Turnitin Publishes No Acceptable AI Percentage, So a Policy Should Not Name One

No. Turnitin's own guidance puts it plainly: there is no "right" or "target" score with the AI writing indicator, any more than there is one with a Similarity Score ³. The company also states that it does not make a determination of misconduct, leaving that decision to educators under their own academic and institutional policies ¹. The 20 percent figure that circulates in policy drafts comes from two places in Turnitin's product writing. Neither place is a standard of conduct.

One is a reporting rule. In May 2023 Turnitin added an asterisk to every AI percentage under 20 in the Similarity Report, in its own words "to call attention to the fact that the score is less reliable" ². Since 8 July 2024 the report goes further. Between 1 and 19 percent it attributes no score and no highlights, showing the asterisk in place of a number ¹⁰. The other is a condition attached to an error rate. Turnitin's document-level false positive rate of under 1 percent is published only for documents its model reads as at or above 20 percent AI writing, in its 2023 post ³; the current FAQ words the same condition as over 20 percent ¹¹. No rate is published for the band below that line.

So 20 is where a vendor's confidence begins, not where a student's honesty ends.

Reading either fact as a threshold inverts it. An asterisk warns that the reading beneath it is weaker. A policy that treats everything above the asterisk as actionable has adopted a reliability marker as a line of conduct.

What the percentage actually counts, sentence by sentence, is set out at what a Turnitin AI score means.

Because a number printed beside a verdict acquires authority the vendor never gave it. Turnitin reported that of 38.5 million submissions processed by 14 May 2023, 9.6 percent came back above 20 percent AI writing ². Carrying that rate over to an institution's own submissions is an assumption about a population Turnitin did not measure. Made anyway, it produces flagged papers in volume every term, and a written threshold makes each of them a case to open.

Vanderbilt University submitted 75,000 papers to Turnitin in 2022 and worked out that around 750 of them could have been incorrectly labeled had the detector been running, then switched it off in August 2023 ⁴.

About 750 papers that could have been wrongly labeled is not a rounding error in a disciplinary caseload.

The rate itself is still published, and it is still conditional. Turnitin's AI writing detection FAQ, updated in September 2026, states the aim of keeping the false positive rate under 1 percent for documents with over 20 percent of AI writing. Behind that aim sits a check: before every update or new model release, Turnitin tests over 700,000 additional academic papers written before the release of ChatGPT ¹¹. Turnitin's chief product officer wrote in May 2023 that below 20 percent there is a higher incidence of false positives ², and the University of Pittsburgh's teaching center records that Turnitin has not disclosed a new false positive rate estimate ⁵. Pittsburgh's own response was to disable the AI detection tool inside Turnitin and to endorse no detection tool at all ⁵.

A clause cannot borrow a vendor's self-reported rate as if it were an audit.

The sentence-level rate is the one that reaches an individual paragraph. Turnitin puts that false positive rate at around 4 percent, so roughly one highlighted sentence in 25 may be human-written ³. The document percentage is assembled out of exactly those highlighted sentences, so its error is not independent of theirs.

The same arithmetic at institutional scale is worked through at what a false positive rate does across a term of submissions.

It supplies the finding before the evidence arrives. A clause that makes any reading above 20 percent unauthorized AI use hands a panel a fact it can verify in seconds. It also hands the panel a conclusion nobody has to reach independently, so the hearing becomes a reading of the report rather than an examination of how the paper was written.

That is the opposite of what the vendor asks for. Turnitin tells instructors to use the information to initiate a conversation rather than to draw a conclusion ³, and the University of Pittsburgh states plainly that AI detectors are not reliable enough to act as proof ⁵.

A threshold clause makes the number sufficient, which is the one property it should never be given.

Appeal rights then have little to grip. The University of Texas at Austin limits a student conduct appeal to three grounds: significant procedural error, new information that was not reasonably foreseeable, and a sanction significantly disproportionate to the violation ⁶. Against grounds written that way, a correctly computed percentage is hard to contest. The arithmetic is not a procedural error, and the reading itself is not new information. What a student can still bring is evidence from outside the report: a draft history, a source check. A threshold clause discourages anyone from gathering either one.

The clause that looked like clarity has quietly narrowed the route out of it.

Process language does the work arithmetic cannot. CUNY's academic integrity policy asks a faculty member who suspects a violation to review the facts and circumstances with the student first, whenever that is feasible ⁷. That sentence names an act. It can be complied with or not, and it survives a change of vendor.

The longer case for keeping the score out of the finding is at why a detector score alone cannot support a finding.

Put the number in the workflow and keep it out of the rule. A detector reading is a defensible instruction to look again, and a policy can say exactly that. At or above a stated reading on a named tool, the instructor reviews the highlighted sentences and checks the draft history and the cited sources. A conversation happens before anything is filed.

The separation matches what Pangram publishes. Pangram Labs states that its document verdict is arithmetic over characters: Human at 90 percent or more human characters, AI at 80 percent or more AI characters, and Mixed for everything between ⁸. Those are decoding rules for a label, published by the company that wrote them, and nothing on that page claims they describe a student's conduct. A number that decides what a report prints is not a number that decides what a student did.

Five clauses do the work a percentage is asked to do:

• The reading that triggers a second look, written as a setting on a named tool rather than as a finding.
• What the reviewer gathers before contacting a student: draft history, cited sources checked for existence, earlier writing from the same course.
• The length below which the institution acts on no reading at all, since Turnitin gives no score under 300 words ².
• Who may see a reading, where it is stored, and how long it stays attached to a student record.
• The grounds on which a student contests the review, written so that a correctly computed number is not by itself an answer.

None of those five needs a percentage in the integrity clause.

Retention is the clause nobody drafts. A reading that stays attached to a record outlives the conversation that produced it. A policy that never says how long it is kept has handed that decision to the vendor's retention terms and to whatever records schedule applies.

A policy can also record what the institution will not do, which is where much of the published professional guidance has landed. The MLA and CCCC joint task force made it a named principle in April 2024: tools for detection and authorship verification should be used with caution and discernment or not at all ⁹. Vanderbilt went further and stated that it does not believe AI detection software is an effective tool that should be used ⁴.

Start from what a detector publishes about itself, and stop where that publication stops. A clause can honestly say that a reading at a named setting opens a review. That is a fact about a configured tool. What a clause cannot honestly say is that a reading establishes AI use, because neither Turnitin nor Pangram publishes a figure that would carry the sentence ¹⁸.

Choosing the setting is itself a decision a policy can document. A drafter looking for a named setting will find three at Human, published as Human's own in-house measurement as of September 2026: accusation-safe flags a paper above 15% machine content, standard above 6% (the default), sensitive above 2%. An institution wanting the fewest flags on human writing names the first one and records why.

Naming the setting is what makes a threshold auditable.

The false-flag figure belongs in the same paragraph. As of September 2026, Human measures 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. That figure is an in-house measurement on Human's own held-out sets rather than an independent audit, and it describes text rather than a student.

Human's documentation is equally explicit about the limit: no setting turns an AI verdict into a Human one, and an AI verdict needs at least 80% of the document to read as machine-written.

The clause that replaces the percentage is the shorter of the two. It names a tool, the setting that tool runs at, the review step a reading opens, and the route by which a student contests it. About what the number proves it says nothing at all.

That route has a design of its own, worked through at how an AI detection appeals process is designed.