Signal Check Research Prototype

Accused of using AI when you didn't

What to do, in order, if a detector score has been used against your work.

Last reviewed 16 August 2026

Being told that software has concluded you cheated is a particular kind of awful, partly because the accusation arrives with a number attached and numbers feel like facts. The first thing worth knowing is that these numbers are wrong often enough that entire universities have stopped using them. You are not in an unusual situation, and you are not without options.

This guide is practical rather than reassuring. It assumes you did not do the thing you are accused of.

This is general information, not legal advice. If the consequences are serious — expulsion, a withdrawn degree, loss of a job or visa status — get advice from someone qualified in your jurisdiction, and do it early rather than after a hearing.

First, before you reply to anything

Preserve your evidence immediately

The single most useful thing you have is the record of how the work came to exist, and some of it can disappear or be overwritten. Before you do anything else:

Version history. Google Docs keeps a full revision timeline under File → Version history. Microsoft Word has AutoSave version history in OneDrive/SharePoint. Do not keep editing the document — that adds revisions and muddies the timeline. Export or screenshot the history now.
Drafts and notes. Earlier files, outlines, annotated PDFs, photographed handwritten notes, reading lists, mind maps.
Messages about the work. Texts to a friend complaining about the essay, emails to a tutor, forum posts, library requests. Timestamps that bracket the writing period are useful.
Browser and library history for the sources you used.
The accusation itself. Screenshot the report, the score, the tool name, and the exact passages highlighted.

Do not panic-confess to something smaller

Under pressure, people often reach for a partial admission — "I only used it to fix grammar" — hoping it defuses things. If it is not true, do not say it. Under most academic integrity policies a partial admission ends the inquiry and converts a contestable allegation into a recorded finding.

Ask for the specifics in writing

You are entitled to know what is actually being alleged. Ask, by email so there is a record:

Which tool produced the result, and what the score was.
Which specific passages were flagged.
What the institution's policy says about detector output as evidence.
Whether anything other than the detector score supports the allegation.
What the process is, and what your rights are within it.

That last question about "anything other than the score" matters more than it looks. Often the honest answer is no — and an allegation resting on a single automated number is much weaker than one supported by, say, a marked change in voice the marker noticed independently.

What the research says, and how to use it

You do not need to argue that detectors are useless in general. You need to establish that a score is not sufficient evidence on its own. The published record does that for you.

61.22%

of TOEFL essays by non-native English writers were misclassified as AI-generated across seven detectors (Stanford, 2023)

26%

of AI text correctly identified by OpenAI's own classifier — which it withdrew for low accuracy

~750

papers per year that a 1% false-positive rate would wrongly flag at a single university

Three points that tend to land, in roughly this order of usefulness:

The vendor of the underlying technology could not make this work. OpenAI released a detector for its own models in January 2023 and withdrew it that July, citing low accuracy. It caught 26% of AI text and falsely flagged 9% of human text.
Institutions have disabled these tools over exactly this concern. Vanderbilt University turned off Turnitin's AI detector in August 2023, noting that the vendor's advertised 1% false-positive rate would have meant roughly 750 wrongly flagged papers across the 75,000 they submitted in 2022.
The error is not evenly distributed. If you write English as a second language, the Stanford study is directly relevant: 61.22% average false-positive rate on TOEFL essays versus 5.19% on essays by US eighth-graders. Nearly one in five TOEFL essays was flagged by all seven detectors simultaneously.

Full citations are at the bottom of this page. Bring the actual sources, not a summary of them — a link to a peer-reviewed paper in Patterns and a university's own published guidance carries weight that a claim in your own words does not.

Making your case

Lead with the record, not the argument

The strongest thing you have is not the research. It is your version history. A revision timeline showing a document accumulating over days, with false starts, restructuring and abandoned paragraphs, is very hard to explain away. Lead with that and use the research to explain why the score should not outweigh it.

Offer to talk about the content

Volunteer to discuss the work: why you structured it that way, what you cut, which source changed your mind, what you would do differently. Someone who wrote a thing can do this indefinitely. It is a far better test than any classifier, and offering it unprompted signals confidence.

Be specific about your own writing habits

If you write in a deliberately plain, organised style — or you were taught to write formally, or you use a grammar checker, or you drafted in another language and translated — say so plainly. These are the exact characteristics that raise detector scores, and naming them turns "suspicious evenness" into an ordinary explanation.

Keep the tone level

The person handling this is often not the person who chose the tool, and may privately share your doubts about it. Anger at the institution rarely helps in the room. "I understand why the report looks concerning; here is the version history, and here is why that tool produces this result on writing like mine" works better than a challenge to their competence.

If it escalates

Find your student advocate. Most institutions have a students' union caseworker, ombudsperson, or academic integrity adviser whose job is representing you. They have seen these cases before and know the local process.
Read the actual policy. Get the written academic integrity procedure. Check what standard of proof applies, whether detector output is admissible, and what the appeal route is. Procedural mistakes are common and are often the fastest route to a reversal.
Put everything in writing. After any meeting, email a short summary of what was said and agreed. This creates a record and corrects misremembering early.
Ask about accommodation policy if a disability, language support, or writing centre involvement is relevant to how your writing looks.

What not to do

Do not run your work through more detectors hoping for a clean result. They disagree with each other constantly, and a second unfavourable score handed over voluntarily becomes evidence against you.
Do not rewrite the work to score lower. It looks like tampering and does nothing about the allegation.
Do not use a "humanizer" tool. Using one on writing you produced yourself is genuinely difficult to explain afterwards.
Do not ignore it. Missed deadlines in an integrity process are usually treated as concession.

On this site's own tool: the analyzer on the home page has no validated accuracy rate and should not be used as evidence for or against anyone — including yourself. A low score from it proves nothing and will not help your case. We say so on every page for the same reason this guide exists.

Sources

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7). arXiv:2304.02819
OpenAI. New AI classifier for indicating AI-written text (January 2023; discontinued 20 July 2023). openai.com
Coley, M. Guidance on AI detection and why we're disabling Turnitin's AI detector. Vanderbilt University Brightspace, 16 August 2023. vanderbilt.edu