Vanderbilt, BU, Arizona, Waterloo and Pittsburgh's Teaching Center Each Published a Page Switching Turnitin's AI Detector Off

Five verified pages say it on the institution's own domain; four carry a date. Vanderbilt's Brightspace group posted the decision on 16 August 2023 ¹. Boston University's IS&T Turnitin FAQ records 3 July 2023 ². Arizona's Dean of Students page, updated 27 June 2023, says the feature was temporarily disabled for all courses until the campus learned more from the vendor, and records no later change ¹¹. Waterloo announced on 5 May 2025 that the tool would be unavailable as of September 2025 ¹⁰.

The fifth page carries no date and goes a step further than the other four: the University of Pittsburgh's University Center for Teaching and Learning states that the tool inside Turnitin is disabled, and that it endorses no AI detection tool at all ³.

Two more institutions belong on a different row of the same ledger.

Yale's Poorvu Center for Teaching and Learning writes on its faculty guidance page that it does not currently support the use of AI detection tools ⁴. MIT Sloan's teaching technology group publishes a page arguing that detectors do not work, and warns that high error rates can lead to false accusations ⁵.

Neither page names a switch. A list filing both under one heading is already guessing.

Trade coverage carries the rest. Inside Higher Ed reported in August 2026 that at least a dozen universities, among them Northwestern, Georgetown and New York University, had disabled Turnitin's AI detection software ⁶. That's journalism rather than an institutional statement, and a briefing labels it that way. A search turns up more, and each new row needs the same check.

Four objections recur on Vanderbilt's and Pittsburgh's pages. The first is plain arithmetic: a false positive rate multiplied by real submission volume. The second is the absence of any published account of how a verdict is reached. The third is a rate the vendor later conceded ran higher than its launch figure, with no replacement published since ³. Equity is the fourth.

Boston University and Arizona add reasons of their own; Waterloo's announcement gives none ¹⁰.

Vanderbilt did the multiplication in public, applying Turnitin's claimed 1 percent false positive rate to the 75,000 papers it submitted in 2022 and writing that around 750 could have been incorrectly labeled ¹. The sum rests on two assumptions: that every submission was human-written, and that the vendor's rate holds on that campus's papers. It runs just as well on a smaller campus, which is why the objection travels, and the method for running it on any submission count is in the false positive arithmetic.

The opacity objection is shorter and much harder to answer. Vanderbilt's page states that "Turnitin gives no detailed information as to how it determines if a piece of writing is AI-generated or not" ¹.

A student accused on the strength of a reading the vendor does not explain has little in the reading itself to rebut.

On equity, Vanderbilt cited research finding that detectors are more likely to label text written by non-native English speakers as AI-written ¹.

Boston University compresses its reasons into one line: the panel went off over a range of concerns, including false positives and a lack of supporting evidence for its efficacy ². Arizona's page is as brief: the feature generated inconsistent results, and the campus did not yet have strong guidance for instructors who wanted to use it ¹¹. Waterloo's announcement records that the decision followed consultation with its associate deans, its undergraduate operations committee and its educational technology steering committee ¹⁰.

Three of the four can be put to any vendor as questions. The conceded rate is Turnitin's own history.

Read Pittsburgh's page before a renewal conversation. It records that Turnitin acknowledged in June 2023 that its AI detection carried a higher false positive rate than the company originally asserted, and that no replacement estimate has been disclosed since ³. The page also notes that accuracy claims of 98 to 99 percent are common among detector vendors, while third-party evaluation finds high false positive rates ³. Many vendors state no rate at all ³.

Its conclusion is the sentence a hearing officer needs: detectors are not reliable enough to act as proof ³. That is the standard argued in why a score alone cannot support a finding.

Some institutions use a detector and write the conditions down beside it, as Portland Community College does on the page that currently limits Pangram to faculty, staff and the students helping evaluate it ⁷. The same page says the tool is not perfect, and that it is being evaluated to see whether it can support instructors in the short term while disciplines adapt to new approaches to assessment ⁷.

Turnitin publishes conditions of its own, and they are easy to miss inside a service a campus has run for a decade. In May 2023 the company added an asterisk to any document score under 20 percent, calling the score less reliable in that range, and raised the minimum length for a reading from 150 words to 300 ⁸. Since 8 July 2024 it attributes no score and no highlights under 20 percent at all ¹². Turnitin also states that it makes no determination of misconduct, and that the decision rests with the educator and institutional policy ⁹.

Put together, those pages show three conditions written down: a stated short-term evaluation ⁷, a floor under document length ⁸, and a vendor statement that the reading decides nothing by itself ⁹.

Guidance can also come from a single school: Inside Higher Ed quotes the Kelley School of Business at Indiana University telling faculty in its AI playbook that services claiming to detect AI use are highly unreliable ⁶.

Each objection converts into a question with a checkable answer: four questions from the campus pages, one per objection, and a fifth from Turnitin's own published conditions rather than from any campus page. Every answer is a document a vendor either has or does not have.

• A false positive rate times real volume produces wrongly labeled papers ¹. Ask for the rate, the population it was measured on, and the same sum run against the human-written share of this institution's submissions, with the assumption stated.
• No published account of how the verdict is reached ¹. Ask for a method document a student and a hearing panel can both read.
• More flags on writing by non-native English speakers ¹. Ask for a separate result on that population.
• A conceded higher rate with no replacement number ³. Ask for a dated figure and a commitment in writing to publish the next one when it exists.
• No score attributed under 20 percent since July 2024 ¹², and nothing under 300 words ⁸. Ask for a stated minimum length and stated behavior on short or edited text.

Two of those rows are procurement questions rather than research questions, and they belong in the packet before a demo is booked.

The list to send a vendor is in the questions to ask a detection vendor, and the contract language that holds those answers in place once the sales call is over is in the detector RFP requirements.

The last requirement is the one nobody writes down in time. A reading is evidence inside a process, and the process has to exist on paper before the tool arrives, not after the first disputed case reaches a dean's inbox on a Tuesday afternoon.

A switch can happen at four levels, and no register at the vendor, an accreditor or a higher education association was found while checking these pages to settle which of them a list counts. Campus-wide, in the learning management system. One school inside a university. A teaching center's guidance. Or a single department's rule. A list counting all four reaches a large number, and a list counting only the first reaches a small one. Both can be honest.

Neither is checkable without the pages behind the rows, and the check is short.

Open the institution's own domain and find the page the list points at. Read what it says. A page with a date and the word disabled or discontinued records a switch; Waterloo's does ¹⁰, and Arizona's records a temporary one ¹¹. A page saying a center does not support detection tools, as Yale's does ⁴, is guidance that leaves the setting untouched, and a page arguing that detectors do not work, as MIT Sloan's does ⁵, recommends a practice to instructors without changing a setting.

Three different facts.

Rows naming an institution without linking that institution's own page are the ones to cut first. Inside Higher Ed's figure ⁶ is a sourced floor: at least a dozen, with the reporting attached rather than a bare number in a table.

For the decision sitting underneath the count, the reasons these universities published are worth more than the tally, and the argument on both sides is set out in whether schools should use AI detectors at all.