A score reports classification, not who wrote the work
Tools use different models, thresholds and labels. A percentage may refer to proprietary confidence, an estimated text proportion or another category. Unless that definition and relevant validation are known, it cannot be converted into a probability that a student used AI. Disagreement between tools further limits interpretation.
Open with observable facts: the preserved file was tested by the named service at the stated time and produced the quoted output. Then state what does not follow. The result does not independently establish either human or machine authorship. It may prompt proportionate review, never replace it.
Preserve the test condition before contacting the student
Keep the exact submitted file unchanged and calculate a checksum. Record whether references, quotations, tables and appendices were included. Save product, visible version, language setting, input route, test date and full result wording. A screenshot provides interface context but cannot substitute for the file.
Check whether uploading the work was authorised. Unpublished student material should not be sent to additional services simply to gather more scores. Review processing terms and institutional approval first. Repeated tests after silent product updates may generate different outputs without clarifying the original concern.
Use language that leaves uncertainty visible
Prefer “output”, “classification” or “signal” unless a locally validated meaning has been established. Avoid “hit” where it sounds like a confirmed identification. Do not say “90 per cent AI” unless you can explain precisely what the provider claims that number represents. A document score does not necessarily describe individual sentences.
Descriptions such as “AI-like style” are also weak. Technical prose, standard methods, translation, language-learning background and intensive editing can influence surface features. Those alternatives do not prove authorship either; they demonstrate why stylistic impression cannot close the question.
Do not coach the student to lower the score. The object under review is the submitted version and its actual history. Rewriting for a detector may destroy useful evidence and turns a learning conversation into optimisation against an uncertain classifier.
State who will attend and what role each person has. Provide relevant material in advance where the real procedure requires it. A detached screenshot without the tested file prevents a meaningful response. Apply approved accessibility and support arrangements without treating them as evidence about authorship.
Invite a coherent account and proportionate records
Ask about topic choice, outline, literature search and central decisions. Relevant records might include notes, reference-manager entries, drafts, feedback, analysis code and named milestones. Each has limits: timestamps do not identify an author with certainty, and a large version history does not prove every sentence.
A subject conversation can illuminate understanding, but must follow the real assessment process. Nervousness, disability, language variation and imperfect recall should not become automatic signs of wrongdoing. Provide required accessibility arrangements and use questions announced or justified by the applicable procedure.
The fair AI-concern review supplies a broader process. The independent-work rubric combines process and product evidence. The higher-education practice hub connects related materials. The single foundation route is the evidence-review guide.
Close with a neutral record, not a verdict
Summarise the technical information, student’s account, material supplied and outstanding questions. Separate quotations from evaluation and allow factual corrections to the note. State the responsible office, deadline and next step only from a verified policy or official communication.
- Preserve the tested file and output.
- Record the score definition and unknowns.
- Check privacy and authority for further testing.
- Explain the signal without asserting proof.
- Invite process explanation and proportionate evidence.
- Do not invent local policy from memory.
- Record uncertainty and the authorised next step.
Responsible communication does not weaken academic standards. It protects them by ensuring that an opaque technical number does not displace evidence, context and an accountable human decision.