Accuracy
How accurate is it?
Our own test, in full.
We tested the AI detector on 123 passages where we knew the answer: 88 written by people, 26 written by AI tools, and 9 written by AI and then edited by a person. These are the results with the verdict rules the site uses today.
| What it was | Passages | What GPT Nada said |
|---|---|---|
| Written by people | 88 | 87 read as human. 1 was called probably machine-written. None was called very likely machine-written. |
| Written by AI, untouched | 26 | All 26 were called very likely machine-written. |
| Written by AI, then edited by a person | 9 | 4 very likely machine-written, 1 probably machine-written, 2 unsure, 2 read as human. |
What we tested
- The human passages are published writing from before AI chatbots existed: 39 Wikipedia extracts, 25 research abstracts and 24 passages from classic novels.
- The AI passages were written for the test by Claude, an AI model, some of them in the style ChatGPT tends to write: essays, history and science answers, a personal statement, a cover letter, reviews and reports.
- The edited passages are AI text that a person then changed by hand: contractions, slips, a personal line, a different ending.
- Each passage was checked on its own, the same way the site checks your text: the quick check, then both detectors on our own computer.
What this test cannot tell you
- We set our thresholds using some of these same passages, so results on new writing will be a little worse than this.
- 123 passages is a small test. It tells you how the detector behaves, not a precise error rate.
- Edited AI text is the hard case. A person who rewrites enough of it can make it read as human, to us and to every detector.
- Our human passages are published writing, not student essays. Careful, formal writing by people is the kind most likely to be wrongly flagged.
- We have not yet tested text taken straight from ChatGPT, Gemini or Copilot. That is our next test.
- It is tested in English only.
A result from GPT Nada is a strong hint, never proof. It should never be used on its own to accuse anyone of using AI.
Check my textFree. No account. Your text is wiped when the check finishes.