Turnitin's 4% Sentence-Level False Positive Figure: What It Meant in 2023

Turnitin's June 2023 explanation described around 4% of sentences already highlighted as AI-written as potentially human-written. That is a statistic about flagged sentences, not a calibrated probability for a particular sentence in your report today. It is also not the percentage of all human-written sentences that will be flagged.

Turnitin's Chief Product Officer Annie Chechitelli described it this way in June 2023:

"Our sentence-level false positive rate is around 4%. This means that there is a 4% likelihood that a specific sentence highlighted as AI-written might be human-written. The incidence for this is more common in documents that contain a mix of human- and AI-written content, particularly in the transitions between human- and AI-written content."

The condition is that the sentence has already been highlighted. A document-level false positive rate instead asks how often human-written documents trigger a detection rule. These use different denominators, so one percentage cannot be converted into the other.

The same June 2023 post also gives a document-level figure:

"Our document false positive rate - incorrectly identifying fully human-written text as AI-generated within a document- is less than 1% for documents with 20% or more AI writing."

The wording is easy to misread: the 20% refers to what the detector reports, not to a known amount of AI writing actually present. Turnitin's August 2024 whitepaper explains a document decision rule using more than 20% of sentence scores above a model threshold. It tests false positives on pre-2019 student papers described as entirely human-written. A false positive is one of those human-written papers crossing the detection threshold.

Fully human-written papers are precisely what a document-level false positive test needs. They are not excluded merely because the detection rule uses a 20% cutoff.

The 2023 post says wrongly highlighted sentences occur more often in mixed documents, particularly at transitions between human and AI writing. More often does not mean exclusively. It does not rule out wrongly highlighted sentences in entirely human-written work.

The under-1% target describes the frequency with which human-written documents trigger the stated detection rule under the vendor's testing conditions. It does not mean that a flagged paper has a less-than-1% chance of being human-written. That reverses the condition being measured.

Both figures can help explain the possibility of error. Neither identifies the author of your particular passage. Examine the flagged text and the writing process rather than using either rate as a verdict.

Turnitin gave a specific data point about where these wrongly flagged sentences show up:

"As explained in my earlier article, there is a correlation between these sentences and their proximity in the document to actual AI writing. 54% of the time, these sentences are located right next to actual AI writing."

"These sentences" refers to human-written sentences that were incorrectly highlighted. In the 2023 account, 54% were next to actual AI writing. This describes an observed concentration of errors, not a rule covering every false positive.

That observation cannot diagnose the cause of a flag in your paper. It neither proves nearby AI writing exists nor excludes an error when it does not. Use the highlighted passages to guide a review; proximity statistics cannot establish authorship.

The same post is explicit that a highlight is a starting point, not a finding — and it says so to instructors, which is who you want reading it:

"Consider highlighted sentences as areas of interest because they're predicted to be close to where AI writing is present. But a small percentage of times, the AI model could get it wrong. So, use the information to initiate a conversation, not to draw a conclusion."

It also rules out the thing students most often assume they are being measured against. There is no number you are supposed to hit:

"Remember that there is no 'right' or 'target' score with the AI writing indicator, just like with a Similarity Score."

None of the above tells you whether you are in trouble, because no published number does. What it does give you is a way to keep the conversation on the report itself rather than on a percentage.

• Ask which passages, and get the report rather than a screenshot. Two highlighted sentences and a document-level number are different objects. The highlights are the only part with a location you can point at.
• Ask which rate is being quoted and what it measures. The 20% cutoff concerns the detector's output. The document-level false positive rate is measured on human-written work; it is not the probability that your particular flagged paper is human-written.
• Take Turnitin's own instruction with you. It tells instructors to use a highlight to start a conversation, not to reach a conclusion. That sentence is addressed to the person holding your report, and it is quoted above word for word.
• Keep what shows how the paper was written — drafts, version history, notes, the reading you did. That is evidence about authorship. A percentage is not.