Which AI Detectors Colleges License, and Which They Switched Off

Turnitin's AI writing indicator, which arrived in 2023 inside the Similarity Report institutions already licensed for text matching ¹. Turnitin's product page now describes it as the Originality add-on: once licensed and enabled, the indicator is linked to the Similarity Report ¹⁵. It reaches Canvas through the Turnitin Advantage integration ². No public registry records which college licenses which detector, which is why the question has no clean answer.

Scale is the one public proxy for share on record here, and it comes from the vendor. Turnitin reported 38.5 million submissions processed for AI writing by 14 May 2023, with 9.6% of those coming back above 20% AI writing and 3.5% between 80 and 100% ¹. That is a volume figure, not a customer list. It says nothing about which institutions the submissions came from, or how many of those campuses had the indicator visible to faculty at the time.

The 2023 arrival matters more than the brand for one practical reason: at that point nothing had to be procured. A campus already licensing Turnitin for text matching met the AI writing indicator inside the same Similarity Report ¹, which is a different event from choosing a detector. The one institutional act available that year was the opposite one, turning it off, which is why so much of the public record from 2023 is a record of switch-offs. Since then the indicator has become a licensed add-on ¹⁵, so a campus that holds Similarity alone has a line item to decide on.

Where a campus runs Canvas, that is where faculty meet the thing, and Canvas is not doing the detecting. Instructure describes the Turnitin Advantage Integration, built on the LTI 1.3 standard, as what makes the full suite of Turnitin tools available inside Canvas ². A campus that holds Turnitin and asks which detector it uses is asking which add-on is licensed and who switched the integration on.

Pangram and GPTZero, each as its own contract rather than a feature of something already installed. Pangram lists an Educational Institution tier for schools, districts and universities as Contact Us, offering unlimited AI checks and plagiarism checks integrated directly into the LMS ⁸. Its university page states FERPA and SOC 2 Type 2 compliance and no training on student papers ⁹. GPTZero publishes accuracy bands ¹¹ rather than a school price. Its pricing page lists enterprise and team plans as Contact Sales ¹⁶.

A separate purchase buys something else too: a vendor whose published evidence a committee can read before it signs. Set the four pages side by side and each vendor commits to something different: a rate, a set, a compliance list, or all three.

Adoption by a conference or a university is a fact about uptake, not a published rate, and the one institutional page on Pangram cited here is worth reading ahead of the vendor's own. Portland Community College tells faculty the detector is not perfect and is being evaluated to see whether it can support instructors in the short term while disciplines adapt to new approaches to assessment ¹³.

Price is the part no peer call produces. Pangram's institutional tier is a quote rather than a number ⁸, so no vendor page supports a comparison of license costs across campuses. What a school pays for Pangram is the same shape of question, with the same shape of answer.

Vanderbilt, Boston University and the University of Pittsburgh's Teaching Center, each on its own public page; Vanderbilt and Boston University date the decision. Boston University disabled the Turnitin AI score panel on 3 July 2023, naming false positives and no supporting evidence for the tool's efficacy ⁵. Vanderbilt announced on 16 August 2023 that it was disabling the detector for the foreseeable future ⁴. Pittsburgh's Teaching Center disabled it inside Turnitin and endorses no AI detection tool at all ⁶.

One of those pages also does the sum a board memo can leave out. Vanderbilt submitted 75,000 papers to Turnitin in 2022, and calculated that around 750 of them could have been incorrectly labeled as containing AI writing had the tool been running ⁴. No peer list supplies that number. It is a function of local volume, not of vendor choice.

Pittsburgh's page carries a procurement detail worth lifting out: it records that Turnitin acknowledged a higher false positive rate in June 2023 and, in the Teaching Center's reading, has not disclosed a replacement estimate ⁶. Turnitin's own June 2023 post restates the document rate with a condition and adds a sentence-level rate ³; a committee should line up the conditioned figures, not the original headline.

Press coverage puts the set wider than the pages documenting it. Inside Higher Ed reports at least a dozen universities, including Northwestern, Georgetown and New York University, having disabled Turnitin's AI detection software ⁷. That is reporting rather than an institutional statement, and a memo citing it needs to say so.

The universities that turned off Turnitin's AI detector are the closest thing to a census this question has.

Three settings decide what a campus is actually running. Turnitin evaluates a document only at 300 words or more, a floor raised from 150 in May 2023 ¹. Readings under 20% get an asterisk and no score at all since 8 July 2024 ¹⁷, in a band Turnitin had already called less reliable ¹. And its document-level false positive rate, published as under 1%, applies to documents the model reads as 20% or more AI writing ³.

What an instructor is shown, and may put to a student, is the highlighted sentence, and Turnitin puts its sentence-level false positive rate at around 4%, roughly one highlighted sentence in twenty-five being human-written ³. On a long paper carrying forty highlighted sentences, that is not a rounding error. It is a figure an appeal can put to the instructor.

The vendor's own instruction sits in the same post: use the information to initiate a conversation, not to draw a conclusion ³. A campus that leaves the indicator on and lets a percentage carry a misconduct finding has adopted a reading Turnitin does not endorse.

Equity has a published figure here too. Turnitin reports that on documents meeting the 300-word requirement, writers learning English received a 0.014 false positive rate against 0.013 for native English writers ¹⁴. That comparison is the vendor's, on the vendor's set, and it belongs in the memo beside the false positive rate rather than after the decision.

Reading a Turnitin AI score starts with knowing which of these settings produced it.

The measured set behind every figure, and the flag count at local volume. Nature's 2026 news feature on detection tools found that only the accuracy rates firms announce from internal testing are current, and that those are not externally verified ¹⁰. A peer list answers neither question. It reports what somebody else signed, and that is the least transferable fact in the file.

Three questions do more work than a vendor name. What set produced each accuracy figure, and who ran the measurement. What the flag count comes to at the campus's own submission volume, the way Vanderbilt worked it out ⁴, with two assumptions said out loud: the rate applies to the human-written share of the pile, not to every submission, and a rate measured on the vendor's set is being assumed to hold on this campus's papers. And what happens to a flagged student between the report and the hearing.

Questions to ask an AI detection vendor covers the first; the false positive arithmetic covers the second.

One vendor already publishes its figures in the shape a committee can line up against Turnitin's page. Human states its operating points as thresholds on machine content: accusation-safe flags a paper above 15% machine content, standard above 6% (the default), sensitive above 2%. As of September 2026 it records 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. Both figures are Human's own in-house measurements on its own held-out sets, not an independent audit, and they belong in the memo under the same heading as any other vendor's internal testing.

A cycle that ends with no purchase and a measured local flag rate has still done its work. What it leaves behind is evidence about the campus's own students, the only population the policy governs, and a clause set the next committee can reuse.