Falsely accused of using AI: what to do
Being told your own work was written by a machine is upsetting, and the accusation often rests on nothing more than a detector’s score. Here is what that score means, what to ask for, what evidence to gather, and how to be ready next time.
Written · Updated
Stay calm, ask exactly what the accusation rests on, and answer with evidence of how you wrote the work: drafts, notes, sources, any revision history, and your own account of the argument. A detector’s score on its own is not proof; the companies that sell detectors say so themselves. For future work, write in something that keeps a timestamped record of the process, such as OBSCURA Authorship, so the evidence exists before anyone asks.
First: find out what it rests on
Ask, in writing, what the concern is based on. If it is a detector, ask which one, what score it gave, and what policy says a score is grounds for action. Most universities’ and publishers’ procedures expect more than a number, and detection vendors advise against using their scores as the only basis for a finding. OpenAI withdrew its own detector in 2023 because it was not accurate enough, and a 2023 study in Patterns found that detectors, on average, flagged more than half of a set of essays by non-native English writers as AI-written, though people wrote every one. If English is not your first language, or your style is plain and careful, say so: that is precisely the writing detectors misjudge.
Gather what you have
- Drafts and notes, with their dates; an outline; photos of handwritten notes.
- Revision history, where the work was written in a tool that keeps one.
- Sources: what you read, and where you used it.
- Your browser and file history around the time you wrote it, if it helps show the work took the time it would.
- An offer to explain it: talk through your argument, or write a paragraph on the topic under supervision. Many disputes end here.
Next time, keep the record as you write
The strongest answer is a record made while you wrote. Authorship keeps every run of typing, deletion and paste with its time, sealed in a hash chain and, if you turn timestamps on, signed at intervals by an independent time-stamping authority. Nothing is uploaded: the record is in your browser, and you decide who gets it. When you are asked, you hand over a .oauthor file and the person checks it free at /verify-authorship, replaying the writing and seeing every paste marked.
Pastes are fine to have: quotations, your own notes, a passage you moved. The record shows them as what they are, so you can explain them, rather than leaving someone to guess.
What a record can and cannot show
A record shows how and when the work came to exist, and that it was not altered after each timestamp.
What no record can do is prove that no AI was involved. Someone could read text from another screen and type it in by hand, and a determined forger could write a program that types for them. A record makes that expensive, because every pause, slip and revision has to be produced in real time and the timestamps fix when each part existed; it does not make it impossible. Anybody who tells you a tool can prove a negative is selling something.
Questions people ask about Falsely accused of using AI: what to do
What should I do first if I am accused of using AI?
Ask in writing what the accusation rests on: which detector, what score, and what policy treats a score as grounds for action. Then gather your drafts, notes, sources and any revision history, and offer to explain the work.
Is a detector score enough to find me guilty?
It should not be. Detection vendors advise against using a score as the only basis for a finding, and research has found detectors misjudge non-native English writers far more often than others. Ask what other evidence there is.
I pasted quotations and my own notes. Will that count against me?
It should not. In an Authorship record a paste is shown as a paste, with its time and length, and a passage you cut and moved within your own text is recorded as a move. You can explain each one.
How can I avoid this next time?
Write in something that keeps a record of the process from the first word, with independent timestamps, so that the evidence exists before anyone asks. OBSCURA Authorship does this on your own device.
- How to prove you did not use AI to write somethingNo test can prove a text was not written with AI. What you can show is how it was written: drafts, history and a timestamped record of the process.
- A record of your writing process that you keepA tamper-evident, timestamped record of how a piece of writing was made, kept on your device, in a file anyone can check for free.
- A Grammarly Authorship alternative that keeps the record with youGrammarly Authorship, GPTZero's replay and Google Docs history keep your writing record in their accounts.
A letter, now and then
An occasional letter when there is something worth reading: the new guides and what changed in the tools. Never more than once a month and nothing else, and you can leave at any time on the unsubscribe page.