Where the ChatGPT watermark applies
OpenAI says the mark rolls out over the coming weeks to eligible ChatGPT and Codex users on all plans, in the EU only. Developers using the API anywhere can turn it on for select models from the announcement date, and it is off by default. OpenAI is not making text watermarking a global default at launch.
So the answer to “does ChatGPT watermark text?” is currently “in the EU, yes, and elsewhere only if an API developer opts in”. That is different from Anthropic, which said it is applying its Claude watermark worldwide.
What textGrain is
textGrain is OpenAI’s name for its method, described in a technical report written with researchers from the University of Pennsylvania and Yale. A secret key is used to sort the model’s next-word predictions, and the model is nudged toward one side of that ordering. Across hundreds of words the nudges form a pattern a detector can find using only the text and the key.
It belongs to the same family as the green-list scheme in Kirchenbauer et al. (2023), explained in our guide to the Claude AI watermark. The practical properties follow from that design: the mark is in the words, so it survives copy and paste; it is invisible without the key; and it says nothing about who typed the prompt.
How well it survives editing
OpenAI’s own test found that replacing 10% of words with synonyms reduced detection from about 92% to 66%. It also said short passages, math answers and translated text are harder to detect.
This is why OpenAI warns that a missing watermark does not prove human authorship: the text may be too short, too heavily edited, or from another company’s system. For the full set of published figures see /guide/textgrain, and for what weakens the mark in practice see /guide/remove-chatgpt-watermark; the mechanism is covered in /guide/does-editing-remove-a-watermark.
Who can detect it, and what a checker can honestly say
At launch, OpenAI is giving detector access only to approved researchers and expert organisations. There is no public key, so no third-party tool can test for textGrain. This product does not hold it and does not pretend to; every result lists the keys it actually tested.
What it can do is run its own named statistical tests, report a separate key-free AI-style likelihood, and, if you rewrite a passage, report what share of your original words changed, the same unit OpenAI uses when it describes robustness. That figure describes your edit, not OpenAI’s detector, and the result says so.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. WatermarkRemoverPro's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.