Humanetext

Falsely Accused of Using AI? How to Defend Your Work

What to do in the first 24 hours, how to assemble evidence that your writing is yours, and how to challenge a detector score without sounding defensive.

AI detectionBy Humanetext Editorial9 min read

The short answer

If you are wrongly accused of using AI, stop editing the document immediately and export its version history - that record of the writing happening over time is the strongest evidence most people have. Then find your institution's policy and its appeal deadline, respond calmly with your evidence first, and offer to discuss the work in person. Raise the detector's unreliability last, not first, because leading with it reads as deflection.

Getting told your work has been flagged as AI-generated is a particular kind of awful. You know what you wrote. You remember writing it. And now you are being asked to prove a negative, usually by someone holding a percentage score they believe is more reliable than it is.

This guide is about what to actually do. Not the theory of detection — we cover that in how AI content detectors actually work — but the practical sequence: what to preserve, what to say, what to send, and what not to do.

First, understand what you are up against

The accusation almost always originates from a number. Turnitin returns an "AI writing" percentage. GPTZero returns a probability. A teacher or editor pastes your paragraph into a free checker and gets back "98% AI." That number then gets treated as a measurement, like a temperature reading.

It isn't. It's a classifier's guess about how statistically predictable your word choices are. Writing that is clear, well-structured, grammatically tidy, and uses common vocabulary scores as more AI-like than writing that is messy and idiosyncratic. That is the whole mechanism. It means the students and writers most likely to be falsely flagged are the ones who write plainly and carefully.

This matters for your defence because it reframes the argument. You are not arguing "your tool is broken." You are arguing "this tool measures predictability, my writing is predictable for reasons that have nothing to do with AI, and here is the evidence of how I wrote it."

The first 24 hours: preserve everything

Do this before you reply to anyone. Evidence degrades, and the strongest evidence is the kind you cannot fabricate after the fact.

Stop editing the document. Every save from this point forward muddies the version history you are about to rely on. Duplicate the file and work in the copy if you need to keep going.

Export the version history. In Google Docs, open File → Version history → See version history, and screenshot the timeline showing named versions and timestamps. Google Docs history is the single most persuasive artefact most people have, because it shows composition happening over time — a paragraph appearing, being cut, coming back reworded. Generated text does not have that shape; it arrives whole.

In Microsoft Word, check File → Info → Version History if the file lives in OneDrive or SharePoint. Locally saved Word files may have autosave versions but they are far less complete.

Screenshot your research trail. Browser history from the days you were working. Library database access logs. The PDFs in your downloads folder with their timestamps. Notes in whatever app you use. Highlighted passages in a Kindle book. These establish that you were engaged with the sources you cited, at the times you claim.

Collect your drafts. Emailed drafts, messages where you sent a paragraph to a friend, a comment thread with a supervisor, the outline you scribbled. Anything time-stamped and prior to submission.

Write down your process while it is fresh. Where you worked, on what days, in what order you tackled the sections, which part you found hardest and why, what you cut. Specificity here is what separates a real account from a constructed one, and you will not remember it as clearly in three weeks.

Understand your institution's actual policy

Before you respond, find the written policy. Most institutions have one, and most people accused have never read it. Look specifically for:

  • Whether a detector score alone is sufficient evidence. Many policies explicitly say it is not — that it may only be used to prompt a conversation. Turnitin's own guidance to institutions says its AI indicator should not be the sole basis for an accusation. If your institution's policy echoes that, quote it.
  • The burden of proof. In most academic misconduct frameworks, the institution must demonstrate misconduct occurred. You are not required to prove innocence, though in practice showing your process is what resolves these fastest.
  • The appeal route and its deadline. These deadlines are often short and strictly enforced. Note the date.
  • Whether you may bring someone. Student unions frequently provide trained advocates for exactly this meeting, free. Use them. They have sat through dozens of these and you have sat through none.

How to respond

Tone does more work here than most people expect. Two responses containing identical facts land completely differently depending on how they are framed.

