Accused of Using AI

Accusations that a student used AI to write an assignment, and how a detection score gets treated as proof when it is not.

The accusation usually arrives as something small

The accusation usually arrives as something small. A comment in the grading portal, a zero with no explanation, or an email from a professor saying only that he would like to talk about your last paper. Sometimes a number is attached — an AI writing score of 68 percent — and sometimes there is nothing but a sentence saying the work “does not appear to be yours.” Either way a file has been opened, and the conversation you are being invited to is part of it.

What makes these cases different from ordinary plagiarism is that there is usually no source document. In a plagiarism case the school can put the original next to your paper. In an AI case there is often nothing but a statistical estimate produced by software and an instructor’s impression that the writing does not sound like you. That is a weaker evidentiary position than most students assume — and weaker than most students behave as though it is. It is also the fastest-moving area in student discipline: institutions are still rewriting integrity codes to name generative AI, so many current cases are charged under rules written for something else.

What does an AI detector actually produce?

A probability, not proof. These tools do not watch you write and do not identify a source. They score text on statistical properties — how predictable each word is given the words before it, how uniform the sentences are, how little the rhythm varies. Prose that is clean, formulaic and free of idiosyncrasy scores as machine-like, because that is what the model was trained to associate with machine writing.

The vendors say as much. Turnitin’s guidance states that the AI writing percentage “should not be used as the sole basis for action” and that Turnitin “does not make a determination of misconduct”; it displays no score at all between one and nineteen percent, because false positives are possible there. OpenAI withdrew its own classifier in July 2023 “due to its low rate of accuracy” — on its own evaluation it caught 26 percent of AI text and wrongly flagged human writing about 9 percent of the time. A score is a reason to look. It is not a finding, and a panel treating a percentage as a match to a source has made an error worth putting in writing at the time.

Why do false positives cluster on certain students?

Because what detectors read as machine-like are the traits of writers who learned English formally, who write to a taught template, or who have been trained out of stylistic risk. A 2023 study published in Patterns by Stanford researchers ran seven widely used detectors against TOEFL essays by non-native English speakers and against essays by U.S. eighth-graders. On the eighth-graders the detectors were close to accurate. On the TOEFL essays the average false positive rate exceeded 61 percent, and all seven unanimously misclassified 18 of the 91 essays. Rewriting those essays with more varied vocabulary dropped the rate to roughly 12 percent — the tools measure linguistic range, not authorship.

The same logic reaches students who write in short, declarative, heavily structured sentences because that is how they were taught to compensate: students with dyslexia using assistive drafting software, students with ADHD working from rigid outlines. An accommodation that produces cleaner prose can raise a score, and that record belongs in the response — see Section 504 and ADA accommodations. Institutions have reacted differently. Vanderbilt University disabled Turnitin’s AI detector in 2023 and published its reasoning, noting that even a one percent false positive rate across roughly 75,000 annual submissions would wrongly flag hundreds of papers. Other schools kept it.

Why is the informal chat with the professor the most dangerous moment?

Because it is not informal. At most institutions the instructor has discretion to resolve the matter in the course or refer it upward, and that choice turns largely on what you say in the meeting. There is no transcript, no advisor, and no notice of the evidence. What exists afterward is the professor’s written summary, which becomes the charging document and gets quoted against every later account you give.

Students lose these cases in that room, usually by trying to be cooperative. “I used it a little, just for grammar” is offered as a partial denial and recorded as an admission. You are not required to answer on the spot. Asking in writing for the specific allegation, the policy provision and the evidence turns an interrogation into a conversation about a document you have both read.

What evidence actually clears a student?

Process, not protest. The most effective answer is a record of the document being built over time, because that is the one thing a language model does not produce.

  • Version history — Google Docs revision history, Word’s version pane — showing incremental edits and timestamps
  • Draft files, outlines and notes saved separately, each with its own creation and modification dates
  • Library, database and browser records showing when you located and read your sources
  • Messages to a professor, teaching assistant, writing center or classmate while the work was in progress
  • Earlier graded work in the same course, which establishes what your writing actually looks like

Two cautions. A paper written in one sitting and pasted from a notes app looks much like one pasted from a chatbot, so thin version history is not evidence against you — but it costs you your best exhibit. And never create or backdate anything; metadata gives it away, and a fabricated exhibit turns a defensible case into an unwinnable one.

