In late January 2026, a New York state judge annulled an academic integrity finding against Adelphi University student Orion Newby, whose World Civilizations paper had scored 100 percent "AI-generated" on Turnitin. The court called the university's decision "without valid basis and devoid of reason" and ordered the violation expunged from his record. Inside Higher Ed reported it as the first time a court forced a university to reverse an AI plagiarism finding.
The ruling is the clearest marker yet of a shift that started in 2023 and finished in the research literature: AI text detectors were sold as an enforcement tool, and the evidence now says they cannot bear that weight. Schools are switching them off, courts are rejecting the scores as sole evidence, and the most rigorous independent testing to date puts commercial detection in a bleak place.
The numbers independent studies actually get
The pattern began with the Stanford study everyone still cites. Liang et al., published in the journal Patterns in 2023, ran 91 essays by non-native English speakers through seven GPT detectors. The tools flagged an average of 61.3 percent of that human writing as AI-generated. Ninety-seven percent of the essays were flagged by at least one detector. None of the essays, all of them real student writing, was AI. The same detectors produced almost no false accusations against essays by U.S. eighth-graders. The authors also showed the tools cut the other way: light prompt-based rewrites pushed genuinely AI-generated text under detection thresholds. Over-flagging the innocent and under-flagging the guilty, with the same software, in the same paper.

A larger academic-integrity team went broader. Weber-Wulff et al., in the International Journal for Educational Integrity, tested 14 tools including Turnitin and PlagiarismCheck. Every tool scored below 80 percent accuracy, only five cleared 70 percent, and the authors concluded the available detection tools are "neither accurate nor reliable," with a systematic bias toward classifying AI text as human.
The RAID benchmark (Dugan et al., arXiv 2405.07940) ran 12 detectors against six million generations across 11 models, 8 domains and 11 adversarial attacks. Detectors were "easily fooled" by adversarial rewriting, sampling changes, repetition penalties and unseen models. Vendor claims of 99 percent accuracy did not survive contact with out-of-domain text.
Then, in May 2026, University of Florida researchers presented "AI Wrote My Paper and All I Got Was This False Negative" at the IEEE Symposium on Security and Privacy. Testing commercial detectors on text that included security-conference papers, they measured false positive rates between 0.05 percent and 68.6 percent across tools and conditions, and false negative rates between 0.3 percent and 99.6 percent. One simple tweak to the generating model's prompt rendered the detectors, in the words of the research group's write-up, basically useless. Co-author Patrick Traynor put the conclusion plainly: "We really can't use them to adjudicate these decisions. People's careers are on the line here."
Why the same metric convicts both groups
Most detectors score text on perplexity, how predictable the word choices are. ChatGPT-style models produce smooth, low-perplexity prose. So, the Stanford team found, do many writers whose first language is not English, whose lexical diversity and syntactic complexity measure lower by design, not deception. The signal the detectors treat as evidence of machines is partly a measure of how sophisticated the writing looks to an English-language model.
That is not a tuning error. It is the detection rule itself. Which means better training data does not obviously fix it, and the tools' failures land unevenly across populations. The Markup reported in 2023 that a Johns Hopkins instructor watched international students get flagged repeatedly while native speakers passed; Turnitin's chief product officer said the company saw no bias in its own research and pledged to publish it. The bias finding has aged well. A 2026 analysis compiles peer-reviewed rates of 5 to 20 percent on native English and up to 61 percent on ESL writing.
The arithmetic schools cannot escape
False positive rate is per document. Enrollment is per semester. Vanderbilt ran the numbers in August 2023 before disabling Turnitin's AI detector: at 75,000 student submissions a year, even the 1 percent false positive rate Turnitin itself advertised would wrongly flag roughly 750 papers annually. With a 61 percent rate on non-native writing, the arithmetic gets grim for any university with a large international student body.
Institutions have been voting with their contracts. Vanderbilt and Georgetown switched the feature off in 2023. The University of Waterloo stopped using it after internal testing flagged human-written text as 100 percent AI. Curtin University in Australia cited fairness to international students. UC Berkeley opted out in 2025 after its own evaluation, and Programs.com's tracker now lists dozens of U.S. colleges that have disabled Turnitin's detector, most citing reliability.
The legal system is converging on the same position.

Newby v. Adelphi did not rule Turnitin inaccurate; it ruled that a 100 percent score alone, against a student with documented tutor edits and two other detectors calling the essay human, was arbitrary. A tracker of the cases shows the next wave: a federal suit in Massachusetts over a teacher's false accusation, a student loss on appeal in Minnesota after a GPTZero flag, and a May 2026 California K-12 case that imports the discrimination framing directly, alleging a boy's essay was re-run through Turnitin without consent after he rewrote it by hand.
What detectors can still do
Nobody credible is arguing that detectors see nothing. They have real signal on raw, untouched model output, which is exactly the least adversarial case. Where they still earn a place is as a screening hint, not evidence: a reason to look at version history, ask for a draft, or talk to the writer. Turnitin has effectively said this itself, stating that its tool should not be the sole basis for adverse action against students. GPTZero's own documentation carries similar caveats, which is a strange posture for a product sold to schools that act on its numbers.
The harder problem for the adjudication crowd is that the frontier models have moved. Output is now more varied, more editable, and routinely passed through rewriting layers, whether for polish or on purpose. The Florida S&P 2026 paper is the academic version of a market fact: the more capable the generators get, the less their text looks like a statistical tell.
Watermarking would fix some of this in principle, but the industry never shipped consistent watermarks, and removal attacks remain an active research area. So the 2026 equilibrium looks like this: detectors will keep scoring essays, keep being wrong at a rate any courtroom would notice, and keep surviving as a convenience for institutions that would rather not build the process-based integrity systems the evidence actually points to.
The schools that turned the feature off are not the ones that stopped caring about cheating. They are the ones that read the numbers and understood that an accusation tool is only as good as its false positive rate, and that at 750 innocent students a year, it is no good at all.
Sources
- Inside Higher Ed, Adelphi Student Wins AI Plagiarism Lawsuit
- CBS News New York, Adelphi student sues over AI plagiarism accusation and wins
- Stanford HAI, AI-Detectors Biased Against Non-Native English Writers (Liang et al., Patterns, 2023)
- The Markup, AI Detection Tools Falsely Accuse International Students of Cheating
- Weber-Wulff et al., Testing of detection tools for AI-generated text, International Journal for Educational Integrity, 2023
- Dugan et al., RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors, arXiv 2405.07940
- University of Florida News, Watching the detectors: Researchers probe efficacy and danger of AI detection tools, IEEE S&P 2026
- UF FICS, Commercial AI Text Detectors Lack Robustness for High-Stakes Environments
- Vanderbilt University, Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector
- Programs.com, Colleges That Turned Off AI Detection: Tracker and Full List
- GradPilot, AI Cheating Lawsuits Tracker
