What to Do When a Student Disputes an AI-Detection Result

A disputed AI flag should trigger a structured review, not a debate about whether the software is “right”. Preserve the original report, identify what it actually says, gather independent process evidence, invite the student to respond and follow the institution's published procedure.

The objective is not to make the student disprove a percentage. It is to decide, on the applicable evidence and standard, whether a specific assessment rule was breached.

Start with the allegation, not the score

Write down the concern in policy language. “The detector says 74%” is not an allegation. A clearer statement is: “There is concern that generative AI may have been used to produce assessed prose without the acknowledgement required by section X of the module guidance.”

Then verify:

• which policy and assessment instructions applied;
• what AI assistance was permitted, prohibited or required to be declared;
• which detector, version and settings produced the report;
• what the percentage or label means in that product;
• which passages were eligible and highlighted; and
• whether the material met the tool's language, length and format requirements.

If those basics cannot be established, the report has limited evidential value.

The evidence checklist

No single item proves authorship. Look for a coherent pattern across independent evidence.

Detector material

• complete report, not only a screenshot of the score;
• original submitted file and extracted text;
• highlighted passages and exclusions;
• report date, model version where available and any known limitations; and
• a written explanation of why the passages created concern.

Writing-process material

• outlines, planning notes and reading notes;
• document history showing substantive development;
• drafts with meaningful changes rather than files created after the allegation;
• reference-manager or library records where relevant;
• source passages supporting factual claims; and
• required declarations of AI or editing assistance.

Contextual material

• the student's explanation of the thesis, sources and writing choices;
• earlier work, used cautiously and only where a fair comparison is possible;
• permitted accessibility, proofreading or language support;
• assessment design and whether answers were likely to be formulaic; and
• evidence inconsistent with the allegation, not only supporting material.

The UK's Office of the Independent Adjudicator says academic judgement should be evidence-based. Where detection software is interpreted, sharing both the report and the academic analysis with the student is good practice.

How to hold a neutral student meeting

The meeting should test understanding and gather information, not set a trap.

Useful questions include:

1. “Can you talk me through how you developed this argument?”
2. “Which sources changed your view, and where are they used?”
3. “What did your first outline look like?”
4. “How did this paragraph change between drafts?”
5. “Which writing or language tools did you use, and for what?”
6. “What did you understand the module's AI rule to allow?”

Give the student the concern and evidence early enough to respond meaningfully. Avoid asking them to explain an undisclosed detector mechanism or to recreate the work under pressure as the sole test. Communication style, anxiety, disability and second-language status can affect an interview without saying anything reliable about authorship.

UCL's current procedure gives an investigatory viva as one possible route for suspected unauthorised generative-AI use. That is an example, not a universal UK rule; institutions must follow their own regulations.

What counts as weak evidence

Treat these as weak in isolation:

• a headline percentage without highlighted text;
• a second free detector chosen after the first result;
• a writing style that feels “too good” or “too formal”;
• a common transition word or phrase;
• a lack of detailed version history where none was required;
• metadata with ambiguous meaning; or
• a student's nervousness in a meeting.

Likewise, a clean detector result does not prove that no AI was used. The process should consider all relevant evidence, including material that weakens the concern.

Documenting the outcome

A sound case note records:

• the exact allegation and rule;
• evidence considered for and against it;
• what the student said and any material supplied;
• the meaning and limitations assigned to the detector report;
• the applicable standard of proof;
• the decision and reasons; and
• review or appeal routes and deadlines.

Keep the finding separate from the penalty. A serious allegation does not justify lowering the evidence threshold. If the evidence is insufficient, record that conclusion without reframing “not proven” as a warning that the detector was probably right.

For UK-specific stages and examples, read UK academic-misconduct procedures and AI evidence.

Frequently asked questions

Should a student have to prove they did not use AI?

The institution should follow its published burden and standard of proof. A student can provide relevant process evidence, but a detector score should not silently reverse the burden into proving innocence.

Is document version history conclusive?

No. It can support an account of development, but histories can be incomplete or manipulated and some students work offline. Assess it with other evidence.

Should I run the work through several AI detectors?

Not routinely. Different tools use different definitions and thresholds, and repeated scanning can create confirmation bias. Use an agreed, documented method and independent evidence.

What if the detector report is the only evidence?

Turnitin's own guidance says its AI report should not be the sole basis for adverse action. If no independent evidence supports a specific policy breach, escalate for procedural advice rather than treating the score as proof.

Sources and further reading

• OIA: Academic disciplinary procedures
• OIA: Casework note on academic misconduct
• Turnitin: How should I review the AI Writing Report?
• UCL: Student Academic Misconduct Procedure
• Genutext: AI detection as evidence