Why an AI detector score is not proof on its own
The same text can get one hundred percent from one detector and zero from two others. In the New York case of freshman Orion Newby, this is exactly what happened: apart from one tool's score, the university had nothing, so the court overturned the penalty and ordered the university to clear the student's record. In Australia, Australian Catholic University came to the same rule on its own, without any court: when a detector signal was the only evidence, it dismissed the accusation right away. The reason is practical. In one year the university handled a record number of academic misconduct cases, and there is no way to work through a stream like that on scores alone.
According to internal documents reviewed by ABC News, a sizable share of the accusations fell apart at the very first check. GradPilot, the site that keeps a list of lawsuits over AI accusations, sums up the case law in one line: no court has found a detector unlawful in itself, and every ruling came down to process. The logic is simple. A detector gives a probability estimate, but it's a person who gets expelled. Something else has to sit between those two things, and the first question of any appeal is whether the university has that "something else."
How to prove you wrote your paper: version history and drafts
In the cases we reviewed, the defense was built on four things: version history, early drafts with notes, independent checks of the text, and the student's account of the work, told the same way from the first conversation to the last. All four have one thing in common: they can't be created after the fact. Version history in Google Docs and Word is the most important of the four. Lawyers who handle these cases in the US tell the Los Angeles Times it is the backbone of a defense, because it shows how the text grew and separates what was written from what was pasted. Drafts and notes show the same thing: how the work unfolded. Keeping them until the end of the semester costs nothing. Independent checks, meaning running the text through other detectors, helped in the New York case: freshman Newby had the results in his file, and by the start of the trial his defense knew more about the disputed essay than the other side did.
The opposite case shows how much a consistent story matters. Graduate student Haishan Yang of Minnesota, accused of using ChatGPT on an exam, changed his story as the case went on, from full denial to "AI was checking my English," and among his files were drafts that looked generated. Those files worked against him. The file works for the side that first learned what was in it.
How to dispute an AI cheating accusation: four appeal questions
Once an AI plagiarism accusation has been made, an appeal is built on four questions courts and ombudsmen have used, time after time, to judge these cases.
- Was the accused shown the actual evidence? The OIA ombudsman, who handles student complaints against universities in England and Wales, sided with students twice in one year on exactly this question: in one case the university told the student only its conclusion and did not show the evidence; in the other, the panel wrote down "searched Google for synonyms" as an admission of AI use, and the recording of the meeting disproved it.
- Is there evidence besides the detector? The New York case fell apart on this question: apart from the score, the university had no evidence.
- Were drafts and work history requested? The English university's rules required requesting them whenever AI use was suspected, the panel did not request them, and the ombudsman treated that as a violation.
- Was the student's language background considered? This is where cases against international students stumble most often.
None of the four questions requires a lawyer. They are asked in writing, in the very first reply to the accusation, and the university's reaction shows right away how solid its case is.
AI detector false positives: who gets flagged more often
Stanford researchers showed back in 2023 that AI detectors of that generation flagged texts by non-native English speakers more often: the algorithms mistook simple vocabulary and short sentences for a sign of AI.
This bias shows up in real cases too. According to a review by the think tank HEPI, most published OIA rulings on such accusations involve international students, and in one of them the ombudsman faulted a university for not checking whether its detector was reliable for a student whose first language isn't English. Newer detectors, including It's AI, are open about the limits of their accuracy: a probability estimate is a reason to ask questions, and the case review is what should answer them. For an international student, language background is a legitimate argument on appeal, and the ombudsman has already accepted it.
When an appeal against an AI accusation fails
An appeal doesn't help when the university has evidence besides the detector and the procedure was followed. If on top of that the student's own explanations contradict each other, there is almost no chance. The Minnesota case of graduate student Yang looked exactly like that. Instructors gave the exam questions to ChatGPT and got near-word-for-word matches with Yang's answers, and the whole process, from notice to hearing, was done by the book. The federal court said it plainly: it gets involved in universities' internal decisions only in extreme cases, and the university's decision looked justified. The school case in Massachusetts ended the same way: the accusation rested on the teacher's judgment, the court found it reasonable, and the student's family did not get a review.
