ChatGPT detector
Paste a few paragraphs and this tool estimates how likely they are to have been written by a language model. It combines a RoBERTa classifier trained to separate human writing from ChatGPT output with stylometric measures: sentence-length variation, vocabulary diversity, repeated openings and the stock phrasing assistants tend to reach for.
Read the result carefully. Text detection is the least reliable form of AI detection there is, and the failure mode matters: these models disproportionately flag writing by non-native English speakers, and by anyone writing in a plain, formal register. A high score is a reason to ask a question, never a reason to accuse.
Drag a file here, or
JPEG, PNG, WebP, AVIF, PDF or .txt · up to 10 MB · you can also paste a screenshot
By uploading you agree to our Terms and Privacy Policy. Your file, IP address, approximate location and device details are stored.
What this checks
- A RoBERTa classifier trained on human-versus-ChatGPT pairs, run across the document in overlapping sections
- Burstiness — humans vary sentence length far more than models do
- Lexical diversity and repeated sentence openings
- Density of assistant-typical phrasing such as 'delve', 'multifaceted' and 'it is important to note'
- Absence of contractions, a weak but measurable register signal
- Disagreement between sections, which often indicates partly edited text
Frequently asked questions
How accurate are ChatGPT detectors?
Research consistently puts them well below the accuracy people assume, and OpenAI withdrew its own classifier in 2023 for exactly this reason. Expect a useful signal on long, unedited samples and close to noise on short or lightly rewritten ones.
Will it flag my writing if English is not my first language?
It might, and that is a documented bias rather than a quirk of this tool. Detectors key on simpler sentence structures and a narrower vocabulary, which correlate with non-native writing as much as with generated text. Weigh the result accordingly.
Can I check work submitted by a student?
You can run the check, but you should not treat the output as evidence. No detector on the market is accurate enough to support an academic-misconduct finding on its own, and false positives land hardest on the students least able to contest them.