Falsely Accused of Using AI? Here's What to Do, Step by Step

The email usually arrives on a weekday afternoon and says something like "I'd like to discuss your recent submission." By the time you're reading this you probably already know what it's about, and you've probably already spent a bad night on it.

So, first, the thing worth knowing before anything else: being flagged by an AI detector is not the same as having done something wrong, and the research backs that up in unusually direct terms. A 2023 Stanford-led study found that commercial detectors classified more than 61% of human-written TOEFL essays — composed under exam supervision, no AI possible — as AI-generated. Being flagged is common. Being flagged wrongly is common.

This is a practical guide to what to do next. It's written from the detector side of the fence, which means it comes with an admission: tools like ours are part of why you're in this situation, and none of us should pretend a percentage deserves the weight it often gets.

Step 1: Don't respond immediately

The instinct is to fire back a defense within the hour. Resist it, at least until you've gathered what's below. A rushed, emotional reply tends to read as defensive, and it burns your best first impression.

A short acknowledgement is fine and buys you time: "Thank you for letting me know — I'd like to gather my drafting materials before we meet. Would [day] work?"

That's it. No defense yet.

Step 2: Collect your process evidence

This is the single most useful thing you can do, and it's why the timing above matters — you want this assembled before the conversation, not scrambled together during it.

You cannot prove a negative about your writing style. But you can show the writing happening over time, and that's what actually persuades people.

Document version history. Google Docs keeps full revision history under File → Version history → See version history. It shows the document growing, sentence by sentence, with timestamps. Microsoft Word has similar version tracking if you were working in OneDrive or SharePoint. This is the strongest single piece of evidence most students have, and many don't know it exists.

Earlier drafts. Any saved file with a different word count than the final.

Notes and outlines. Handwritten notes photographed on your phone count. Messy is good — messy is human.

Browser or library history from your research. Shows you reading the sources you cited.

Anything with timestamps. Emails to yourself, messages to a classmate about the assignment, a library checkout record.

The pattern you're establishing is that this document has a history. AI-generated text doesn't have one.

Step 3: Ask what the accusation actually rests on

At the meeting, it's entirely reasonable to ask this directly and calmly:

What specifically triggered the concern?
Is this based on a detector score, or on something else in the work?
Can I see which sentences were flagged, not just the overall percentage?

That last one matters more than it sounds. Flagged sentences scattered evenly throughout a document usually indicate the tool is penalizing your general writing style. Flags concentrated in one section suggest something different. If the flags are everywhere, that's an argument in your favor and worth naming.

If the answer is "the detector said 89%," then you now know the accusation rests entirely on a statistical estimate — and Step 4 is directly relevant.

Step 4: Know the research, and cite it accurately

You don't need to lecture anyone. But being able to reference specific published findings changes the tenor of a conversation considerably, because it moves you from "student denying it" to "person who knows the literature."

The Stanford study. Liang, Yuksekgonul et al., published in Patterns (Cell Press, 2023, Article 100779). Seven commercial GPT detectors tested against human-written text. They handled US eighth-grade essays with near-perfect accuracy — and flagged over 61% of supervised, human-written TOEFL essays as AI-generated. The researchers traced the cause: the detectors were penalizing lower linguistic complexity, not detecting machines.

Vanderbilt's decision. In August 2023 Vanderbilt University disabled Turnitin's AI detector and published its reasoning. At Turnitin's claimed 1% false positive rate applied to roughly 75,000 annual submissions, around 750 papers could be wrongly flagged. Vanderbilt concluded that AI detection software "is not an effective tool that should be used."

If English is not your first language, this is your central point. The 61% finding is specifically about non-native English writers. It's peer-reviewed, it's in a Cell Press journal, and it describes exactly the situation you're in. That isn't an excuse — it's the documented behavior of the tool being used against you.

Step 5: Put it in writing afterwards

After any meeting, send a short email summarizing what was discussed and what was agreed. Not adversarial — just a record.

"Thanks for meeting today. To summarize my understanding: [what was discussed], and the next step is [whatever it is]. I've attached my document version history and earlier drafts as mentioned."

If this escalates later, having a contemporaneous written record of what was said matters a great deal. If it doesn't escalate, you've lost nothing.

Step 6: If it escalates

Find your institution's academic integrity policy. Read the actual procedure document. Many now include specific language about AI detection evidence, and some prohibit accusations based solely on detector output.

Ask whether your institution has a stated position on AI detectors. Given that a number of universities have disabled these tools, the answer is sometimes surprising and sometimes helpful.

Find out if you have a student advocate or ombudsperson. Most institutions have someone whose job is exactly this, and students routinely don't know it.

Bring someone with you to any formal hearing if the procedure permits it.

Keep the tone factual. The strongest position in a formal process is the calm one with documentation attached.

One thing worth saying plainly

If you did use AI in a way your institution prohibits, this guide isn't a strategy for getting away with it, and I'd rather say that outright. Version history won't fabricate a drafting process that didn't happen.

But if you didn't — if you wrote it, and a percentage says otherwise — then the number is wrong, this happens frequently enough to be well-documented in peer-reviewed research, and you're entitled to say so with evidence.

Quick reference

Situation	What to do
Just received the email	Acknowledge briefly, request time to gather materials. Don't defend yet.
Preparing for the meeting	Collect document version history, drafts, notes, outlines, research history.
In the meeting	Ask what triggered it, whether it's detector-based, and to see sentence-level flags.
English is your second language	Cite the Patterns study: over 61% of human-written TOEFL essays were flagged as AI.
Accusation rests only on a score	Note that a detector output is a probability estimate, not evidence of authorship.
After any meeting	Send a summary email creating a written record.
It escalates formally	Read the integrity policy, find a student advocate, bring documentation.

Frequently asked questions

How do I prove I didn't use AI to write my essay?

You can't prove a negative about your writing style, which is the fundamental unfairness here. What works instead is process evidence: Google Docs or Word version history showing the document being written over time, earlier drafts, notes, and outlines. A document with a visible editing history looks very different from generated text.

Can I be expelled based on an AI detector result?

Policies vary by institution, and some do treat detector output as actionable. But the score itself establishes nothing about authorship — it's a statistical probability with no source behind it. If you're facing formal proceedings, read your institution's written integrity policy and find out whether it addresses AI detection evidence specifically.

What if my professor won't accept the research about false positives?

Put your position in writing, attach your process evidence, and ask about the formal appeals procedure. Most institutions have an academic integrity process with review stages beyond the individual instructor, and a student advocate or ombudsperson whose role is to help students navigate exactly this.

Why do AI detectors flag non-native English speakers so often?

Most detectors rely heavily on perplexity — a measure of how predictable word choices are. Simpler vocabulary and more regular sentence structure produce low perplexity, and so does AI-generated text. The 2023 Patterns study confirmed the mechanism directly: improving word choice in TOEFL essays reduced misclassification, while simplifying native-speaker essays increased it.

Should I run my own work through a detector before submitting?

It can tell you whether you're likely to be flagged, which is useful for preparing. But don't rewrite genuine work to game a detector — you'd be distorting your own writing to satisfy a tool that's already documented as unreliable on exactly this. Knowing your risk is useful; writing to the algorithm isn't.

Sources

Liang, W., Yuksekgonul, M., et al. (2023). GPT detectors are biased against non-native English writers. Patterns (Cell Press), Article 100779.
Vanderbilt University Brightspace (2023). Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector.

Accusations aren't limited to coursework. If it's happened with a job application, see AI detection on resumes and cover letters; if you work in content and it's happened to a deliverable, AI detection for SEO agencies covers the same problem in a commercial setting.