Navigating AI Detection Scores: A Guide for Educators

A percentage in an AI report is not automatically the probability that a student cheated. Start with what that specific report measures, then examine the submitted work, the assignment rules, and the student's writing process.

Reviewed September 18, 2026. This guide separates the detector result from an academic decision and provides a practical review worksheet. Follow your institution's policy; the workflow below is an editorial aid, not a substitute for its procedures.

Understanding AI Detection Scores

In Turnitin's current report, the percentage describes the share of qualifying prose identified as likely AI-generated or AI-modified. It is independent of the similarity score. A result of 30% therefore should not be restated as a 30% chance of misconduct. Check the denominator, supported language, and report type before comparing numbers. See Turnitin's report guide.

Different products can use different scoring conventions. Do not average percentages from unrelated detectors as if they were interchangeable measurements. Save the original report, including the date, version when available, highlighted passages, and any stated limitations.

The Limitations of AI Detection Tools

These tools can offer a review signal, but they are not definitive. False positives occur, and performance changes with the detector, writing sample, language background, and benchmark. Turnitin's own guidance says an AI writing score should be used as one data point rather than a conclusion.

Turnitin suppresses exact scores and highlights between 1% and 19%, showing an asterisk because that range has a higher incidence of false positives. Its guide also says the model is not reliable for some non-prose formats. An asterisk is not a hidden disciplinary threshold; a missing report is not evidence of misconduct.

A 2023 study by Liang and colleagues found substantial misclassification of non-native English writing in the detectors and samples it tested. That is a reason to check fairness and context, not a current accuracy estimate for every product. Keep the study date, sample, and detector scope attached to any statistic you use.

Best Practices for Educators

Given these limitations, educators should adopt a holistic approach when interpreting AI detection scores:

1. Use Detection Scores as Indicators, Not Proof: Treat AI detection scores as one piece of evidence rather than definitive proof of AI authorship. Consider the context and other factors before making judgments.

2. Consider the Student's Writing History: Compare the flagged text with the student's previous work. Significant deviations in style or quality may warrant further investigation.

3. Engage in Dialogue: If a submission is flagged, discuss the findings with the student. This conversation can provide insights into their writing process and clarify any misunderstandings.

4. Stay Informed About Detection Tool Limitations: Regularly update your knowledge on the capabilities and shortcomings of AI detection tools to make informed decisions.

5. Implement Clear Policies: Establish and communicate clear guidelines regarding the use of AI tools in coursework to set expectations and maintain academic integrity.

A Review Worksheet for One Flagged Passage

Use a copy of the assignment rather than repeatedly rewriting the student's original. Record these items before drawing a conclusion:

• Rule: what assistance did this assignment allow, and where was that explained?
• Report: what was scanned, what did the percentage mean, and which passage was highlighted?
• Evidence: what drafts, notes, source annotations, or version history are available?
• Content question: can the student explain the highlighted claim, its source, and the choices made in revising it?
• Alternative explanation: could permitted translation, feedback, formulaic task wording, or a required template explain part of the pattern?
• Next step: what does the institution's process require, and what remains unresolved?

An absence of saved drafts should be recorded as missing evidence, not automatically converted into proof of authorship by AI. Students may have different working habits. Apply the same process consistently instead of asking only some students to produce records that were never required.

Example: A Score and a Source Problem Are Different

In a fictional case, a report highlights a paragraph in a history essay. The teacher notices that its cited source supports a narrower claim than the paragraph makes. The useful first question is concrete: "Which passage supports this sentence?" The student can then explain the source choice and revise the unsupported inference under the course rules.

The source problem exists whether the detector score is high, low, or unavailable. A change in score after revision does not establish who wrote either version. Keep the assessment of evidence separate from any investigation into permitted tool use.

Plan the Assignment Before the Next Report

Specify which stages allow AI assistance, whether a disclosure is required, and what process evidence students should retain. A short outline, annotated source, or explanation of one revision can make expectations clearer when it is part of the assignment from the outset.

For a broader explanation of scoring, see how AI detectors work. For a source-preserving editing exercise that can be used where assistance is permitted, see ten practical revision tips and the AI Humanizer workspace. The assignment's rules should decide whether a tool belongs in the process.