Falsely accused of using AI? What the detector data shows
You are not alone, and the data is on your side. In our test of seven free AI detectors, 117 of 126 human-written texts, all published before ChatGPT existed, were flagged as AI by at least one detector. Even the most careful detector flagged 6.3% of human texts. A detector score is not proof, and your writing process is the strongest evidence you have.
Tested
Detectors flag human writing all the time
Our test used 126 human texts from Wikipedia, Wikinews, Stack Exchange, arXiv and Project Gutenberg, all published before 30 November 2022, the day ChatGPT came out. 117 of them were flagged as AI by at least one of seven free detectors. The detectors also flagged famous texts: all seven called Genesis 1 AI-generated, and six flagged the US Constitution.
So a flag on your work says only that your writing resembles what one detector learned to call AI.
What to gather
Collect anything that shows your work taking shape. Most writing tools keep a version history with timestamps. Drafts, outlines, handwritten notes, the sources you read and messages where you discussed the topic all help. If you can, explain your choices in the text in conversation; people who wrote something can usually talk about why.
What to ask
Ask which detector was used and what its published false positive rate is. Point out that detectors disagree with each other: in our test, the share of human texts flagged ranged from 4.8% for the OpenAI RoBERTa detector to 83.3% for TMR. Our false positives page and accuracy guide give numbers you can cite, and every score is on our data page.
Formatting can change the result
Some detectors react to formatting. TMR flagged every multi-paragraph human text in our test, and far fewer when the same words were run together as one paragraph, as our paragraph break test shows. If the result depended on how your text was pasted in, that is worth raising.
Questions people ask
Can I prove I did not use AI?
Your best evidence is the writing process, such as document version history, drafts, notes, sources and browser history from your research. Detectors themselves cannot prove authorship either way.
Why would my own writing be flagged as AI?
Detectors flag text that looks formulaic or predictable. In our test they flagged human academic abstracts, news reports and even famous texts such as the US Constitution, which six of seven detectors called AI-generated.
Which detector did the school use, and does it matter?
It matters a lot. In our test the false positive rate ranged from 4.8% to 83.3% depending on the detector. Ask which tool was used and what its published error rate is.