OpenAI to watermark ChatGPT text in EU, offers API opt-in worldwide

OpenAI says it will add an invisible watermark to eligible ChatGPT and Codex text in the European Union over the coming weeks, and that API customers anywhere can now opt in to the same mark. The company says it is responding to the EU AI Act. It is not making text watermarking a global default.

The plan is in an Oct. 5 OpenAI post. “We are not making text watermarking a global default at launch,” the company writes.

The rule in English

The EU AI Act tells companies that make AI systems that generate text, images, audio or video to mark that output so a machine can detect it as AI-made. Article 50(2) says providers “shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.” Providers’ technical solutions must be effective, interoperable, robust and reliable “as far as this is technically feasible.” Under Article 113, the article applies from Aug. 2, 2026.

OpenAI’s post does not cite the article by number. It says the act “requires generative AI providers to make generated text identifiable in a machine-readable way.”

What OpenAI is doing

• ChatGPT and Codex in the EU: Eligible text output gets an invisible watermark “across all plans in the EU only,” starting over the coming weeks.
• API, worldwide: Starting Oct. 5, API customers “around the world” can opt in to watermarked text for select models. Watermarking stays off by default in the API. OpenAI says it is working with cloud partners to offer the same option in the coming weeks.
• Detector: Approved researchers and expert organizations can apply for access. OpenAI says it is not making the detector public at launch “given the risk of missed watermarks and false positives.”

The watermark, which OpenAI calls textGrain, adds “an invisible statistical signal to the model’s word choices.” The company says it plans to release the technology as open source.

How well it works, by OpenAI’s numbers

OpenAI says textGrain matched or beat other methods it tested, including Google’s SynthID for text. It also lists limits:

• At a 1% false-positive target, the detector caught about 80% of watermarked 200-token passages and about 95% of 400-token passages on psychology questions. Math text scored “substantially lower.”
• Swapping 10% of words for synonyms in 400-token passages cut detection from about 92% to 66%. Swapping 25% cut it to 17%.
• On benchmarks for its Astra model, OpenAI says it saw no meaningful quality difference with and without the watermark.

Those are company tests. OpenAI cites no outside check of them.

Who it binds, and who it exempts

The AI Act duty falls on providers, which here means OpenAI. OpenAI’s choice binds eligible EU users of ChatGPT and Codex; the post describes no opt-out. It does not bind ChatGPT users outside the EU. API developers decide for themselves; OpenAI says opt-in lets customers “decide how watermarking fits their transparency obligations.” Image and audio checks through openai.com/verify stay public.

What a parent, teacher or founder should do differently

• Parents and teachers: A missing watermark is not proof a person wrote something. OpenAI says text can be “too short, edited, or translated for detection to work reliably.” Schools also can’t run the detector; it is not public.
• Founders building on the API: The mark is off unless you turn it on. If you serve EU users, read Article 50 yourself; OpenAI is leaving the API call to customers.
• Workers and writers in the EU: OpenAI says a mark “does not identify the user” and does not measure how much a person edited the text.

What’s next

• The date ChatGPT and Codex marks actually reach EU users.
• OpenAI’s promised update to its textGrain technical report and the open-source release.
• Whether outside researchers who get detector access confirm OpenAI’s detection rates.

— Simone Adler