Full conversation or extract: a decision matrix
These options are not a hierarchy of honesty. A carefully described redacted extract may provide better evidence than an unstructured, unsafe conversation dump. At the same time, “representative” must not become a way to show harmless exchanges while concealing interactions that materially shaped the research.
When a full captured conversation earns its place
A complete conversation is most informative when a small number of exchanges drove a central methodological operation. It exposes follow-up questions, failed responses and corrections that a single prompt would hide. Export promptly where possible, and record the product, displayed model information, access date and file format. Give the record an identifier that the methods section and tools register can cite.
Use precise language about completeness. A file may contain the whole conversation still stored under one chat, yet exclude deleted branches, edited prompts, attached documents or sessions conducted elsewhere. “Complete export of saved conversation C-04 on [date]” is more defensible than “complete evidence of all AI use” unless a wider claim can genuinely be demonstrated.
A conversation still needs interpretation. Mark material inputs and responses without silently rewriting the retained original. Add a short note explaining what entered the work, what was changed and how it was checked. A hundred pages of dialogue without a contents key or adoption notes transfers the burden of reconstructing the method to the examiner.
Select representative extracts by an explicit rule
Extracts are appropriate for large sets of comparable activities, such as keyword suggestions or sentence-level alternatives. Define the selection method before choosing attractive examples. Options include the first, a typical and a failed case; every interaction associated with a given chapter; or every response from which material was adopted. Report the covered period and number of logged activities only when your record supports those figures.
Explain omissions. For example: “Appendix C presents R-03, R-11 and R-17 as examples of an unchanged suggestion, a heavily revised suggestion and a discarded response. The activity log records the remaining operations serving the same purpose.” When privacy or rights require removal, identify the category of redaction without indirectly revealing the protected information.
A prompt-only appendix often leaves the most important question unanswered. Without a response and an adoption note, a reader cannot see whether the interaction affected the work. Include at least the response type, selection decision and human verification. An AI use log provides a structured place for these links.
Privacy, confidentiality and rights constrain disclosure
Redaction involves more than removing names. A combination of location, date, role and an unusual event may identify someone. Filenames and embedded metadata can also expose sensitive details. Inspect screenshots, HTML exports and PDF properties. State that redaction occurred, identify the broad reason and explain whether the methodological meaning of the record remains intact.
A full prompt may reproduce copyrighted text or images supplied as input. A precise description and a shorter permitted extract can be more appropriate than copying the material into a public appendix. This page is not legal advice; it identifies why maximum disclosure and responsible disclosure can point to different formats.
Give the appendix a readable, cross-referenced structure
Open with a short legend: purpose, selection rule, covered period, tools, displayed version details, notation for omissions and the local instruction followed. Give each activity a consistent block containing its ID, date, research task, prompt, relevant response, author evaluation, adoption and verification. Define any colour or annotation rather than assuming it is self-explanatory.
Where permitted, preserve the original files separately from the version formatted for submission. Line wrapping and redaction may alter the appearance of the appendix. Do not call a PDF tamper-proof simply because it is a PDF. A checksum or integrity procedure should be claimed only when it was actually applied and can be verified.
Connect IDs to the AI tools register and methods section. A cross-reference might read: “Keyword activity T-02 is represented by extracts C-03 and C-11; all publications cited were subsequently opened in the disciplinary database.” This makes both the system contribution and the scholarly source check visible.
Seven questions before submitting prompt evidence
- Does the local rule require or permit this precise form of appendix?
- Are methodologically important interactions included, rather than only convenient examples?
- Is the sampling rule transparent and consistent with the activity log?
- Does each item distinguish prompt, response, adoption, modification and verification?
- Have protected materials been handled permissibly and redactions labelled?
- Do identifiers point to the correct methods passage, register entry and retained file?
- Does the appendix describe its limits without overstating completeness or resistance to alteration?
The European Commission’s living guidelines frame transparency and reproducibility as elements of responsible generative AI use in research. UNESCO also emphasises a human-centred and privacy-aware approach. Together, these principles support a reasoned evidence choice rather than indiscriminate publication. The AI transparency hub provides the connected workflows, while academic writing and evidence explains the appendix’s place in the broader scholarly record.