What should I preserve tonight?

Before you answer the email, reopen the document, or tidy anything up. Every additional edit overwrites part of the record you need.

  • Export or screenshot the full revision history of every draft, and store it outside the document
  • Download local copies of the assignment, drafts and notes without re-saving the originals
  • Save the syllabus, prompt and course AI policy exactly as posted, with the date you captured them
  • Save the accusation in full — email, portal comment, detector report, highlighted PDF
  • Write your own dated account of how you wrote the paper, before the meeting reshapes your memory of it

What does my school’s policy actually prohibit?

This is where a surprising number of these cases come apart. Does the institution’s integrity code — the code, not the syllabus — name generative AI or unauthorized assistance in terms that reach what you are accused of? Did the course have a stated AI policy, where was it stated, and was it consistent across the syllabus, the prompt and what was said in class? Was the rule in effect when you submitted?

The syllabus-only problem is real. Many institutions have delegated AI rules to individual instructors, so one course permits AI for brainstorming, another bans it, a third says nothing, and a fourth carries a policy contradicting the department’s. A student charged under a rule that appeared only in a lecture slide, changed mid-semester, or reads “discouraged” while enforced as a prohibition has a genuine notice argument. At a public institution that sounds in due process; at a private one it is a breach of the school’s own published rules. Either way it is an argument about the rule rather than the software, and often the stronger one.

What if I did use AI, or used some of it?

Then the questions become how much, for what, and what the policy actually forbade. Generating a paragraph and submitting it as your own is different from checking grammar, different from asking a chatbot to explain a concept before you wrote anything. Many codes draw those lines. Many instructors do not, at first.

Where a violation did occur, the work shifts from contesting the finding to controlling what follows it: the sanction, whether a transcript notation attaches, whether the finding can be held in abeyance, and what the school says if a licensing board asks in five years. Second findings escalate sharply, and in graduate and professional programs an integrity finding also feeds a professionalism or fitness file, which is often where the durable damage is done. The wider framework is on the academic misconduct and honor code page.

What I can help with

  • Reading the accusation against the integrity code and the course policy in force when you submitted
  • Preserving version history, drafts and metadata before the record degrades
  • Preparing you for the instructor meeting, or responding in writing instead
  • Building the process record that shows how the document was actually written
  • Challenging a detector score treated as a finding rather than as a lead
  • Raising accommodation, language and writing-history issues in a form a panel can act on
  • Working to limit the sanction, the notation, and what the school discloses later
  • Preserving procedural objections for appeal and for any review beyond the campus

Why timing matters

Two clocks run at once. Response and appeal windows in integrity cases are commonly measured in days, and they are enforced whether or not you had three exams that week. The evidence runs on its own clock: version history is finite, cloud platforms trim revision data, and every time you reopen a file you write over part of what would have cleared you.

The students who end up with real options stop editing, preserve everything, and find out what the file contains before they explain themselves. A flat-fee Full Read + Game Plan is described on the Fees and Scope of Services page.

Where AI-detection cases stand in 2026

Three years into the generative-AI era, the legal landscape for these accusations has settled into a recognizable shape, even though no Florida court and no Eleventh Circuit panel has yet ruled directly on whether a detector score is competent evidence of misconduct.

The tools have not gotten more reliable in the way that matters. OpenAI withdrew its own AI-text classifier in 2023 for low accuracy and has not replaced it. Turnitin’s AI indicator ships with the company’s own caution that it should not be the sole basis for an integrity finding, and a number of universities disabled the feature after false-positive complaints. Peer-reviewed work has shown that detectors flag writing by non-native English speakers at far higher rates than writing by native speakers, and that simple edits change the score dramatically. None of that means a detector report is worthless to a school; it means the report is a reason to ask questions, not an answer.

