Accused of Using AI When You Didn't: What to Do (2026)

First: this happens to people who did nothing wrong

If you are reading this the day you got the email, start here: an AI detector flagging your work is not evidence that you did anything. It is a statistical guess about writing style, produced by a tool that never saw you write and cannot know who did.

This is common enough that institutions have started backing away from these tools. Vanderbilt University disabled Turnitin's AI detection in 2023, reasoning that even a small error rate applied across tens of thousands of papers produces hundreds of wrongful accusations a year. Several teaching centres now advise staff that a detector score cannot be used as evidence on its own.

That does not make the meeting less frightening. It does mean you are not arguing from a weak position, and the rest of this page is about how to argue from a strong one.

Why detectors flag honest work

These tools measure how much your writing resembles patterns common in machine-generated text. They look at things like how much your sentence lengths vary, how evenly your paragraphs run, how often certain stock phrases appear, and how predictable your word choices are.

Read that list again and notice the problem: those are the traits academic writing is taught to have. A well-structured thesis has even paragraphs, formal register, consistent terminology and careful transitions. The better you follow the guidance you were given, the more machine-like you score.

Two groups get hit hardest. Students writing in a second language, whose more regular sentence construction and smaller working vocabulary read as machine-like to a detector. And careful, methodical writers who edit heavily — because editing removes exactly the irregularity these tools treat as proof of a human hand.

What to do, in order

1. Do not reply immediately. Acknowledge receipt, ask for the specific allegation in writing, and ask what the deadline is. You are entitled to know what is being claimed before you respond to it.

2. Ask exactly what was flagged.A percentage is not an allegation. Ask which passages were flagged, by which tool, and what that tool reports as its false-positive rate. This single request reframes the conversation from “defend yourself” to “let us both look at the evidence”.

3. Gather your process evidence before you write anything else. This is the part that wins, and it is easier the sooner you do it. See the section below.

4. Find your institution's actual policy. Most have a written academic integrity procedure specifying what constitutes evidence, what your rights are, and who decides. Read it before the meeting. Procedural points are often the strongest ones available to you.

5. Ask whether you may bring someone. Many institutions allow a student representative or advisor. Most students never ask.

6. Be precise about any tool you did use. If you ran a grammar checker, say which one and on what. Precision reads as honesty because it is difficult to fake.

The one question to ask

If you take one thing from this page, take this. Ask:

What false-positive rate does this tool publish, and on what corpus was it measured?

A false-positive rate is how often the tool gives that same verdict to writing a human definitely produced. Without it, a score of 80% is uninterpretable — it might mean the tool is confident, or it might mean the tool gives 80% to a quarter of all theses ever written.

Most vendors do not publish this figure for the conditions that apply to you: your language, your document length, your discipline. Asking for it is not a rhetorical trick. It is the question that determines whether the number in front of you means anything at all, and it is entirely reasonable for the person being accused to ask it.

We publish ours on our accuracy page, broken down by language and document length, along with the corpora behind each figure and the cases where we refuse to give a verdict at all.

The evidence that actually helps

Process evidence beats argument. In rough order of persuasiveness:

• Version history. Google Docs and Microsoft Word both retain revision history showing the document growing over time. A thesis that accumulated over weeks looks nothing like one that was pasted in. Export or screenshot it before anything expires.
• Your notes and drafts. Handwritten notes, outlines, abandoned paragraphs, the messy second draft. Imperfection is evidence.
• Your reading trail. Annotated PDFs, library loans, reference manager entries with dates, highlighted sources.
• Supervision record. Emails and meeting notes where you discussed the argument as it developed. A supervisor who watched the work take shape is your strongest witness.
• Your ability to discuss it. You can explain why you cut a section, why one source disagreed with another, what you would do differently. Nobody who did not do the work can.

Do this today rather than the night before the meeting. Cloud version histories do not last forever.

What to say in the meeting

Stay factual and specific. Three things worth saying plainly:

“I wrote this. Here is the version history showing it being written.” Lead with the evidence rather than with your feelings about the accusation, however justified those are.

“What false-positive rate does this tool publish for documents in this language and of this length?” Ask it neutrally. You are asking how to interpret a number, which is a reasonable thing to want to know.

“I am happy to discuss any part of this work in detail.” Then do. An oral defence of your own argument is the evidence a detector cannot produce and a person who did not write it cannot fake.

If the outcome goes against you, ask about the appeals process in the same meeting, and ask for the reasoning in writing.

Check before you submit, not after

Everything above is harder than the alternative, which is knowing in advance. Running your own check before submission tells you whether your writing sits in a range that tends to get flagged, and gives you a report — with a published error rate — that you already hold if the question ever comes up.

Our free AI check covers the first 1,500 words with no account and no signup, and states our measured false-positive rate next to the verdict. If it flags something, it shows you the exact sentences rather than a number you cannot act on.