How it works
The text is split into parts of about 300 words, and each part is checked by TMR, an open-source RoBERTa model trained to tell human from machine writing. On the independent RAID benchmark (672,000 texts from many AI models, including disguised ones) it scores an AUROC of 99.3%, with about 2.6% of human texts wrongly flagged.
The overall score is the average of the parts, weighted by their length, so you can see whether a whole text or only some paragraphs look AI-written.
Read the result carefully
Teachers and editors: use it to decide what to look at more closely, then talk to the writer - never accuse anyone on a score alone.
- No detector is perfect - treat the score as a signal, not proof.
- Very formal, edited, translated or template-like human writing can be flagged.
- AI text that was heavily rewritten can pass as human.
- The model is trained on English; other languages give unreliable scores.
Questions
Is this AI checker free?
Yes - unlimited checks of up to 5,000 words, no sign-up. The model runs in your browser.
Is my text uploaded or stored?
No. The model is downloaded to your browser once (about 126 MB) and checks the text on your device.
Which AI models can it detect?
It was trained on text from many models, such as GPT, Llama, Mistral and others. Newer models can be harder to detect.