Triage the omission without editing the evidence
Establish four facts first. Has the assignment been formally submitted? Is replacement explicitly available under the current rules? Can you recover a contemporaneous record of the AI activity? Did any unverified statement, citation, calculation, code or interpretation enter the work? Preserve what exists before trying to answer from memory.
An incomplete use statement and a factual error are not the same problem. Your notice may need to address both, but keep the findings distinct. The recipient should be able to see what was missing from the process record and whether the academic content also changes.
Before submission: rebuild the final package
Freeze the present draft and work on a copy. Review the full workflow rather than searching only for generated sentences. Relevant assistance may have shaped brainstorming, an outline, search vocabulary, translation, language revision, software, data analysis or figures. Connect each material interaction to the place it influenced.
Recover exports, chat history, notes and document versions without altering them. Record the service, displayed model detail, date, purpose, retained output and subsequent human check. Where a fact rests only on generated text, find an appropriate original source or remove the claim. Run calculations and code tests again against the actual submitted inputs.
- Read the assessment brief and local rules that define permitted use and disclosure.
- Assign an identifier to every recoverable consequential interaction.
- Mark uncertain reconstructed facts explicitly.
- Verify quotations, references, numerical results, code and interpretations.
- Insert the disclosure in the designated statement, methods section or form.
- Create a new final file with a distinct version name.
- Check that appendices and internal locations match that version.
The AI-use declaration framework turns the recovered evidence into a readable statement. It is better to report that a model label was not preserved than to guess one. The correction must not introduce a fresh inaccuracy.
After submission: report, do not quietly replace
Closed submission changes the route. Save the exact submitted file, receipt, timestamp and existing interaction evidence. Keep those originals read-only. Draft any corrected disclosure or marked-up manuscript as a new document so that the sequence remains visible.
Find the official procedure for your assessment. The correct contact may be a course lead, examinations office, supervisor, online portal or designated integrity officer. Explain the omission concisely and ask which formal route should be used. Do not assert that replacement must be allowed; institutions and assessment settings apply different rules.
If you have already received a query, observe the channel and deadline stated in it. Answer the issue asked, retain originals and distinguish records from recollection. Do not send confidential research material in an ordinary message unless the recipient confirms an appropriate transfer method.
Write a factual correction notice
Identify the work and submission date. Then describe the precise omission: which AI activity was absent or inaccurately characterised? Give the service, visible model information, access date, task, affected locations, extent of retained material and human verification where supported. End by asking for the approved next step.
A neutral structure is: “My submission of [date] omits the following information from its tools declaration: I used [service] on [date] for [specific task]. The affected locations are [sections]. I retained [defined material] and checked it by [method and evidence]. I have preserved the submitted file and available interaction record unchanged. Please advise which established route I should use to file this correction.”
Avoid minimising phrases such as “only tested briefly” if suggestions shaped the question or argument. Avoid unsupported self-accusation too. State observed conduct, evidence and uncertainty. The responsible body applies its rules to determine permission, relevance and procedure.
Worked case: translation assistance found after submission
While archiving a dissertation, a student notices that the tools statement omitted AI-assisted translation of two paragraphs they had written. Submission closed the previous day. The student preserves the uploaded document, receipt and chat export. In a separate location list, they mark both paragraphs and compare them again with the original-language text and cited sources.
The correction notice names the tool, date, limited task, passages and verification. It does not demand a particular outcome or replace the file without permission. The student does not attach a chat containing restricted draft material until a secure route is confirmed. The institution can now distinguish a missing process disclosure from any substantive amendment it may request.
A group assignment needs an additional reconciliation against the member log. The group contribution and tool record identifies who conducted and reviewed each activity. The AI transparency hub provides disclosure methods for other uses. An AI text assessment addresses patterns in a document; it cannot reconstruct the real use history or replace a correction notice.
Preserve evidence without rewriting the story
- Keep original submissions, exports and receipts read-only.
- Date every new correction file and give it a distinct version identifier.
- Separate timestamped evidence from remembered details.
- Do not clean a version history to make the route look simpler.
- Verify or remove references that came only from generated output.
- Log each contact, channel, date and response status.
- Record the final decision and which correction, if any, was accepted.
The aim is a prompt, inspectable correction rather than an artificially complete retrospective dataset. An honest unknown, an unchanged original and a clear notice provide stronger evidence than invented precision.