What Turnitin claims
Turnitin has publicly described its document-level false positive rate as below 1% for documents containing substantial portions of AI text, and has been more cautious about sentence-level reliability, acknowledging higher error rates there. The company also suppresses low percentage scores entirely because its own testing found them unreliable. Those are meaningful engineering disclosures, and they are also, inevitably, measurements on Turnitin’s own test corpora.
What independent testing found
University validation studies and academic research tell a messier story. The Washington Post and several university writing centers documented honest student essays flagged at meaningful rates. Vanderbilt disabled the feature after internal review, writing publicly that it could not verify the claimed false positive rate and that the harm of a false accusation outweighed the benefit. Peer-reviewed work, most prominently Liang et al., measured false positive rates above 50% on essays by non-native English speakers across multiple commercial detectors. The pattern across all of it: accuracy on clean test sets is real, and accuracy on the hard cases that fill actual classrooms is substantially worse.
Vendor test sets are curated: clearly human text versus clearly machine text. Real submissions include formal academic register, heavy grammar-tool use, formulaic lab reports and second-language writing, all of which look statistically machine-like. The divergence is not fraud; it is the gap between laboratory and classroom.
What a percentage should mean in practice
Turnitin’s own guidance says the score is the start of a conversation, not evidence. Take that seriously in both directions. A 90% AI score on a student whose in-class writing matches the submission, who has drafts and version history, is most plausibly a false positive and the process evidence settles it. The same score with no drafts, a voice mismatch and invented citations is a different conversation. The score never changes; the surrounding evidence is what makes it mean something.
If you are on the wrong end of a flag
Move calmly and in writing. Request the specific report and tool version. Assemble drafts, outlines, version history and your prior writing. Cite Turnitin’s own guidance that scores are not sole evidence, and your institution’s policy language. Ask what false positive rate the institution validated before relying on the tool; most have no answer, and the question lands. The full escalation playbook, including the research to cite, is in our false positives guide. And for how the engine works under the hood, see what AI detector Turnitin uses.
How to run your own validation
Institutions and skeptical individuals can replicate the accuracy question without trusting anyone's marketing, ours included. Assemble three folders. Known human: essays written before late 2022, pulled from archives, ideally including second-language writers, because that is where detectors fail hardest. Known AI: fresh output from the current versions of two or three major chatbots, on prompts matching your real assignments. Known mixed: AI drafts revised by humans for ten to twenty minutes each. Run every document through the tool you are evaluating, record verdicts against ground truth, and compute two numbers separately: the false positive rate on the human folder and the miss rate on the AI folder. Institutions should demand both numbers per population, not blended into one accuracy figure, because a tool can hit 95% overall while flagging a quarter of your international students. One afternoon of this beats every vendor page ever written, and putting the results in your integrity policy makes the policy defensible in a way no citation to a vendor can.