What works: calm, specific, cooperative, and focused on evidence. You want to be the person who makes this easy to resolve, not the person who makes it a fight.

What backfires: leading with an attack on the detector's accuracy. Even though the accuracy criticism is correct, opening with it reads as deflection. Establish your own evidence first, then address the tool's reliability as a secondary point.

A workable structure for a written response:

Thank you for raising this. I wrote this paper myself and I'd like to show you how.

I drafted it in Google Docs between the 3rd and the 11th. The version history is attached — you can see the introduction was rewritten three times, and the section on X was originally the conclusion before I moved it. My notes and the annotated sources I used are also attached.

I'm happy to discuss the argument in person, including the parts I struggled with. I'd also welcome writing a comparable piece under supervision if that would help.

I understand the detector returned a high score. My understanding is that these tools measure how statistically predictable writing is rather than how it was produced, and that plain, structured prose scores highly for that reason. I'd be grateful if the score could be considered alongside the process evidence above.

That last paragraph does the reliability argument without making it the centrepiece.

Offer a verbal defence

This is the strongest card most people hold, and it is underused. Offer to discuss the work in detail, unprepared, in person.

Someone who wrote a piece can tell you why the third section exists, what they cut and why, which source they found least convincing, and what they would change. Someone who generated it cannot do any of that convincingly, because there was never a decision to remember. Most experienced assessors know this, and a fifteen-minute conversation frequently ends the matter.

Prepare by rereading your own work and reconstructing your reasoning. Not memorising it — you want to be able to think about it out loud, including the parts you are unsure of. Being willing to say "honestly that paragraph is weak, I ran out of time" is more convincing than a flawless recitation.

If you did use AI assistance

Partial use is where most real cases sit, and honesty is the only strategy that survives contact with scrutiny.

If you used a tool to fix your grammar, restructure a paragraph, or brainstorm an outline, say so plainly and describe exactly what it did. Many policies permit some of this, particularly for non-native speakers using language support. A precise account of limited use is defensible. A denial that later collapses is not, and the cover-up is treated far more harshly than the original conduct at every institution I am aware of.

Draw the line clearly for yourself: assistance with expression is a different thing from substitution of authorship. Our position on that distinction is set out in AI writing tools and academic integrity.

The reliability argument, in detail

When you do need to make the case that the score is unreliable, these are the points that hold up.

False positive rates are not zero, and the base rate makes them worse. Even a detector with a genuine 1% false positive rate, run across a cohort of 500 submissions, will falsely flag around five innocent students. Scale that across a semester and the number of wrongly accused people is not small. Vendor accuracy claims are also typically measured on clean benchmark datasets that look nothing like real student writing.

Non-native English speakers are systematically disadvantaged. Published research has found detectors flag writing by non-native speakers at dramatically higher rates than native speakers' — the pattern is consistent enough that some institutions have restricted detector use on those grounds. If English is not your first language, this is directly relevant and worth raising. We go into the mechanism in why AI detectors fail non-native English speakers.

Detectors disagree with each other. Run the same passage through three detectors and you will frequently get three different verdicts. If you can demonstrate that — screenshots of the same text scoring 95% on one tool and 8% on another — it undermines the premise that any of them is measuring something real.

Established human text gets flagged. People have fed detectors the US Constitution, the King James Bible, and Shakespeare, and received high AI-probability scores. This makes the point vividly, but use it sparingly. It can read as a gotcha rather than an argument, and an assessor who feels mocked digs in.

Detectors cannot cite evidence. A plagiarism checker points to a source document. An AI detector points to nothing — it cannot show you the thing your work supposedly came from, because no such thing exists. That asymmetry is worth naming.

Practical protection going forward

Not fair that this is necessary, but it is cheap insurance.