The advice from student forums to "just file an appeal" doesn't work as well as it sounds. An appeal helps when there is something to show.
What to do today, before any accusation comes
Even a dispute the student wins costs a lot: the defense in the New York case cost the Newby family six figures, Communications of the ACM writes. Preparing in advance is cheaper, and it doesn't take much time.
The steps for a student are simple. Turn on version history wherever the work is written: in Google Docs it is on by default, and in Word it works when autosave to the cloud is on. Make a separate folder for every assignment and put notes and early drafts in it. Keep the folder until the end of the semester, even when the grade is already in. If the work went through any checker, save the results. Explain how the work was written, and keep the explanation the same every time.
There is something unfair here: an honest person has to collect proof of their honesty in advance. But detectors aren't leaving universities, there are more cases every year, and rulings in these cases turn on documents. Documents collected in advance are better protection than anything said out loud.
For an instructor, the same preparation saves time too. Checking the four appeal questions before filing a report takes a few minutes. A case that falls apart six months later takes much more, and after a false accusation students trust the instructor less, and the checks themselves too.
If a letter from the dean's office does come, there's no need to panic. The outcome of these disputes is decided by preparation, and all of it happens in advance. In the cases we reviewed, the winner was the side that knew its own weak spots better.
FAQ
Can a student be expelled because of an AI detector score alone?
No ruling so far has upheld a punishment that rested on an AI detector score alone. In the cases decided through early 2026, a New York court threw out a penalty backed only by a score, and Australian Catholic University dismissed score-only accusations without a hearing. The punishments that survived court review, like the Minnesota expulsion, had separate evidence behind them: matching ChatGPT outputs and inconsistent explanations from the student. A detector score starts a review; on its own it has not been enough to end one.
How accurate are AI detectors on student essays?
Independent tests put AI detectors well below their marketing. A peer-reviewed study of 14 detection tools, cited by Communications of the ACM, found none reached 80% accuracy, and one in five AI-generated texts passed as human. A separate evaluation of 805 samples, cited by the UK think tank HEPI, measured 39.5% average accuracy on unedited AI text, falling to 17.4% once the text was lightly edited. This gap between claimed and tested accuracy is why a score works as a signal for a closer look, never as a verdict.
Can Grammarly or translation software trigger AI detectors?
Yes. Grammar checkers, paraphrasing tools, and translation software rewrite wording, and wording is what AI detectors read. The essay at the center of the 2026 New York lawsuit had been written with Grammarly assistance before a detection service flagged it as fully AI-generated. Universities also treat these tools unevenly: at Australian Catholic University, confirmed misconduct cases included the undisclosed use of AI tools to paraphrase or translate content. A student using any rewriting tool should check what the course policy says about it before submitting.
What should a student do first when accused of AI cheating?
Answer in writing and ask to see the evidence before giving any explanation. University procedures usually give the accused the right to review the case materials and to bring an adviser to the hearing: in the Minnesota federal case, the student had an advocate and a full opportunity to present evidence, and the court treated that as proper process. A written reply also fixes the record. Statements made in a first panicked conversation cannot be unsaid later, and hearing panels quote them.
Should I run my essay through an AI detector before submitting it?
A pre-submission check removes surprises, which is what decided the reviewed cases. It shows whether any passage reads as machine-written before an instructor sees it, and the saved report becomes one more dated document in the student's folder. Paragraph-level tools such as the It's AI detector show which specific passages raise the score, which matters more than the overall number: a flagged paragraph can be rewritten or explained while the work is still in the student's hands.
Can you sue a university over a false AI cheating accusation?
Students have sued and won, but courts step in reluctantly. The New York student won through a state-court procedure for reviewing arbitrary university decisions, while the Minnesota federal court repeated that it interferes with academic rulings only in extreme circumstances. The usual path is shorter: a written internal appeal first, then an external complaints body such as the OIA ombudsman in England and Wales. Court is what remains when the file is strong and the university still refuses to move.
