Accused of using AI on an essay: what to do
Preserve everything and change nothing, then ask what is actually alleged before you explain yourself. Almost every way this gets worse starts with something written or deleted in the first hour. Your version history, drafts, notes and chat logs are the account of how the essay came to exist, and they decay fast. The situation is stressful and it is routine for the people handling it. Save the record tonight, then reply.
First: preserve the record
Do this before you think about the reply, because the evidence decays.
- Version history. Google Docs keeps a timestamped edit history; Word keeps versions when the file lives in OneDrive or SharePoint. Export or screenshot it now. Histories can be lost when files are moved, copied or converted.
- Drafts and outlines. Every earlier file, including the messy ones. Especially the messy ones.
- Reading trail. Downloaded PDFs, annotated papers, library loan records, browser history for the days you worked.
- Notes. Handwritten pages, phone notes, voice memos, messages to friends about the assignment.
- Chat logs. If you used a chatbot for anything, keep the conversation. Deleting it is the single worst move available and it is the one people reach for first.
Copy all of it somewhere separate from your working folder. Then leave the submitted file alone.
Second: find out what is actually alleged
"Using AI" covers several very different accusations, and they carry different evidence and different consequences. Ask which one you are facing.
Undeclared AI use. The claim is that a machine wrote some or all of the text. Usually triggered by a detector score, a marker's impression, or both.
Fabricated sources. The claim is that references in your bibliography do not exist or do not say what you cited them for. This is a factual allegation that can be checked in minutes, and it is the more serious one because it does not rest on anyone's judgement about style. If this is the accusation, the guide you want is what to do when your professor says your sources don't exist, and you should start auditing your reference list tonight.
Plagiarism. A similarity finding, which is a different report from the AI indicator and means something different. See what the Turnitin AI score actually measures if the two have been mixed together in the email you received.
Ask for the evidence in writing. Published procedures at many institutions give you a right to see it. Marquette's academic misconduct hearing overview and Indiana University's student code procedures both set out notice, disclosure and response rights. Find your own institution's equivalent page and read it before you say anything substantive.
If you wrote it
You cannot prove a negative, and you do not have to. What you can do is give a detailed, checkable account of how the essay came to exist, which is a different and much more achievable thing.
Lead with process. Version history that shows a document built over several sessions, with false starts, restructures and deleted paragraphs, is the strongest artefact most students have. It is not proof, and you should not oversell it as proof. It is evidence of a writing process, and it is much harder to assemble after the fact than people assume.
Be able to talk about your own essay. Why you framed the argument that way. Which source changed your mind. What you cut and why. Where a particular statistic came from. In practice this conversation persuades more often than any document, because it is the part nobody can outsource.
Know the limits of the detector evidence. You do not need to argue detector accuracy, and leading with it can read as deflection. But it is fair to note, calmly and once, what the tool's own maker says about it. Turnitin states that its indicator does not make a determination of misconduct and should not be the sole basis for action. Vanderbilt University disabled the feature entirely in 2023, partly over false-positive volume at scale.
If you write in your second language, say so. Research on AI detectors published in Patterns found that detectors misclassified a majority of TOEFL essays by non-native English writers as AI-generated while scoring native-speaker essays accurately. That study tested other detectors rather than Turnitin, so present it as context about how these tools behave, not as a claim about the specific tool used on you.
A published case shows how these points land. In case CS072504 at the Office of the Independent Adjudicator, which reviews student complaints against providers in England and Wales, an international student was called to a viva after Turnitin flagged a module assignment as high in AI-generated content, then to a misconduct panel, which found two offences, unacknowledged AI material and collusion, and imposed a module retake, a 10% cut to the overall degree mark and a reflective essay. In July 2025 the OIA found the complaint partly justified. The provider "had not shown what evidence led the panel to conclude that the student had used AI to the extent that they committed academic misconduct", had misrepresented the viva (the transcript showed Google used for synonyms, not AI used to paraphrase), had not asked for the notes, draft plans and earlier versions its own procedures called for, and had not considered "whether Turnitin's AI detection might be less reliable for non-native English speakers". It recommended hearing the allegations again from scratch. The collusion finding stood, so read it for what it shows: the score was a trigger, the evidence was the process, and the provider lost the AI finding because it could not show its evidence.
If you did use AI and did not declare it
Then the advice is genuinely different, and pretending otherwise would not help you.
The first thing to establish is what your course actually permits, because the range is wide. Some courses ban AI outright, some allow it with disclosure, some allow it for brainstorming and editing but not drafting, and plenty have no clear rule at all. Read the assignment brief and the module handbook, not just the university policy. If the rule was vague, that matters, and it is a fair thing to say.
The second thing: what makes these cases go badly is almost never the original use. It is what happens next. Fabricated evidence, invented drafts, a story that changes between the email and the meeting. The most-cited fabricated-citation case in the world, the 2023 sanctioning of two New York lawyers who filed a brief full of ChatGPT-invented cases, turned on precisely this. The judge's ruling rested on their conduct after being questioned rather than on the original mistake. Integrity processes are built by people who have seen every version of the cover-up, and they respond to it much more harshly than to the underlying act.
The third thing: get advice from someone whose job this is before the meeting. A student union representative, an academic adviser, a student advocacy service. They know your institution's process, they know what outcomes are typical for a first case, and they will tell you whether admitting the use early tends to help in your specific system. Many procedures do treat cooperation and acknowledgement as mitigating. Some formalise it. That is worth knowing before you decide what to say.
What we will not do here is tell you what to say in the room. Your situation, your policy, your call. What is clear from how these processes actually work is that an accurate account, given early, produces better outcomes than a story that has to survive scrutiny it cannot survive.
The middle ground, which is where most people are
Very few essays are entirely machine-written or entirely untouched. Most students who get flagged did something in between: asked a chatbot to explain a concept, used it to tidy a clumsy paragraph, generated an outline, had it suggest sources. Whether any of that counts is a policy question, and it varies by course, which we go through in is using AI to write essays cheating?
If this is you, specificity is your friend. "I used it to rephrase two paragraphs in the discussion and to check my grammar; the argument, the structure and the sources are mine" is a claim someone can evaluate. "I barely used it" is not. Vagueness reads as evasion even when it is just embarrassment.
What the meeting is actually like
Formats differ, but the published procedures share a shape. You get written notice of the allegation. You are usually entitled to see the evidence beforehand and to bring an adviser or support person. The meeting itself is typically a small panel or a single officer, who states the allegation and the evidence, then gives you a chance to respond and ask questions. You are normally present for the discussion but not the deliberation. Decisions commonly turn on whether the allegation is more likely than not to be true, a lower bar than a criminal standard. Then there is an appeal route with a deadline.
It is a meeting, not a trial. Bring your evidence organised, printed or on a laptop. Answer what is asked. Do not fill silences with speculation about detectors or about other students. If you do not remember something, say you do not remember rather than guessing, because a guess that turns out wrong costs you credibility you cannot get back.
Note the appeal deadline the moment you get the outcome. Appeals are usually tightly time-limited and typically require a specific ground, such as procedural irregularity or new evidence, rather than simple disagreement.
Afterwards
Whatever the outcome, the thing worth changing is the paper trail. Write in a tool that keeps a history. Keep your notes. Record where each source came from and the exact line that supports each claim, as you write rather than at the end. That habit protects you from the AI question, from the plagiarism question, and from the far more dangerous one about references that do not exist. If you want to know what human markers notice, as opposed to what software flags, what actually gives a ChatGPT essay away is the honest version of that list.