  • Draft in something with version history. Google Docs by default. The history is automatic and it has resolved more of these cases than any other single artefact.
  • Don't compose in ChatGPT's interface, even for your own words. Pasting your own draft in for feedback creates a record that looks bad out of context.
  • Keep your notes. Messy, dated, handwritten or otherwise.
  • Avoid paste-bombing. Writing a whole essay elsewhere and pasting it into the submission document in one action produces a version history that looks exactly like generated text arriving whole. Paste in sections as you go, or draft in place.
  • Save the untidy stages. The version with the terrible opening paragraph is evidence.

What not to do

Don't run your work through a "humanizer" to lower the score after an accusation. It changes the text after the fact, which is the worst possible look, and if the submitted version differs from what you defend, you have created a new problem. Rewriting tools — including ours — are for improving how a draft reads before you submit it, not for laundering a score afterward. We are explicit about this in our editorial standards.

Don't ignore the deadline. Appeal windows are short and lapsed appeals are rarely reopened.

Don't go it alone if the stakes are real. If the potential outcome includes a failed module, a mark on your record, or expulsion, get an advocate. Student unions, ombudsman offices, and in serious cases education solicitors exist for this.

Don't escalate on the first email. Most of these resolve at the lowest level when the accused person turns up with evidence and a reasonable manner. Save formal escalation for when informal resolution has actually failed.

The wider picture

The uncomfortable truth is that reliable detection of AI text is not currently possible, and there are good theoretical reasons to think it will not become possible. Institutions have adopted detectors anyway, because the alternative — rethinking assessment — is expensive and slow.

That leaves individuals absorbing the cost of a tool that does not work as advertised. Until policy catches up, the practical defence is the same as it has always been: keep your working, show your process, and be able to talk about your own ideas.

Most people who are falsely accused and respond with evidence and composure are fine. It is genuinely stressful and genuinely survivable, and the score is a lot weaker than the person quoting it believes.

Common questions

Can an AI detector prove I cheated?
No. AI detectors do not detect authorship and cannot produce the source your work supposedly came from, the way a plagiarism checker points to a matching document. They estimate how statistically predictable your word choices are, which is a proxy rather than evidence. Turnitin's own guidance to institutions states its AI indicator should not be the sole basis for an academic misconduct finding, so a score alone is a reason to ask a question rather than to reach a conclusion.
What evidence proves I wrote something myself?
Version history is the strongest, because it shows the document being built over time rather than arriving whole. Google Docs keeps this automatically; Microsoft Word only does when the file is saved to OneDrive or SharePoint. Beyond that: dated research notes, earlier drafts, browser and library history from the days you worked, and messages where you sent a paragraph to someone. Being able to discuss your own argument in detail is often what actually ends the case.
What should I do first if I am accused of using AI?
Stop editing the document. Every save from that point muddies the version history you are about to rely on. Export or screenshot that history before anything else, then find your institution's written policy and note the appeal deadline, which is often between five and fifteen working days. Only then reply, leading with your evidence rather than with criticism of the detector.
Should I rewrite my work to lower the AI score?
No, and this is the most damaging mistake people make. Changing the text after an accusation is indefensible, and if the version you defend differs from the one you submitted you have created a second and worse problem. Rewriting tools, including ours, are for improving a draft before you submit it, never for altering work that has already been flagged.
How long do I have to appeal an AI accusation?
Appeal windows are usually short and strictly enforced, commonly five to fifteen working days from the date of the finding. The exact deadline is in your institution's academic misconduct policy. Find it before you draft any response, because a lapsed appeal is rarely reopened regardless of how strong your evidence is.
Can I bring someone with me to the meeting?
Usually yes, and you should. Student unions frequently provide trained advocates for exactly this meeting at no cost, and they have sat through dozens of these hearings while you have sat through none. Check your institution's policy for who is permitted to accompany you, and arrange it before the meeting rather than asking on the day.
What if I did use AI for part of my work?
Say so plainly, and describe exactly what it did. Many policies permit grammar correction and language support, particularly for non-native speakers. A precise account of limited use is defensible almost everywhere. A denial that later collapses is not, and every institution treats the concealment more harshly than the original conduct.

Keep reading