Courts have been reluctant to intervene where the school’s policy was clear. The most-discussed decision so far, from a federal court in Massachusetts in late 2024, declined to stop a high school from disciplining a student who had used a generative tool on a project, largely because the school’s written policy prohibited exactly that and the student had notice of it. The lesson for college students is the opposite of what the headline suggests: the case turned on the policy, and most cases I see turn on whether the school’s policy actually prohibited what the student did, whether the syllabus said anything at all, and whether the student was told in advance. A vague or contradictory policy is your best friend.

In Florida, the procedural rules do real work. At a state university or Florida College System institution, section 1006.60 of the Florida Statutes requires the school to give you, at least five business days before any disciplinary proceeding, all known information relating to the allegation, including anything exculpatory. The detector report, the settings used to generate it, the instructor’s notes, and any drafts or metadata the school pulled are all “information relating to the allegation.” If you did not receive them, that is a procedural defect. Florida’s appellate courts also require a misconduct finding to rest on competent, substantial evidence — and a percentage from a tool that its own vendor says should not be used alone is a thin foundation. At a private school the same arguments run through the student handbook, which Florida courts treat as a contract.

The burden of proof is usually preponderance, set by the code. That is lower than you would like, but it also means the school has to show that it is more likely than not that you used a prohibited tool in a prohibited way, not merely that a program thinks your paragraph is “AI-like.” Version history, research notes, prior writing samples, and a walk-through of how you built the piece routinely outweigh a score.

How to challenge the detector report itself

Ask, in writing, for the complete report and the settings used to generate it; the name and version of the tool; whether the instructor ran the same tool on the student’s earlier work, and what it returned; and the school’s written policy on how detector output may be used. Ask for the same for any comparison the instructor ran against a generative tool’s output. Then build the affirmative record: the document’s version history exported from the platform you wrote it in, the browser and research history for the drafting window, the sources you cited and where you found them, and two or three earlier writing samples for a style comparison. If your first language is not English, say so and put the published research on detector bias in the file. If you used a permitted tool — a grammar checker, a citation manager, a translation aid — identify it and show what it did and did not do. The goal is a file in which the detector score is the least persuasive item.

Where to go next

Accused of using AI on an assignment: the first 48 hours
AI watermarking and content credentials: what they mean for students accused of using AI
Suspension appeals: academic vs. disciplinary, and what to put in the letter
Academic misconduct and honor-code cases
Plagiarism accusations
Exam cheating and remote-proctoring flags
Due process at public institutions

Common questions about AI academic integrity accusations

Can a school punish me based only on an AI detector score?

Many try, and whether it holds up depends on the code and the panel. The score is a probability estimate, not a match to a source, and the leading vendor states in its own guidance that the percentage should not be the sole basis for action. Ask on the record what evidence exists besides the score.

How do I prove I wrote it myself?

With process evidence rather than argument. Version history from Google Docs or Word, dated drafts and outlines, research notes, messages about the assignment while you were writing it, and earlier graded work in the same course together show a document being built over time. Preserve it before you touch the file again, and never backdate anything.

Should I meet with my professor before I know what the evidence is?

Not without preparation. That meeting is usually where the case is decided, because the instructor typically has discretion to resolve it in the course or refer it, and your statements there become the record. Ask in writing for the allegation, the policy provision and the evidence, then decide what to say.

Is using Grammarly or a translation tool a violation?

It depends entirely on the policy, and many policies have not caught up. Grammar checking, spelling, translation, citation managers and accessibility software are permitted in most courses and prohibited in a few. If an approved accommodation involves assistive drafting software, say so early and document it.

What if my professor never had an AI policy?

Then the charge has to rest on the institution’s integrity code, and the question is whether that code, as written when you submitted, reached what you did. Silence in a syllabus is not automatically permission. But a rule announced after the fact, stated inconsistently, or phrased as discouragement and enforced as prohibition is a real notice problem.

Will an AI finding follow me to graduate school or a licensing board?

It can. Integrity findings frequently appear in a dean’s certification, and graduate, professional and licensing applications ask directly about discipline. Whether a notation attaches, how long it remains, and what the school says when asked are usually negotiable at resolution and much harder to change afterward.

Facing this now?

Deadlines in these matters are often short. Reach out for a confidential consultation about your situation.