Establish the rule that genuinely applied to the task
Identify what students could access before submission. Which AI activities were permitted, limited or excluded? What disclosure was required? Does the course statement align with academic regulations, the task and marking criteria? An expectation written later must not be treated as a clearly communicated prior rule.
Preserve the submitted file, timestamp, appendices and relevant portal record. Investigate a copy. Record the trigger without turning it into a finding: “the detector marked paragraph four” or “style changes in section two” is more accurate than “AI cheating”. Check responsibility and the approved procedure before confronting a student informally.
Treat detector output, style and metadata as bounded signals
An AI detector classifies patterns under particular conditions. A reviewable record needs product, date, text state, language, settings, definition of the score and marked segments. Error behaviour varies with genre, length, language, editing and system state. Do not set an automatic decision threshold or claim accuracy without validation relevant to the actual task.
A style shift may reflect source genre, feedback, translation, extensive revision or a template. Metadata and version history document technical events, not necessarily the individual at the device. A large insertion warrants a specific question but is not a provenance finding. Search actively for alternative explanations and record how each was evaluated.
Process records answer different questions. Drafts may reveal argument development, research logs a source trail, tests programming activity and an AI-use log disclosed assistance. They also have limits. Agreement among independent records linked to the disputed passage is more informative than any isolated item.
Enable a response to identified passages and evidence
Provide the applicable rule, questioned locations and accessible supporting material through the institutional route. Ask open questions: How did this passage develop? Which source or intermediate state connects to it? Which tools were used and how were outputs checked? Avoid wording that presupposes an admission.
A disciplinary discussion can reveal understanding, but it should not become an improvised extra assessment without a policy basis. Use comparable questions grounded in the submitted work. Record the question, response, material shown and residual uncertainty. Anxiety, language differences or imperfect recollection are not provenance evidence by themselves.
Where confidential data, disability, multilingual writing or access barriers matter, involve the appropriate specialist service. Do not demand complete private accounts or conversation histories without established necessity and authority. The guide to preserving evidence after an AI concern explains proportionate process records from the student perspective.
Refer a balanced record to the authorised decision-maker
Create a case summary containing policy state, submission ID, numbered passages, methods used, evidence supporting and weakening the concern, alternative explanations, the student response and unresolved points. Separate system output, observation and inference. Include checks that produced no finding so the file does not create an artificially seamless chain.
Stay within role. An educator can document disciplinary observations while a panel or other body may own the formal determination. Preserve the version and recipient of the referral. Explain subsequent steps and designated contacts under the applicable procedure.
Quality check for a multi-source process
- The rule communicated before submission is identified.
- The report and submitted file are unambiguously connected.
- Every question has a locatable passage.
- Detector output is neither proof nor automatic exoneration.
- Alternative causes were investigated seriously.
- The response could address accessible evidence.
- Contrary evidence and uncertainty remain visible.
- The authorised role makes the decision.
The higher education practice hub connects institutional workflows. The passage framework for similarity reports develops source review. The one foundation connection is academic writing and evidence.