AI detector false positive rates: every published number in one table

The answer

There is no single false positive rate for AI detectors. There is a vendor figure, measured on the vendor’s own terms, and a set of independent measurements on specific populations that come out far higher. Both are published and both are quoted in integrity hearings, usually by different sides. The table below puts them together with their scope, so that a committee can see what each number does and does not claim.

Why the question comes up

Detector use has grown with detector output. Turnitin reported on February 24, 2026 that 15 percent of submissions between October 2025 and February 2026 were more than 80 percent AI, up from 3 percent at launch in 2023, and that 19.4 percent of US higher-education submissions in the second quarter of 2026 were heavy-AI. A Guardian freedom of information request found about 7,000 UK academic misconduct cases involving AI in 2023 to 2024. Every one of those numbers is a human being reading a percentage next to a student’s name, and the false positive rate is what decides how many of them are wrong.

Source: Turnitin press materials, February 24, 2026; Guardian FOI, 2024; checked 2026-09-08.

Every published number

Source: turnitin.com; Liang et al., Patterns, 2023; Common Sense Media, 2024; Waterloo announcement, September 2025; Giray, 2024; Guardian, 2024; Nouse, March 2026; all checked 2026-09-08.

How to read the table

The vendor number and the independent numbers are not in contradiction; they measure different things. Turnitin’s figure is a document-level rate across a broad corpus, scoped to documents the tool already scores above 20 percent. The Stanford figure is a rate on one population, non-native English writers, across several detectors. The Common Sense and Nouse findings are about who ends up accused, not about the detector’s arithmetic. Read together they say something simple: the average rate can be low while the rate for particular students is very high, and the students who bear the difference are the ones writing in a second language, writing formally, or writing with a neurodivergent profile.

The base rate arithmetic

A rate under 1 percent sounds like a rounding error until you multiply it. Take a mid-sized campus and assume 20,000 written submissions run through the detector in one term. That is the stated assumption; adjust it for your own institution.

• At exactly 1 percent, 20,000 submissions produce 200 papers flagged that were not written by AI.
• Use Turnitin’s own February 2026 figure that 15 percent of submissions are heavy-AI, and 17,000 of the 20,000 are not. At 1 percent, that is 170 wrongly flagged papers from the honest group alone.
• Even at half the stated ceiling, 0.5 percent, the honest group yields 85 flagged papers a term, or roughly 170 a year.

Every one of those papers is a student who receives a message, gathers drafts, and asks for a meeting, or does not and accepts a penalty. None of them can be identified from the score, because the score is the same for a true positive and a false one. That is the problem an integrity office has to design around, and it is why several institutions have turned the indicator off; the list with dates is on this site.

What to do with a score

1. Treat it as a reason to look, not as a finding. Write that into policy so that students and instructors read the same rule.
2. Show the student the report and ask for their account of the work: drafts, notes, version history, and a conversation about the argument.
3. Weigh the population effects. A non-native writer or a student with an accommodation is more exposed to a false flag, and the table above says by how much.
4. Where a record of the writing session exists, read it before the score. It answers typed versus pasted directly.

What a writing record adds, and its limit

A detector reads finished prose. A keystroke-level record of the writing session shows how that prose came to exist: what was typed and what was pasted, how long the pauses ran, which sentences were rewritten, and the order in which the argument formed. It has no false positive rate in the detector sense because it does not classify anything; the instructor reads it and reaches their own view. Its limit is just as clear. It shows how text entered the editor, not whether the ideas were the student’s, and a student who drafted elsewhere and pasted in will show as pasting. It is evidence for a conversation, not a verdict. Used that way, it replaces the score with something both parties can look at together.

What not to do

• Do not quote a single rate without its population and its source. The number changes the argument depending on which one you pick.
• Do not treat a second detector as a check on the first. Two probabilities that disagree tell you the tools disagree, not who wrote the paper.
• Do not let the score be the only evidence in a finding. In Newby v. Adelphi University (New York, early 2026) a court found a finding resting on a Turnitin 100 percent flag to be without valid basis and ordered it expunged.

Source: Justia, 2026 NY Slip Op 26021; checked 2026-09-08.