There is no public ChatGPT watermark detector. OpenAI's textGrain watermark exists, but only in the EU, and the detector for it is restricted. This site does not read the watermark.
What textGrain is
OpenAI announced textGrain on 5 October 2026. It is an invisible statistical signal built into word choices, so the text reads the same to a person. API customers worldwide can opt in for select models, and it stays off by default in the API.
Where ChatGPT text carries it
- EU. OpenAI will add the watermark to eligible ChatGPT and Codex text in the EU, across all plans. It said this would happen over the coming weeks and published no exact start date. OpenAI's help article ties it to the EU AI Act and the EU Code of Practice on Transparency of AI-Generated Content.
- Rest of the world. OpenAI said, "We are not making text watermarking a global default at launch." It has made no statement about other regions, so ChatGPT text outside the EU should be treated as unmarked.
- API. Customers can turn it on in project or organization settings.
Images and audio work differently. They can be checked publicly through OpenAI's verification page and Content Provenance API. Text is the only type with a restricted detector.
Who can check it
OpenAI has not released the text detector. Applications are open and access is case by case, initially to approved researchers and expert organizations. Qualifying uses are research on provenance and detection reliability. Commercial fraud prevention, reselling, and academic-integrity checking of student work fall outside the program. A teacher cannot get this detector for a class set.
What a watermark cannot prove
OpenAI states that the absence of a detected watermark does not prove human authorship. A watermark also does not identify the user, measure how much a person contributed, establish ownership or verify accuracy. OpenAI's own robustness figures show that rewording weakens the signal.
Checking a text without the watermark
A style-based detector reads how a text is written instead of looking for a key. ChatGPTDetect does this: it reports whether the text reads as AI-written, highlights the passages that carry the signal, and shows how closely the writing matches ChatGPT compared with other families. Veread, Fileverity's text model, wrongly flagged 0.5% of human-written text in seven languages in tests and detected all fully AI-written OpenAI-family texts; see the facts page. Teachers will find more in AI detection for teachers.
Sources: OpenAI, "EU text provenance" (5 October 2026); OpenAI Help Center, "Provenance signals in OpenAI-generated content" (checked 10 November 2026).