Turnitin's public FAQ (updated October 2026) reports a document-level false positive rate below 1%, and a Turnitin blog, June 2023 put the sentence-level rate at approximately 4%. Both figures come from Turnitin's internal testing of academic writing. The FAQ does not include the size of that test set, the composition, or the specific model configurations tested.
Quick answer
Turnitin currently reports a document-level false positive rate below 1% (AI writing FAQ, updated October 2026); its most recent published sentence-level rate is near 4% (Turnitin blog, June 2023). The model has been updated since 2023 but the methodology behind the numbers has not been published externally. The sentence-level rate and the disclosure gap are the two issues that have not moved and that continue to drive institutional reviews.
The 2026 landscape does not turn on whether the classifier improved. It turns on whether the improvements have been documented well enough for a research university to defend integrity cases built on the scores. Here is the current state.
The current Turnitin claim
The company has separately published a false positive explainer and additional support material. The explainer does useful work on how to read the score. It does not add methodology detail beyond what the FAQ provides.
The classifier has improved. The disclosure has not. That is the gap institutions are still working around.
The Working Educators editors
What changed since 2023
The classifier has been updated more than once. Turnitin has referenced improvements in handling paraphrased text, mixed human-AI writing, and shorter passages. The document-level false positive figure has stayed at or below 1% across updates, per the company's public statements.
The user interface has changed. The AI score display, the way the flag is exposed to faculty, and the reporting features around it have all iterated. Some of those changes address the criticism Vanderbilt and others raised about how the flag was surfaced to instructors without adequate context.
What did not change
Turnitin has not published a sentence-level figure since the near-4% rate in its June 2023 blog post. On a five-hundred-word essay, a 4% sentence-level rate means several individual sentences may be incorrectly flagged even when the document-level judgment is correct. That is a real problem when faculty use highlighted sentences as evidence in a hearing.
The methodology disclosure has also not changed in the way institutions asked for in 2023. Turnitin has not published a paper describing the classifier's training set, evaluation set composition, or the specific per-population accuracy figures that would let a research university defend its use. That was the crux of the Vanderbilt objection and it remains open.
The sentence-level rate still matters
The document-level rate is the number the marketing usually cites. In classroom practice, the sentence-level rate is what faculty actually see. The interface highlights specific sentences as AI-generated, and those highlights end up quoted in emails to students, in referrals to integrity offices, and in the record of any resulting hearing.
A sentence-level rate near 4% distributed across the population of college essays means large volumes of individually mis-flagged sentences per term. That is manageable if the score is treated as triage. It is not manageable if the score is treated as evidence, and the interface does not always make the distinction clear.
Where that leaves 2026 institutions
Some institutions have kept Turnitin's AI detection enabled and paired it with policy language that limits how scores can be cited in hearings. That is a workable middle position. Others have followed Vanderbilt in disabling the feature and moved integrity work upstream to assignment redesign and disclosure norms.
A third group has begun evaluating alternatives that publish methodology. Proofademic's side-by-side comparison against Turnitin is one entry point. The Proofademic research paper (May 2026) reports a document-level false positive rate of 0.2% on human-authored academic texts and 0.5% on non-native English writing, with the methodology described in enough detail for external reviewers to evaluate.
Neither of these routes is a silver bullet. What both have in common is that they do not stake integrity on a single classifier score. That is the shift the 2026 evidence base supports, regardless of which vendor an institution ends up preferring.
Frequently Asked Questions
Has Turnitin's false positive rate improved since 2023?▼
The document-level rate has stayed at or below 1%, per the vendor's AI writing FAQ (updated October 2026). Turnitin put the sentence-level rate near 4% in a June 2023 blog post and has not published a newer sentence-level figure. Both figures come from internal test sets that have not been externally documented.
Is Turnitin still the default at most universities?▼
Yes, though the share of institutions treating the AI score as adjudicative evidence has fallen. Many institutions now use it as triage and pair it with policy limits on how scores can be cited in hearings.
Why does the sentence-level rate matter more than the document-level rate?▼
Because faculty interact with the sentence-level highlights, not the top-line document score. Individually mis-flagged sentences end up quoted in referrals and hearings even when the document-level judgment is correct.
What would close the disclosure gap?▼
A published methodology paper describing the classifier's training set, evaluation population, and per-subgroup accuracy figures. Turnitin has not published this to date. Some competing vendors have.
Should our institution disable Turnitin's AI detection?▼
That depends on how you use the score. If it is treated as triage and paired with assignment redesign and disclosure norms, it can stay. If it is being cited as evidence in integrity hearings without a corroborating record, the sentence-level rate is a real risk.
The Bottom Line
Turnitin's AI detection is not the worst tool in the category and it is not the best. The 2026 story is that the classifier has iterated while the disclosure has not, and the disclosure gap is what continues to drive institutional reviews.
The safest posture at renewal time is not on or off. It is a policy that treats the score as triage, pairs it with assignment-level defenses, and refuses to build integrity cases on any single classifier's judgment. That posture would hold regardless of which vendor an institution chose.
