Which models it knows
Writing from today's models has habits that a trained detector can pick up: very even rhythm, safe word choices, tidy transitions. We test our Engine against texts from twelve current models and report how often each one is caught on the accuracy page. When a model is harder to catch, the table shows it.
How to read the result
- Likely AI means the score is higher than all but about 1 in 100 human texts in our tests.
- Unclear is common and honest: the text could go either way. Look at the highlights.
- Likely human means the score is lower than all but about 1 in 100 AI texts in our tests.
A Verdict is evidence to weigh, not proof. More in How it works.
Questions
Does it detect ChatGPT, Claude and Gemini?
We test against texts from twelve current models by nine companies, including OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek and Qwen. Results per model are on our accuracy page; some newer models are harder to catch than others.
What if the AI text was edited or paraphrased?
Light editing usually leaves the signal. Heavy rewriting weakens it. We measure paraphrased AI text separately and publish that rate too.
Can I check text in Spanish?
Yes. English and Spanish each have their own thresholds, measured on texts in that language.