What textGrain does
OpenAI describes textGrain as an invisible statistical signal added to the model’s word choices. A detector looks for that signal to assess whether a passage contains an OpenAI watermark. Nothing is added to the text as characters, so stripping hidden Unicode does nothing to it.
Rollout is staged. ChatGPT and Codex users in the EU on every plan get it over the coming weeks, API customers anywhere can opt in for select models from 5 October (it stays off by default), and OpenAI is working with cloud partners to offer it for model outputs accessed through them. OpenAI also says it plans to release the technology as open source.
The published detection figures
At a target false positive rate of 1%, OpenAI reports its detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. Detection was substantially lower for mathematics, where there is less freedom in word choice.
For edits, tested on 400-token passages: replacing 10% of words with synonyms reduced detection from about 92% to 66%, and replacing 25% reduced it to 17%. OpenAI’s own conclusion is that editing can substantially weaken the signal.
Read together: length helps the detector, constrained content and edits hurt it, and the relationship with edits is steep rather than gradual. Those figures come from OpenAI’s ideal-conditions evaluation, and OpenAI itself says strong performance there does not guarantee reliable detection in everyday use.
Who can run the detector
Approved researchers and expert organisations can apply for access, granted case by case under the EU Code of Practice. It reports whether it finds an OpenAI watermark without identifying the user or revealing prompts. OpenAI is not making it public at launch because of the risk of missed watermarks and false positives.
That means there is no public textGrain key, and a tool claiming a textGrain verdict has either been approved or is guessing. WatermarkRemoverPro does not hold the key, never reports a textGrain result, and lists the keys it did test on every analysis.
What textGrain does not tell you
OpenAI lists the limits itself: a watermark does not measure how much a person contributed, establish ownership or responsibility, identify the user, or verify accuracy. A missing watermark does not prove human authorship, since the text may be short, edited, translated, from an unsupported model, older than the rollout, or from another company’s system.
For how this compares with Anthropic’s mark see /guide/chatgpt-vs-claude-watermark, and for the method class they share see /guide/does-editing-remove-a-watermark.
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.