Begin with a reliable use record
“AI was used” tells a reader very little. A useful record connects a named system and available version information to purpose, date, affected file or passage, the human decision and the verification performed. Your university’s current policy determines mandatory disclosure. These first tools capture information during the work rather than reconstructing it from memory after submission.
Separate different forms of writing support
Brainstorming, outlining, language editing and translation alter academic work in different ways. One generic disclosure cannot describe all four. Preserve the starting material, suggestion, accepted change and your review. The guides below show how to retain and explain the contribution for which the submitting author remains responsible.
Control literature, code and data work
An AI output can suggest search terms, code or analysis ideas, but it does not establish correctness. Open literature yourself, test code in the documented environment and compare analyses with the data and method plan. Group work adds another requirement: identify who operated the tool, checked outputs and approved each consequential decision.
Never confuse generated output with evidence
Fluent prose can contain invented references, distorted findings or quotations that do not support the claim. Treat the output as working material rather than an authoritative scholarly source. Trace factual statements to reliable originals and compare summaries section by section. Preserve enough of the checking route to explain both errors and corrections.
Protect confidential material and correct omissions
Transparency does not authorise entering personal, unpublished or contractually protected content into an external service. Check approval, purpose, minimisation and the institutionally accepted environment first. If a disclosure is missing, reconstruct it honestly and contact the competent office according to submission status. Never invent a chat history or silently rewrite the record.
Use the collection as one workflow
Create the use log and tools register before the first session. Then open the guide for the specific task. Verify every retained contribution through the source, quotation or summary workflows and carry only checked material into the methods description and final declaration. Before submission, reconcile logs, final files and disclosure. Use the protection and correction routes for confidential material or retrospective gaps.
Choose proportionate documentation
Match detail to the significance of the AI contribution. A discarded idea list requires less than code, translation or analysis retained in the submission. Consider effect on the claim, ability to check the output and need to reconstruct the stage. A full transcript may be inappropriate where it repeats protected material; a justified extract can instead preserve purpose, relevant input, decision and verification.
Avoid both extremes. A generic declaration without a connection to the work is not traceable, while publishing every chat can create privacy and rights problems without explaining your judgement. Link each use to a specific stage and reconcile the final declaration with the internal record.
Four questions before retaining an output
- Provenance: What input, source or file shaped it, and was processing permitted?
- Accuracy: Which original source, calculation or test confirms it?
- Contribution: Which selection, revision and disciplinary decision did you make?
- Disclosure: Where will the contribution become visible under current local rules?
If one answer is missing, do not carry the output into the submission unchanged. Research, test or rebuild it from verified material. A detector score cannot answer these questions or establish provenance and permitted independent work; process evidence is the stronger basis for transparent practice.