The AI detector false-accusation ledger: every documented case, detector and outcome

Every false accusation by an AI writing detector that survived in dated public reporting from 2023, the tools' first mainstream year, ended with the accused students cleared. Not one of those reversals came from the software: students beat the flags by producing drafts and file histories, or a newsroom forced a review 1 2. Four documented case clusters at three universities, one of them an entire class. Each case made the news once and then scattered, and the viral roundup that proves the appetite, BuzzFeed's listicle of accused students, is screenshot posts naming no detectors and no outcomes, several of them predating ChatGPT 3. This ledger is the assembly: a dated, source-linked list of every case joined to public reporting, plus the institutions that walked the tools back. It is maintained as new cases land.

The scale behind the flags: Turnitin's software serves more than 16,000 academic institutions 2, and nearly 950,000 international students study at US schools that routinely warn them a misconduct finding can bring suspension or expulsion and undermine their visa status 2. Turnitin put its false-positive rate under 1 percent 4 5. The Washington Post's own test put the tool near 50 percent on a smaller sample 6. The Patterns study averaged 61.2 percent false positives across seven detectors, none of them Turnitin's, on essays by non-native English writers 7 2.

The cases on the record

A row enters this ledger only when dated public reporting or an institutional statement names the institution, and it counts as a false accusation only where that reporting documents the evidence that cleared the student or an exoneration by the institution. Detectors appear only where the reporting names them, and outcomes are stated exactly as reported, never inferred. Each row carries the date, the institution, the detector, the evidence that cleared the student, and the outcome as reported.

• Spring 2023, University of California, Davis (William Quarterman, a senior history major): a professor failed him after running his essay through GPTZero, and the case went to the university's Office of Student Support and Judicial Affairs 1. He was cleared following media coverage of his case and a thorough investigation 8 6.
• Spring 2023, University of California, Davis (Louise Stivers, a 21-year-old political science student): Turnitin's AI indicator, which administrators later said was in beta, flagged her paper and pushed her into an integrity review during her final semester 8. She was cleared by the administration, but the investigation stays on her record, she has to self-report it to law schools and state bar associations, and the school never apologized 1.
• May 2023, Texas A&M University-Commerce: a professor pasted his class's work into ChatGPT, accepted its claim that it had written the papers, and threatened failing grades with diplomas in jeopardy. No student was prevented from graduating, and all but one, who admitted using the software, were exonerated 6.
• Spring 2023, Johns Hopkins University: in a communications course taught by Taylor Hahn, Turnitin labeled more than 90 percent of one student paper as AI-generated and the majority of a second paper; the students cleared themselves with drafts and PDFs, and Hahn documented a semester-long pattern of false positives that hit international students hardest 2.

By tool, the record names Turnitin in two clusters, GPTZero in one, and ChatGPT itself in one, used as a detector despite not being one.

A claimed 1 percent, and what measurement found

Turnitin shipped the detector in April 2023 with less than 24 hours' notice to customers, no option to disable it at launch, and the claimed 1 percent false-positive rate 5. The measured record runs the other way. The Washington Post's own test, on a smaller sample, put Turnitin itself near 50 percent 6. Across seven detectors, none of them Turnitin's, essays by non-native English writers drew an average false-positive rate of 61.2 percent, with about a fifth of papers unanimously misclassified, while the detectors almost never flagged native speakers' writing 7 2. Common Sense Media's 2024 testing found false-positive rates of 20 percent on writing by Black students, 10 percent for Latino students and 7 percent for White students 6.

Turnitin's chief product officer, Annie Chechitelli, said the tool was trained on writing by English speakers in the US and abroad as well as multilingual students, so it should not show the bias the Patterns study identified, and the company's own research so far pointed to no such bias. Patterns co-author Weixin Liang replied that the tool still needs third-party evaluation before anyone can trust real-world performance 2.

Vanderbilt did the arithmetic no vendor prints: at 1 percent false positives against the 75,000 papers the university submitted to Turnitin in 2022, around 750 student papers could have been wrongly labeled in one year, which is why it disabled the tool 5.

The walk-backs, and how slowly the lesson traveled

Six public retreats mark the year. Two were vendors shutting their own tools: OpenAI discontinued its AI classifier at the end of July 2023 over its low rate of accuracy, and Quill.org and CommonLit shut their AI Writing Check, saying generative AI tools are too sophisticated for detection 2. Four were universities: Cambridge University and other Russell Group members opted out of Turnitin's AI detection in April 2023 as unreliable 6; the University of Pittsburgh's teaching center told faculty at the end of June that it did not support any AI detector after testing a handful, warning that false positives carry risks including legal sanctions 2; Vanderbilt disabled Turnitin's detector for the foreseeable future in August 5; and the University of Texas at Austin opted out in September 2023, five months after Cambridge 4 6.

Then there is the uptake gap. When Turnitin shipped an institutional switch to hide the AI indicator in June 2023, only two percent of its customer institutions had flipped it by the end of July, according to the company 2. On that figure against the more than 16,000 institutions, the arithmetic gives at least 320 switches flipped, leaving the other ninety-eight percent with the indicator on 2. The vendor's own number matches the complement: Turnitin's chief product officer, Annie Chechitelli, told The Register in September 2023 that 98 percent of its customers are using the feature 4.

The ledger is maintained: rows are added as dated reporting lands. The pattern the first year already holds: the error rate is a vendor's claim, the clearing evidence is the student's to produce, and the side-by-side record is the only part anyone can audit.