Which AI Detector Teachers Use: Turnitin by Contract, GPTZero by Signup

Turnitin and GPTZero. What separates them is how they arrive. At Pittsburgh, Turnitin reached faculty through the license the university already held, which is why its teaching center could disable the AI writing report for the whole campus at once ⁶. GPTZero reaches the same teacher through a signup page, free, in a browser tab, with no purchase order anywhere near it ⁵.

One is a procurement decision, the other a browser tab.

The vendors' own counts tell the same story from the other end. Turnitin reported 38.5 million submissions processed for AI writing by 14 May 2023, of which 9.6 percent came back above 20 percent AI writing ². That is contract volume, not signup volume.

GPTZero counts the other kind. Its own site claims 1 million educators and use in more than 3,500 colleges ⁴. Its FAQ describes a free account that raises a single scan to 150,000 characters, or a batch of 250 files ⁵. The free path goes nowhere near a purchasing department, and the same FAQ says GPTZero also works with over 100 organizations in education, so an institutional route exists beside the signup ⁵.

Sixty-eight percent of US public middle and high school teachers reported using an AI content detection tool regularly in the 2023-24 school year, up 30 percentage points on the year before ¹. The figure comes from a nationally representative Center for Democracy & Technology survey of 460 teachers, fielded in late 2023. Thirty points in twelve months. Adoption moved faster than any evaluation of the tools could.

In the same survey only 25 percent of teachers said they were very effective at telling whether an assignment was written by a student or by generative AI ¹. The report offers that lack of confidence as one possible explanation for the adoption ¹.

Students describe the practice from the other side. Among teens whose teachers had discussed rules for AI, 29 percent said at least one teacher had run their writing through detection software ⁸. That reading comes from an Ipsos survey of 1,045 young people aged 13 to 18, fielded in spring 2024 for Common Sense Media ⁸.

Neither number names a product. Both describe a habit that spread before anyone had settled which tool was worth having.

What the vendors tell a teacher to do with it is start a conversation; what the survey records is how often that conversation ends in trouble. Turnitin's guidance on its sentence-level error rate ends on a plain line: "use the information to initiate a conversation, not to draw a conclusion" ³. GPTZero's FAQ is blunter, saying "these results should not be used to punish students" ⁵.

What the survey records is rougher than the guidance. Teachers who used a detector regularly were much likelier to report students getting in trouble over AI, 72 percent against 48 percent of the teachers who didn't use one ¹. Forty percent of the same teachers agreed that a student got in trouble for how they reacted when an adult at the school raised alleged misuse of generative AI ¹. That's a conversation going wrong, not a score going wrong.

A reading can start that conversation. It can't settle it, because a detector scores text: it holds no record of who typed the document, and it can't see the draft history a student who did the work can produce on request.

Because the AI detector inside Turnitin is an institutional switch. The University of Pittsburgh's teaching center threw it, disabling the tool for the whole campus, and then went further, stating that it endorses no AI detection tool at all ⁶. A teacher at a school that made that call has no sanctioned detector to choose, whatever a vendor page offers.

Pittsburgh's stated reason is the one a procurement file should carry. Headline claims of 98 to 99 percent accuracy are common among detector vendors, third-party evaluation finds high false positive rates, and many of those companies decline to publish a false positive rate at all ⁶.

Pittsburgh is not the only campus to have decided this. A school weighing the same question can start from the universities that turned off Turnitin's AI detector and read what each one published.

The vendor question and the policy question are not the same question.

A teacher who wants to know what is on offer should ask the teaching center before the sales page, with two questions: whether the AI writing report is enabled in the school's Turnitin license, and whether a personal account with an outside detector is allowed to receive student work at all.

Read the conditions attached to each headline figure, because they change what it means. Turnitin puts its document-level false positive rate below 1 percent, a rate it states for documents with 20 percent or more AI writing ³. Its sentence-level rate is around 4 percent, so roughly one highlighted sentence in twenty-five was written by a person ³.

Length matters as much. Turnitin raised its minimum from 150 words to 300 words in May 2023 ², and since 8 July 2024 it attributes no score at all under 20 percent, printing an asterisk instead ⁹.

GPTZero's published 99.1 percent figure for human articles applies inside its high confidence band, alongside 98.4 percent on AI articles ⁵. Both companies reward reading in full. The working is set out in what a Turnitin AI detection accuracy claim actually measures and in how accurate GPTZero is on its own published testing.

A confidence band is not a rate.

Outside tests have not reproduced the vendor rates on every population. Seven detectors were run over 91 TOEFL essays written by non-native English speakers and wrongly flagged more than half of them, an average false positive rate of 61.3 percent ⁷. The authors strongly caution against using such detectors in evaluative or educational settings, particularly when assessing the work of non-native English speakers ⁷.

Teens' own reports add a second population, this one by survey rather than by test. Black teens report being wrongly flagged at about twice the rate of their White and Latino peers ⁸.

The cost of a wrong flag lands unevenly.

One more number, for scale, with its conditions attached like the two above. Human's own in-house measurement, on its own held-out sets and not an independent audit, is 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%, measured September 2026. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. The figure describes documents, never a student.