OpenAI to watermark ChatGPT and Codex text in the EU, detector restricted

OpenAI is starting to mark the text its models write. In a post published October 5, it describes a text watermark called textGrain, which it will roll out in ChatGPT and Codex in the European Union, offer as an opt-in in its API, and test through a restricted detector. OpenAI ties the move to the EU AI Act, which it says requires generative AI providers to make generated text identifiable in a machine-readable way. This article covers what is changing, how well the watermark works by OpenAI’s own numbers, and what it cannot tell you. We read the post and OpenAI’s technical report; we did not test the watermark or detector.

What is OpenAI changing about text watermarking?

Where	What happens	When
API, worldwide	Customers can opt in to watermarked text for select models. Off by default.	From October 5
ChatGPT and Codex, EU	An invisible watermark is added to eligible text output, on all plans. Not a global default.	“Over the coming weeks”
Cloud partners	OpenAI says it is working to offer watermarking for its models used through their services.	“In the coming weeks”
Detector	Researchers and expert organizations can apply. Access is granted case by case. Not public.	Applications open from October 5

OpenAI’s image and audio tools are separate and stay public. Images carry Content Credentials (OpenAI says it is C2PA conformant) and, in supported images and audio, invisible SynthID watermarks, and can be checked on openai.com/verify or through its Content Provenance API.

How does OpenAI’s textGrain watermark work?

OpenAI says textGrain adds an invisible statistical signal to the model’s word choices, and the detector looks for that signal. The technical report, whose authors include OpenAI staff and academics from the University of Pennsylvania and Yale, gives the mathematics. In plain terms: the model’s choice of each next word is tied to a secret key through an optimal-transport calculation, with a setting that limits how much sampling randomness the watermark may use up, and the detector needs only the text and the secret key. OpenAI says the report will be updated and that it plans to release the technology as open source.

How well does the textGrain detector work?

OpenAI tested on watermarked English answers from the ELI5 question set, at a target false-positive rate of 1%.

Test	Result
200-token passages, psychology-type content	about 80% detected
400-token passages, psychology-type content	about 95% detected
400-token passages, 10% of words replaced with synonyms	detection down from about 92% to 66%
400-token passages, 25% of words replaced	down to 17%

Two limits follow. Shorter or more constrained text is harder: OpenAI says detection was substantially lower for content like mathematics, where there is less freedom in word choice. And editing erodes the signal quickly. OpenAI itself says strong results under ideal conditions do not guarantee reliable detection in everyday use, and gives that as a reason to keep the detector restricted. It also says text may be too short, edited or translated to detect.

Does watermarking hurt model quality?

OpenAI reports no meaningful performance difference for Astra, its latest frontier model, with and without the watermark. Its table covers eight benchmarks, with the runs labelled “Astra, max”; for example 49.57 vs 49.76 on the Artificial Analysis Intelligence Index, 72.80% vs 71.68% on DeepSWE v1.1 and 94.44% vs 93.94% on GPQA Diamond (unwatermarked vs watermarked). Some scores went up and some down. The post gives no run-to-run variation, so we cannot tell how much of the gap is noise. It also says textGrain matched or exceeded other methods it tested, including SynthID for text; that comparison is OpenAI’s own and we did not find it in the technical report.

What can’t a text watermark tell you?

OpenAI lists the limits itself. A watermark does not measure human contribution, establish ownership or responsibility, identify the user, or verify accuracy. And the absence of a watermark does not prove human authorship, since the text may be short, edited, translated, from an unsupported model, from before watermarking, or from another company’s tools.

What to keep in mind

These are our observations:

• It is regional. The default watermark applies to EU users of ChatGPT and Codex. Readers elsewhere, including in Turkey, would only see it if an API customer opts in, according to OpenAI’s description.
• It is not an AI detector for teachers or editors. The detector is not public, only reports an OpenAI watermark, and a 25% rewrite nearly erases the signal, so it cannot settle whether a person wrote a text.
• Other vendors are not covered. It marks only OpenAI’s output, so a clean result says nothing about text from other tools.
• The numbers are OpenAI’s. The detection rates, quality comparison and SynthID comparison come from OpenAI, and outside researchers are exactly the group it is inviting to test them.

Editorial review: October 5, 2026 (AI). Prepared from OpenAI’s post “Our approach to EU text provenance rules” (October 5, read in full through a browser) and the accompanying textGrain technical report (October 5, text extracted and read). We did not test the watermark, the detector or the models. All detection, quality and comparison figures are OpenAI’s. The post’s two interactive charts are not reproduced; their figures are given in the text. The cover is OpenAI’s own card, reproduced with credit and a link to the source. “What to keep in mind” is ours.