Vanderbilt 2023: why a university disabled Turnitin's AI detector, with the numbers it used

This is a statement record, like the Turnitin claims record in the studies index: it documents what an institution decided and the arithmetic behind the decision, with the vendor's numbers as the institution cited them. No test was run; Vanderbilt's post describes "several months of using and testing this tool, meeting with Turnitin and other AI leaders, and talking to other universities" without printing a result (Vanderbilt Brightspace, 16 August 2023, read 9 November 2026).

One detector, discussed rather than measured: the AI writing indicator Turnitin switched on for its customers on 4 April 2023 with, in Vanderbilt's words, "less than 24-hour advance notice, no option at the time to disable the feature, and, most importantly, no insight into how it works" (Vanderbilt Brightspace, 16 August 2023, read 9 November 2026).

Detector as described in the two Vanderbilt posts and the Turnitin post they cite, read 9 November 2026 (Vanderbilt Brightspace, 4 April 2023).

The only samples in this record are counts of papers, not scored documents (Vanderbilt Brightspace, 16 August 2023; Turnitin, 23 May 2023, read 9 November 2026).

Counts as stated by each party in the linked posts; neither count is a sample this record can inspect.

The metric Vanderbilt used is expected false positives: a stated document false positive rate multiplied by a year's submissions. Turnitin's post defines the rate it publishes as the share of human-written documents flagged as containing AI writing, "one out of every 100 human-written documents", for documents scored over 20% AI (Turnitin, 23 May 2023, read 9 November 2026).

The table sets Vanderbilt's figures beside the Turnitin statements it relied on (Vanderbilt Brightspace, 16 August 2023; Turnitin, 23 May 2023, read 9 November 2026).

Vendor values are Turnitin's statements; Vanderbilt's 1% and 750 are the university's reading of them. Vanderbilt did not publish any test result of its own.

Vanderbilt gave four reasons. The first is the arithmetic above, which the post pairs with press reports of false accusations "including multiple instances that involved Turnitin" and with the Stanford finding that detectors flag non-native English writers more often, the result this index carries as Liang 2023 (Vanderbilt Brightspace, 16 August 2023, read 9 November 2026). The second is opacity: "Turnitin gives no detailed information as to how it determines if a piece of writing is AI-generated or not." The third is the market: other vendors had pivoted or shut down, a reference to the OpenAI classifier withdrawn on 20 July 2023 that the index records as OpenAI classifier 2023. The fourth is privacy, the risk of "taking student data and entering it into a detector that is managed by a separate company". The post's conclusion is categorical: "we do not believe that AI detection software is an effective tool that should be used."

Turnitin's post of 23 May 2023 supplies the counter-reading without answering Vanderbilt, which it predates by twelve weeks. It concedes that "real-world use is yielding different results from our lab", that documents scored under 20% AI show "a higher incidence of false positives", and that false positives cluster in the first and last sentences of a document; it responds with an asterisk on scores under 20%, a 300-word minimum and a change to how opening and closing sentences are aggregated (Turnitin, 23 May 2023, read 9 November 2026). The post also says that "behind each false positive instance is a real student who may have put real effort into their original work", the same framing Vanderbilt used to reach the opposite decision. No Turnitin statement addressed to Vanderbilt's decision was found on turnitin.com on 9 November 2026.

Vanderbilt's April post had already set the frame: faculty had "no insight into how it works", false positives were "not just a possibility, but an inevitability", and the tool should be "a guideline, not a grading metric" (Vanderbilt Brightspace, 4 April 2023, read 9 November 2026). The August post replaced that guidance with a list of alternatives: early communication with students, disclosure and citation of AI use under APA and MLA guidance, in-class writing, and comparison with a student's earlier work.

The results as evidence question, added 9 November 2026, reads the 16 August 2023 and 4 April 2023 posts as two of the seven primary documents in the index on acting on a detector score.

Vanderbilt states none as limits, but the post is explicit that its decision rests on the vendor's own figure and on reported cases rather than on a Vanderbilt measurement, and that privacy and transparency weighed alongside accuracy (Vanderbilt Brightspace, 16 August 2023, read 9 November 2026). Turnitin's post states that its metrics "may change", names the 20% threshold, the 300-word minimum and the 4% sentence rate as the figures most likely to move, and says the 1% rate "is small" but cannot be reduced to zero (Turnitin, 23 May 2023, read 9 November 2026).

Evident from the sources: Vanderbilt rounds Turnitin's "less than 1%" to 1% and applies it to all 75,000 papers, although Turnitin's rate is stated for documents scored over 20% AI and says nothing about the share of a year's papers that would reach that score; the 75,000 figure is for 2022, before the detector existed; and the 800,000-paper validation has no published composition.

What this record does not say: 750 is an upper-bound estimate from a vendor's rounded claim and a prior year's submission count, not a count of Vanderbilt students who were wrongly flagged.

One card, as a dated institutional decision rather than a measured rate: the Turnitin card notes that Vanderbilt disabled the indicator on 16 August 2023 and why (Vanderbilt Brightspace, 16 August 2023, read 9 November 2026).

The record carries no false positive, false negative, accuracy or robustness figure, so it enters no numeric ranking. It belongs beside the claim vs evidence gap ranking, which measures Turnitin's published claim against the tests in this index, and it illustrates how the vendor's figure reads at the scale of one institution. How this record is sourced explains why vendor claims and institutional statements are dated and kept apart from study results, and the home page shows which AI detectors exist and who has tested them.

Related records: the Turnitin claims record holds the full set of Turnitin's dated statements; Weber-Wulff 2023, Walters 2023 and Perkins 2024 are the tests of Turnitin in this index; Liang 2023 is the non-native writer study Vanderbilt cited.