Compare evidential strengths by export type
“Evidential strength” here does not mean forensic or legal authenticity. It describes which aspects of an academic workflow an artefact can reasonably document: visible sequence, context, timing, technical labels and connection with the final work. The institution decides which form satisfies an assessment requirement.
A six-step capture workflow
- Identify relevance: retain interactions that influenced prose, method, code, data or selection decisions.
- Use the available native export: prefer it before relying solely on copied text or images.
- Record context: note the product, provider, displayed model information, date and academic purpose.
- Assign an ID: choose a neutral code such as C-007 and use it in the activity log.
- Record adoption: state which response was rejected, revised or incorporated.
- Add verification: identify source checking, tests, comparison or disciplinary editing.
Perform these steps close to the interaction. A later reconstruction may lose a displayed version label or the order in which alternatives appeared. If the system supplies no export, combine a contemporaneous log with the permissible visible records. Never call a partial capture the complete conversation.
A practical filename is `C-007_2026-09-04_search-terms.pdf`. It describes an activity without exposing a participant’s name. The same ID in the AI use log supplies the chapter, selection process and original-source check. The exported conversation therefore does not need to carry the whole method by itself.
Bridge the gap between a response and the submission
The largest evidential gap often lies after the chat. Showing a response does not reveal what entered the finished paper. Add an adoption statement. “Response 4 produced five candidate keywords; I selected three and searched them in Database X” documents a different contribution from “I fact-checked and rewrote two sentences from Response 4 for section 3.2.”
For writing changes, retain an authorised version comparison. For code, connect the chat ID with a file revision, human changes, execution environment and test record. For analysis, document parameters, input category and independent validation. When a model proposes literature, the scholarly trail leads to publications you opened and assessed, not to the conversation as if it were a bibliographic source.
An AI tools register can summarise several conversations that served the same function. IDs keep each underlying file retrievable. The resulting chain is readable: the register states function, the log records activity, the export shows the captured interaction, the methods section explains significance, and the submitted work reveals final adoption.
Do not infer integrity from the file extension
Where an approved workflow calls for it, a cryptographic checksum can detect later byte-level changes. Its claim is narrow: the checked file matches the bytes used to create the reference value. It does not prove who conducted the conversation or whether content was removed first. If you use a checksum, record the algorithm, value, creation date and location of the independent reference.
Keep an unchanged working copy and a separately redacted submission copy when that is permitted. Log the redactions. Version control may reveal changes to supporting text, but it does not establish identity or completeness on its own. A modest statement tied to observable evidence is stronger than a technical-sounding claim the procedure cannot sustain.
Plan for sensitive material before making copies
Conversations may include personal data, confidential research, unpublished ideas, copyrighted content or credentials. Do not generate additional cloud copies without considering the approved storage location, access controls, required retention and permitted disclosure to examiners. More copies can increase exposure without adding meaningful evidence.
Screenshots and screen recordings easily capture names, tabs and notifications outside the target conversation. Exports may include metadata or embedded attachments. Review more than the visible prompt text. A redacted copy should signal the omission and broad reason, while any unredacted record belongs only in a location where holding it is authorised and necessary.
UNESCO’s guidance foregrounds privacy and human responsibility. The European Commission’s living guidelines caution researchers about sensitive and unpublished information in generative systems. These principles show why a complete-looking public appendix is not a success if it compromises protected content.
A final check for both the record and the claim
- The export type and creation route are stated.
- Tool, displayed version, date, purpose and activity ID agree across records.
- The link between response, human revision and final submission is explained.
- Completeness, timestamp and integrity claims stay within what can be evidenced.
- Facts and citations lead to original publications rather than a model response.
- Sensitive content and metadata follow the controlling institutional process.
- The submission format matches current requirements for the specific assessment.
The AI transparency hub connects related documentation tasks. The foundation guide to academic writing and evidence places chat records within the wider source and method trail. Use the decision guide on documenting prompts in an appendix to decide which of the preserved material should actually accompany a submission.