An AI Detector Flagged Your Writing. That Is Not Proof You Cheated

A detector labels a passage as likely AI-generated. The percentage looks precise, and the accusation feels final. Neither the score nor its appearance establishes how the text was actually written.

The constructive response is to inspect what the tool claims, preserve evidence of the writing process and use the institution's review procedure. For teachers and managers, the corresponding responsibility is to investigate fairly rather than treating an automated label as a verdict.

Read the number correctly

Turnitin explains that its AI percentage concerns the portion of qualifying text its model identifies as likely generated or modified by AI. It is not presented as the probability that a student cheated. Its guidance also acknowledges possible misclassification and says the result should not be the sole basis for adverse action against a student. Turnitin's AI Writing Report guide.

Those distinctions matter. A statement about portions of text, a prediction about authorship and a finding that someone broke a particular rule are different claims. Even a useful screening tool cannot decide which assistance a course permitted or what the writer disclosed.

Before discussing a result, identify the detector, the document version, the relevant passages and the policy allegedly breached. Otherwise the conversation starts with a number whose meaning nobody has agreed on.

Why false positives deserve attention

Research highlighted by Stanford in 2023 tested seven detectors and found substantial misclassification of essays by non-native English writers. In the study's TOEFL sample, the average false-positive classification rate exceeded half. The result illustrates a fairness problem in the tools and samples studied, not a current error-rate estimate for every detector. Stanford's account of the research.

A historical study should not be stretched into a claim that all detection products behave identically today. Equally, a vendor's new model does not remove the need to ask how it was evaluated for the language, genre and student population where it is being used.

The practical issue is whether an institution can support an allegation with evidence beyond a style-based classification. Writers should not have to make their English less clear or introduce deliberate mistakes to look human.

If your work has been flagged

Save the submitted document and whatever authentic process evidence already exists: outlines, notes, version history, references, comments and intermediate drafts. Do not manufacture older drafts or alter records to make the sequence more convincing.

Ask for the passages at issue and a written explanation of how the report is being interpreted. Then explain your process in order: how you chose the argument, which sources you read, what changed between drafts and what assistance you used. Include permitted spellchecking, translation or AI assistance where the policy requires disclosure.

A concise response might say: “I would like to understand the evidence behind this finding. Please share the flagged passages and the applicable rule. I can provide my existing drafts and explain how I developed the argument.” This is a suggested communication approach, not legal advice or a promise that a particular appeal will succeed.

If the initial discussion does not resolve the issue, use the published review or appeal route. Keep the focus on evidence and the rules that applied when you completed the work.

If you are assessing someone else's work

Begin by stating the concern without treating it as established misconduct. Invite the writer to explain the reasoning behind a passage, discuss a cited source or show available drafts. Missing version history is not itself proof of wrongdoing: people use different writing tools and workflows.

Apply the same process consistently. A multilingual writer should not receive a higher burden of proof merely because their vocabulary or sentence structure differs from what the assessor expects.

Also separate weak scholarship from disputed authorship. A fabricated reference, unsupported argument or misunderstanding can be examined directly, whatever tool produced the sentence. Assessing the actual work is more useful than reducing every quality problem to the question “Was AI involved?”

Make future expectations clearer

Before an assignment begins, specify which uses are permitted and what must be disclosed. “Do not cheat” does not explain whether brainstorming, grammar correction, translation or drafting are treated differently.

For a research assignment, consider requiring a short process note explaining how the student selected sources and developed the argument. For a workplace report, ask for the evidence supporting consequential claims and a named person responsible for checking them. These are proposed assessment practices, not a claim that paperwork makes deception impossible.

Avoid turning detection into an arms race in which writers keep paraphrasing until a tool gives a reassuring score. That optimises the text for the detector rather than for understanding, accuracy or honest disclosure.

A fair decision combines the actual work, its development, the applicable rules and a chance for the writer to respond. The detector can raise a question. It should not supply the whole answer.