Orion Newby handed in an essay for his world civilizations class. His professor gave it a zero.
The professor had run the paper through Turnitin, a tool used by schools to examine student writing. According to a New York court decision, he reported that the tool had assigned the essay an “AI-generated score of 100%” and concluded that the work had been produced by artificial intelligence. Newby denied using AI to write it. 1
The accusation became an academic integrity violation. Newby was told to attend a plagiarism workshop and warned that his registration could be placed on hold if he did not. He appealed; the university upheld the finding. In January 2026, a New York state judge annulled the finding and ordered the university to remove it from his record. 1
The court identified problems with the evidence and the university’s process. Its decision did not establish a universal error rate for AI detectors. It did show what can happen when an institution must turn a suspicious score into a decision about a person.
Schools, workplaces and contests are trying to decide when AI use crosses a line. Some choose to prohibit it. But a ban is only the beginning: enforcing one fairly requires understanding what the technology can reveal, what it cannot, and what an accused person needs to challenge a decision.
A score can start an accusation
Newby’s professor had more than a number on a screen. The court record says he also discussed Grammarly with Newby; Newby denied using its generative AI to write the essay. Newby’s family tried other detection tools, which returned a different assessment. Neither a high score nor a competing low score can independently settle who wrote a piece of text. 1
The way the university handled the case mattered too. The judge pointed to a later email from the professor that undermined the accusation, a failure to give Newby an opportunity to consult an adviser of his choice, and an appeal handled by the same official who issued the original determination. The court found the violation and the decision denying his appeal lacked a valid basis. 1
That distinction matters for anyone tempted to turn the case into a simple story of one bad number. A detection result entered a process involving people, rules, competing evidence and consequences. Each part could have changed what happened next.
What the percentage actually says
Turnitin’s own guidance describes its AI writing percentage as the share of qualifying text—principally long-form prose—that its system identifies as likely AI-generated or AI-generated and then altered. It is not a measure of the probability that the student cheated. The company says its model can misidentify human and AI writing and should not be the sole basis for adverse action against a student. 2
A classifier makes an estimate from patterns in the material it examines. It does not watch someone write. A false positive is human work flagged as AI-generated; a false negative is AI-generated work the tool fails to flag. The balance depends on the detector, the writing, the threshold chosen and the conditions in which it is used.
Researchers at Stanford and elsewhere tested seven detectors against essays written by people taking the TOEFL English proficiency exam. In that experiment, the tools incorrectly classified more than half of those essays as AI-generated on average. That finding raises a fairness concern for people writing in a second language; it is not a measurement of every detector’s error rate today, or of every student’s risk. 3
Scale changes what even a small error rate means. Suppose an institution screens 1,000 essays known to be human-written and a particular tool has a 1% false-positive rate on work like those essays. About ten would be flagged incorrectly on average. That example is arithmetic, not a claim that any specific detector has that rate in that setting. The proportion of incorrect flags among all flags would also depend on how many essays actually used AI and how often the tool catches them.
The chatbot does not remember writing every essay
A detector is not the only way an authority might try to make this call. In 2023, an instructor at Texas A&M University–Commerce asked ChatGPT whether it had written student papers and challenged students over the answers. The university said no students failed the course or were barred from graduating because of the incident. 4
There is an irony in asking a chatbot to police a rule against using a chatbot. OpenAI says ChatGPT cannot reliably determine whether text was AI-generated or whether it wrote the text, and may make up an answer when asked. A confident response to “Did you write this?” is not a record of authorship. 5
An assistant may have access to search, provenance tools or other external evidence. Those are separate operations, and their results need to be examined on their own terms. A fluent explanation is not proof that the assistant performed the checks it describes.
A missing watermark is not a human signature
There is another test that can sound like an AI detector but answers a different question. A watermark detector looks for a signal that a compatible system embedded in generated or edited content. Google’s SynthID, for example, supports watermarks for images, video and audio produced by participating systems. 6
If a tool finds a supported watermark, that can be evidence that a compatible AI system generated or processed some of the content. It does not automatically establish who used the tool, how much of the final work was generated, or whether that use broke a rule.
