Beyond 'We Used ChatGPT': An AI Provenance Passport for Structured, Risk-Proportionate Transparency in Scientific Publishin (Preprint)

UNSTRUCTURED Artificial intelligence (AI) is increasingly involved throughout the scientific research life cycle, including literature discovery, protocol development, data processing, coding, statistical analysis, image production, manuscript drafting, and peer review. Publishing policies generally respond by requiring authors to disclose whether an AI tool was used and, sometimes, to identify the tool and its purpose. These declarations are necessary but insufficient. A statement such as "ChatGPT was used to improve the manuscript" does not establish which model or deployment was used, what material was supplied to it, which scientific decisions were affected, how its outputs were incorporated, or whether those outputs were independently verified. Existing reporting guidelines, contributor taxonomies, model documentation frameworks, and research-object standards address important parts of this problem but do not provide a life-cycle–wide and risk-proportionate record that can be linked selectively to supporting evidence. This Viewpoint proposes an AI Provenance Passport: a structured, machine-readable, and versioned record of consequential AI participation in research and publishing. The Passport would record six domains: system identity and access conditions; research stage, task, and autonomy; input and data exposure; output disposition; human verification and responsibility; and integrity, privacy, and reproducibility controls. It would use a three-layer architecture comprising a concise public summary, interoperable metadata, and restricted audit evidence, with reporting depth determined by the combined profile of materiality, autonomy, and risk. Reciprocal Passports should also document consequential AI use by reviewers, editors, and publishers. The proposal is a framework for evaluation rather than a validated integrity intervention. Multisector pilots should assess completion burden, reporting accuracy, reviewer comprehension, detection of privacy and methodological risks, equity effects, and costs. Provenance cannot establish that a scientific claim is true or prevent deliberate misconduct. It can, however, replace vague declarations with a more specific evidentiary account of what AI did, under whose supervision, and with what safeguards. Scientific publishing should therefore move beyond binary disclosure toward proportionate, interoperable, and auditable AI provenance.