How it works
What happens when you send a text, how far you can trust the result, and what happens to your text.
The flow
-
1
Send your text
Paste it on the home page or upload a PDF, DOCX or TXT file. For PDFs we rejoin hyphenated line breaks and strip page numbers and the reference list.
-
2
Choose settings
Under “Analysis settings” you can set the language and choose which analyses run. By default everything that is ready runs.
-
3
Review the marked passages
The text is split into passages and scored. Click one to see its score and explanation.
-
4
Apply or reject suggestions Soon
Revision suggestions will appear here once they are ready.
-
5
Download a report Soon
We're working on downloading results as a PDF report.
Method
Behind it is a multilingual language model called XLM-RoBERTa. We trained it on about 27,000 academic texts: human-written abstracts from DergiPark and arXiv, and texts we had various AI models write or edit.
For each text the model gives three probabilities: human-written, AI-edited or mixed, and fully AI-written. Long texts are split into passages that are scored separately; the overall result also shows what share of the text carries AI-like patterns.
Tested on texts it never saw in training
| AI text caught | False alarm on human text | |
|---|---|---|
| Turkish | 94% | 2.7% |
| English | 66% | 3.0% |
At the “very likely” threshold. The test texts came from an AI model the detector never saw during training; results on real users' writing may differ.
Limits
- The result is a probability. No detector can prove who wrote a text; it shouldn't be used on its own to accuse anyone.
- False alarms are higher on writing by native English speakers. We're working on it.
- Results are less reliable for texts under 150 words and for individual passage scores.
- We can't read scanned (image) PDFs. Two-column journal layouts may come out scrambled.
What happens to your text?
The text and file you send for analysis are processed in the server's memory and discarded once you get the result. We don't write them to disk, we don't log their contents and we don't train the model on them.
The only way we use a text for research is if you send it yourself, with explicit consent, from the “Contribute” page. Details are in the privacy notice. Privacy notice
FAQ
Does it cost anything?
Analysis is currently free and needs no account.
Is the result definitive?
No. Treat it as a hint. That's also why we show which passages were marked instead of a single percentage.
What are the dotted underlines?
Formulaic phrases such as “furthermore”, “it is important to note” or “bunun yanı sıra”. People use them too, so on their own they prove nothing. We show them because they appear very often in AI writing.
Why does it say short texts are less reliable?
The more words the model sees, the more accurate it is. Below 150 words the result is more open to chance.