The detector challenge
Can you fool GPTTrace?
Take AI-written text, rewrite it, humanize it, edit it by hand — whatever you like — and see whether GPTTrace still spots it. Texts that get past us become training data, so the detector keeps getting harder to beat.
Why we built this
AI “humanizers” and paraphrasers are an arms race. The best way to stay ahead is to train on the texts that actually beat us, so every submission that gets through is added to GPTTrace's training and test data, and the next version is measured on whether it catches them.
This page is the one exception to GPTTrace's usual rule that what you check is never stored. Everywhere else on the site, text is checked in your browser or in memory and then discarded.
Frequently asked questions
What happens to the text I submit?
Unlike every other check on GPTTrace, challenge submissions are stored, because that is the point: texts that fool the detector are used to train and test future versions. We store the text, the AI you say wrote it, what you did to it, and the score. We do not store your IP address, a cookie or an account with it. Nothing is stored unless you tick the consent box.
What should I submit?
Text that was generated by an AI and then changed in any way you like: an AI humanizer, a paraphraser, prompting the AI to write like a person, translating it back and forth, or editing it yourself. Please don't submit anything personal or confidential, and don't submit text you wrote entirely yourself — the challenge is about AI text.
Does a “fooled” result mean the text is human?
No. It means this version of GPTTrace scored it below its “Likely AI” line. Detectors give evidence, not proof, and every detector can be beaten by enough rewriting. That's exactly why we collect the texts that get through.
Which model scores the challenge?
The same large model that powers the GPTTrace API (Ettin-1B, version v8s, open weights on Hugging Face), with the calibrated threshold described on our benchmark page.