A negative result is even easier to misunderstand. “No watermark detected” does not mean “made by a person.” The content might have come from a system that does not use the watermark being checked. A signal might also be missing or too weak to find after editing or other changes. A tool can be confident about its own test and still be unable to answer the larger question of authorship.
“No watermark detected” does not mean “made by a person.”
OpenAI makes this distinction for its text watermark: it says an absent mark does not prove human authorship, and that a detected signal does not measure how much human judgment went into a passage. Its text system also has limitations involving length and editing. Those are claims about that system, not guarantees about all watermarks. 7
Visual content adds another complication: people may ask a chatbot to judge an image without knowing whether it checked a watermark at all. In a Columbia Journalism Review test using authentic news photographs, every model tested mislabeled at least one real photo as AI-generated. Some responses claimed verification steps that the researchers found had not been performed. That was a limited test of particular systems, not a measurement of every visual detector. 8
A fair process needs more than another score
Even a review by a person can fall short if that person treats the original flag as a conclusion. In one case reviewed by the independent adjudicator for higher education in England and Wales, an international student received a zero after a detector flag and an oral examination. The adjudicator upheld the student’s complaint: the institution had not shared its evidence before the examination or given the student a fair opportunity to explain the work. It recommended that the case be reconsidered. The finding concerned the fairness of the process; it did not establish who wrote the assignment. 9
Other institutions have decided the risk is too high for the tool’s role. Vanderbilt disabled Turnitin’s AI detector in 2023, explaining that false positives could affect many students when the tool is applied at scale. Institutions still have an interest in assessing students’ work, but they can choose how to gather evidence and what consequences follow. 10
A defensible policy starts before a flag appears. It should say what assistance is permitted, who may see a report, what independent evidence is required, and how a student can examine and challenge the case. It should account for people who cannot produce a complete editing history and for the cost of an investigation that continues after an accusation proves unsupported.
None of that makes unauthorized AI use impossible. It does make the institution responsible for the leap from a tool’s limited output to a judgment about someone’s conduct. If an organization wants to ban AI, understanding that leap is part of the job.
References
- Matter of Newby v. Adelphi University. (2026, January 28). New York State Law Reporting Bureau. https://www.nycourts.gov/reporter/3dseries/2026/2026_26021.htm
- Using the AI Writing Report. (n.d.). Turnitin Guides. https://guides.turnitin.com/hc/en-us/articles/22774058814093-Using-the-AI-Writing-Report
- GPT detectors are biased against non-native English writers. (2023). Patterns. https://doi.org/10.1016/j.patter.2023.100779
- Texas A&M Professor Misused ChatGPT to Detect AI Text Generation in Student Submissions. (n.d.). MIT AI Risk Repository. https://airisk.mit.edu/ai-incidents/538-texas-a-m-professor-misused-chatgpt-to-detect-ai-text
- Does ChatGPT tell the truth? (n.d.). OpenAI Help Center. https://help.openai.com/en/articles/8313428-does-chatgpt-tell-the-truth
- Google SynthID. (n.d.). Google DeepMind. https://deepmind.google/models/synthid/
- Our approach to EU text provenance rules. (n.d.). OpenAI. https://openai.com/index/eu-text-provenance/
- Why AI Models Are Bad at Verifying Photos. (2025). Columbia Journalism Review. https://www.cjr.org/tow_center/why-ai-models-are-bad-at-verifying-photos.php
- AI and academic misconduct - CS072502. (2025, July). Office of the Independent Adjudicator for Higher Education. https://www.oiahe.org.uk/resources-and-publications/case-summaries/ai-and-academic-misconduct-cs072502/
- Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector. (2023, August 16). Vanderbilt University. https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/