits its limits, and a producible process record. Mechanism. The instrument is one optional attribute on the existing CRediT roles: a tier stating what a named human will certify, including a label for AI-executed work a human approved but did not verify. We also float, for the standards conversation, a candidate role for the human who directs and gates the AI; the instrument does not depend on it. Case. We applied the standard to one real AI-produced manuscript, against our own interest. The adversarial review caught four substantive errors the human director had missed. This is an existence proof that the failure mode is real; it does not show that AI review beats human review. Takeaway. Disclosure can establish that a tool was present. It cannot establish that the work was done. How this was made: AI-produced, human-directed and accountable. A human set the question, the standard of proof and the frame, directed the work, and gates what is published; AI drafted and ran the verification apparatus described in the paper. Independent analysis, not peer reviewed.