From 5 October API customers can switch the mark on for selected models; it is off by default. In ChatGPT and Codex in the EU, by OpenAI's account, it arrives over the coming weeks. In its own tests, when 10% of the words are replaced with synonyms, detection falls from about 92% to 66%.
In short- 5 October 2026: OpenAI announced a text watermark (textGrain) in response to the EU AI Act. In the API it is opt-in for selected models; in ChatGPT and Codex in the EU only, over the coming weeks.
- According to OpenAI, at a 1% false positive rate: about 80% detection on a 200-token passage and about 95% on 400, for psychology; substantially lower for mathematics.
- In the same tests, replacing 10% of the words with synonyms drops detection from about 92% to 66%, at 25% to 17%. The detector is not public for now: approved researchers and expert organizations only.
Replace every tenth word with a synonym and OpenAI's mark is found in about 66% of cases instead of about 92%. Replace every fourth and about 17% remain. The numbers are OpenAI's own, for passages of 400 tokens.
The facts: on 5 October 2026 OpenAI published its approach to text watermarking in response to the EU
AI Act. From the same day
API customers worldwide can opt in to a
watermark for selected models; it stays off by default. Over the coming weeks OpenAI will add an invisible watermark to eligible
ChatGPT and
Codex text output on all plans, in the European Union only; there is no global default at launch. The technology, textGrain, adds an invisible statistical signal to
the model's word choices, and the detector looks for that signal. According to OpenAI, textGrain matched or exceeded the other approaches tested, including SynthID for text. At a target false positive rate of 1%, the detector finds the mark in about 80% of 200-token passages and about 95% of 400-token ones, for content such as psychology; for mathematics, where word choice is narrower, detection is substantially lower. Replacing 10% of the words with synonyms drops detection in a 400-token passage from about 92% to 66%, and 25% to 17%. According to OpenAI the mark does not meaningfully change the results of the Astra model on its
benchmarks, for example 49.57 points without the mark and 49.76 with it on the Artificial Analysis Intelligence Index. The detector is not public: researchers and expert organizations can apply, and access goes to approved ones, initially case by case, in accordance with the Code of Practice on Transparency of AI-generated Content. All of this applies to text only; the tools for images and audio, openai.com/verify and the Content Provenance API, stay public. Article 50(2) of Regulation (EU) 2024/1689 requires providers of AI systems that generate synthetic audio, image, video or text to ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated; the technical solutions must be effective, interoperable, robust and reliable as far as this is technically feasible. Article 50 applies from 2 August 2026 (Article 113). For systems placed on the market before that date, the deadline for paragraph 2 is 2 December 2026 (Article 111(4), added by Regulation (EU) 2026/1744).
A wristband at the door gets checked in a second, and nobody rewrites it with synonyms. Text gets rewritten.
The absence of a mark does not prove a person wrote the text.
That sentence is OpenAI's, and it is the most valuable thing on the page. The mark does not measure how much a person put in, does not prove who owns it, does not point to the user, and does not say whether the text is true. At most, it can show that an OpenAI system generated or processed part of the passage.
The law wants technical solutions that are robust and reliable "as far as this is technically feasible". How feasible that is, OpenAI shows with its own numbers: about 95% for a 400-token passage in psychology, substantially less for mathematics and for edited text. Whether that is enough for Article 50, we do not decide.
In the API the switch is yours, and it is off by default. Before you promise a client that your text is "marked", read OpenAI's section on what the mark does not tell you.
The visual is generated code art. No third-party images.