What Turnitin Has Publicly Said About Its False Positive Rate

Turnitin is the most-used AI detector in higher education, and it is also the vendor that has publicly acknowledged the most about its own limits. Between April 2023 and early 2024, the company revised its false positive claims, added disclaimers to the AI score, and asked institutions to treat the number as a starting point rather than proof. Reading what Turnitin has actually said, in its own words, is one of the most useful things a student facing an accusation can do.

What Turnitin has publicly said about its false positive rate

When Turnitin launched its AI writing detection feature in April 2023, the company stated a false positive rate of less than 1% at the document level. That number, published on the Turnitin website and repeated in company blog posts, was the headline pitch to universities. Within weeks, Turnitin began qualifying it.

In a May 2023 update, Turnitin's Chief Product Officer Annie Chechitelli wrote that the company had found a higher-than-expected false positive rate on documents where AI-generated text made up less than 20% of the total. In response, Turnitin added an asterisk to any AI score below 20%, indicating the result may be less reliable. The company acknowledged in the same update that the sub-1% figure was measured at the document level and that false positives occur more frequently at the sentence level.

The sentence-level problem the company acknowledges

A document-level false positive rate answers a narrow question: how often does Turnitin call an entire document AI-generated when it was not? It does not answer the question most instructors actually ask, which is whether specific highlighted sentences were written by AI. Turnitin has said openly that sentence-level accuracy is lower than document-level accuracy, which is why the tool aggregates sentence predictions into an overall score.

The practical consequence: an instructor who focuses on the highlighted passages, rather than the overall percentage, is looking at the least reliable output the tool produces. If your accusation rests on specific highlighted sentences rather than the overall percentage, the peer-reviewed research on what detector scores can and cannot establish is directly relevant.

What Turnitin tells institutions the score is not

In its guidance to administrators, Turnitin has repeatedly stated that the AI score is not evidence of misconduct on its own. The company's published position is that the indicator is a signal for further review, not a finding. From the Turnitin AI writing detection FAQ and its guidance materials for educators:

"Our AI writing detection capability is designed to help educators identify text that might be prepared by a generative AI tool. Our AI writing assessment is not designed to be used as the sole basis for adverse actions against a student."

That language matters in a hearing. Turnitin is the vendor whose tool produced the score. When the vendor tells institutions the score should not be the sole basis for a finding, an institution that treats it that way is acting against the tool developer's own guidance.

What changed after 2023

Turnitin has updated its detector multiple times since launch. The public position on limitations has become more, not less, cautious. The company continues to publish accuracy numbers, but pairs them with clear statements that the tool is one input among several and that human review is required. Several universities, most notably Vanderbilt, disabled Turnitin's AI detection feature in 2023 citing accuracy concerns. Vanderbilt's decision and the reasoning behind it are covered in a separate analysis of the Vanderbilt AI detection decision.

How to use Turnitin's acknowledgments in a response

Turnitin's own statements are the strongest source for the argument that a score is not proof. External researchers can be dismissed as academic critics. The vendor cannot. A response letter that references Turnitin's published guidance should:

• Cite the specific Turnitin statement that the AI writing indicator is not intended as the sole basis for adverse action
• Note the acknowledged elevated false positive rate below the 20% threshold, if your score falls there
• Distinguish between the document-level and sentence-level accuracy claims when the accusation rests on specific highlighted passages
• Ask what additional evidence, beyond the score, the institution is relying on
• Reference the institution's own procedural requirements for evidence of a violation

These points work best when paired with process evidence: drafts, version history, browser records, and notes that document how the paper was actually written. The vendor's acknowledgment establishes that the score alone is not enough. Your process evidence establishes what actually happened. For a full breakdown of the artifacts that matter and how to preserve them, see the guide on gathering evidence in the first 48 hours, and the procedural rights FAQ for what you can ask the institution to disclose before a hearing.

What the acknowledgments do not do

Turnitin's statements are useful but limited. They do not tell an instructor to ignore the score. They do not overrule campus policy. They do not, on their own, get an accusation dismissed. What they do is establish, from the source least sympathetic to your argument, that a detector reading is not proof. That establishes the ground you need before the rest of your defense (process evidence, policy citations, and, if relevant, research on detector bias) can do its work.

If you are drafting a written response, NotBot generates a personalized defense package that cites Turnitin's published limitations, references the specific detector used in your case, and mirrors your institution's procedural language. For post-finding cases, the appeal package reframes the same evidence around the grounds an appellate body will actually consider.