OpenAI adds invisible text watermarks in the EU and opens an API opt-in

What OpenAI is rolling out

According to OpenAI, its approach is phased and reflects both the requirements of the EU AI Act and the limits of the technology. The company says text watermarking and detection remain early technologies with significant limitations, and that views about their benefits and responsible uses are still developing.

• API: Starting today, API customers globally can opt in to text watermarking for select models. It remains off by default. OpenAI says it is also working with cloud partners to offer watermarking for OpenAI model outputs accessed through their services in the coming weeks.
• ChatGPT and Codex: Over the coming weeks, OpenAI will add an invisible watermark to eligible text output for users across all plans in the EU only. It is not making watermarking a global default at launch, and says the regional approach gives it room to learn from real-world use and feedback.
• Detector: Applications for access to the text watermark detector are open. Access is initially limited to approved researchers and expert organizations, granted case by case in accordance with the Code of Practice.

The detector reports whether it finds an OpenAI watermark, without identifying the user or revealing their prompts or conversations. Because of the risk of missed watermarks and false positives, OpenAI is not making it publicly available at launch.

How textGrain works and where it struggles

The technology is called textGrain. It adds an invisible statistical signal to the model's word choices, and the detector looks for that signal to assess whether a passage contains an OpenAI watermark. OpenAI points to a technical report that will be updated in the coming weeks, and says it plans to release the technology as open source. In its evaluations, textGrain matched or exceeded the other approaches it tested, including SynthID for text.

OpenAI stresses that strong performance under ideal conditions does not guarantee reliable detection in everyday use. Its own figures show the gaps:

• At a target false positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages, compared with about 95% of 400-token passages, for content such as psychology. Detection was substantially lower for content such as mathematics, where word choice is less flexible.
• In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced it to 17%.

OpenAI also reports no meaningful performance differences with and without watermarking on the benchmarks it uses for Astra, its latest frontier model. On the Artificial Analysis Intelligence Index, Astra scored 49.57 points without a watermark and 49.76 with one. On GPQA Diamond the figures were 94.44% and 93.94%, and on DeepSWE v1.1 they were 72.80% and 71.68%.

What a watermark cannot tell you

OpenAI lists several conclusions that a detection result does not support:

• It does not measure human contribution. It can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing or creativity went into it.
• It does not establish ownership or responsibility, and does not show whether use was lawful or disclosure was required.
• It does not identify the user, and does not link a person, organization, account, prompt or conversation to the text.
• It does not verify accuracy.
• The absence of a watermark does not prove human authorship. Text may be too short, edited or translated for reliable detection, may come from an unsupported model, may predate watermarking, or may have been produced by another company's tools.

Part of a layered provenance approach

OpenAI says no single provenance technique is enough, so it combines open standards, durable watermarking and verification tools. It adds Content Credentials to supported image outputs, is C2PA conformant, and embeds invisible SynthID watermarks in supported images and audio. Image and audio verification stays publicly available through openai.com/verify and the Content Provenance API. The new restrictions apply only to text.

For users in the EU, ChatGPT and Codex text will soon carry a hidden signal. API customers elsewhere can decide whether watermarking fits their own transparency obligations. OpenAI says it expects to revisit each part of the approach as technology, standards and evidence evolve, and will expand detector access when results can be interpreted responsibly.