On October 5, 2026, OpenAI announced a phased plan to add its textGrain watermark to eligible ChatGPT and Codex text output in the European Union. The consumer rollout is for the EU; API customers worldwide have a separate opt-in for selected models, with watermarking off by default.
OpenAI’s phased EU rollout
The announced consumer feature covers eligible text output across ChatGPT and Codex plans in the EU. OpenAI said the rollout would proceed in phases over the weeks following its October 5 announcement. It described the consumer rollout as EU-only at launch, not a global default.
OpenAI also opened applications for access to its text detector. At launch, access was on a case-by-case basis for approved researchers and expert organizations, rather than a public checker.
How textGrain works
OpenAI describes textGrain as an invisible statistical signal embedded in a model’s word choices. It does not add a visible label or a string of hidden characters. Copying the same wording does not itself change the word choices carrying that signal, while replacing words can make detection less likely.
API watermarking is a separate opt-in
The API option has a broader geographic scope than the consumer rollout: customers worldwide can opt in for selected models. The setting is off by default, so the announced API arrangement is not an automatic watermark on every API-generated text.
What OpenAI’s editing evaluation found
In OpenAI’s evaluation of 400-token passages, its detector identified the watermark in about 92% of unedited passages. Detection was about 66% after replacing 10% of the words with synonyms, and 17% after replacing 25%. Each figure came from that evaluation, which used a 1% target false-positive rate; they are not universal accuracy rates for any text or detector.
OpenAI also reported different results by passage length and subject. At the same 1% target false-positive rate, it detected about 80% of 200-token psychology passages and about 95% of 400-token psychology passages. Detection on mathematics passages was substantially lower. Shorter or constrained text can therefore be harder for this method to detect, and the editing results show how synonym changes affected those specific test passages.
What a watermark result can establish
OpenAI says a positive detection does not identify the user, reveal a prompt or conversation, measure how much a person contributed, establish ownership or responsibility, or verify whether the text is accurate. A negative result does not establish human authorship. The signal is a clue about generated text, not a verdict on who wrote it or whether its claims are true.
