Direct answer
For teachers at an institution with Turnitin, use the built-in AI indicator as a flag, never as proof. For independent cross-checks use GPTZero's free tier plus one of ZeroGPT or Winston AI as a second opinion. Treat any single-detector flag as the start of a conversation with the student, not a conviction. The Stanford HAI 2023 study documented more than 60 percent false positive rates on essays by non-native English writers across 7 detectors.
AI detectors are useful signals and dangerous verdicts. Used as a trigger for a conversation about process and sourcing, they help teachers surface cases that deserve a closer look. Used as proof of misconduct, they produce predictable false accusations against the students least able to defend themselves. This piece covers which detectors to use, how to read the scores, and how to write an integrity policy that is both useful and fair.
Which detector to use
If your institution has Turnitin: the built-in AI indicator is the right default. Teachers already use Turnitin for plagiarism; the AI indicator sits in the same workflow. See Turnitin AI detection 2026 for what the indicator actually measures.
If you need a free cross-check: GPTZero's free tier is the most-used free detector in educational contexts. Running the same passage through two free detectors and treating agreement as the stronger flag reduces the single-detector false-positive risk. See GPTZero guide and ZeroGPT review.
If your institution does not have Turnitin: Winston AI, Copyleaks, or Originality.ai are the paid options most often adopted by schools, each at roughly $18 per month entry pricing. See best AI detectors in 2026, Winston AI review and Copyleaks review for the deep dives.
How to read the score honestly
All mainstream AI detectors read the same underlying signal: how predictable the text is (perplexity) and how uniform the sentence lengths are (burstiness). That signal overlaps with the signature of writing by non-native English speakers and by anyone whose prose style is clean and plain. The Stanford HAI 2023 research found more than 60 percent of essays by non-native English writers were flagged as AI by one or more of 7 leading detectors; 19 percent were flagged by all 7 detectors simultaneously, versus 0 percent for native-speaker essays. Any policy that uses detector scores as proof will produce this pattern of false accusations systematically against international students.
The practical reading: a high score is a flag worth investigating. Low scores do not prove the text is human-written. High scores do not prove the text is AI-generated. The appropriate next step after a flag is a conversation with the student about process, sources, and voice.
Writing an integrity policy that works
- State the detector's role explicitly: a flag, not a verdict. Students should know in advance that a detector score alone will not drive sanctions.
- Set a specific threshold for investigation, not for conviction. For example: scores above 70 percent on Turnitin's AI indicator trigger a short meeting; sanctions require independent evidence collected in that meeting.
- Build an appeal process that references the Stanford research by name. Non-native English students who are flagged should be able to point to an institutional policy acknowledgement of the known detector-bias problem.
- Document AI permission at the assignment level. "AI permitted with disclosure for drafting; prohibited for assessed sections" is clearer than a blanket policy.
- Train teachers in reading student work, not in using detector tools. The sustainable integrity signal is a teacher who knows their students' voices well enough to spot drift, not a classifier.
Frequently asked questions
What is the most accurate AI detector for teachers?
On vendor figures Turnitin's AI indicator and Winston AI both report high accuracy on their own test sets. On independent evaluation the gap between the leading detectors narrows considerably and all share the same bias patterns. The practical choice is driven by what your institution already has: in a Turnitin school, use Turnitin; otherwise GPTZero free plus a paid cross-check works at low cost.
Can I use a detector score as evidence of misconduct?
You can use it as a flag that triggers an investigation. You should not use it as proof of misconduct on its own. Policies that treat scores as proof produce predictable false accusations against non-native English writers, as the Stanford HAI 2023 study documented.
How do I tell AI-written work from a thoughtful human draft?
Two signals are more reliable than any detector: low burstiness (sentences all roughly the same length) and absence of specific detail that only this student would know. A thoughtful human draft usually carries at least one piece of concrete, personal detail. See natural AI writing for the structural markers.
Should I bother with AI detection at all?
As a flag, yes. As a process, use it alongside assignment design that is harder to produce from a generic prompt: specific-source citations, in-class draft components, oral defence for high-stakes submissions. The AI detection explained guide covers the broader argument.