OpenAI is opening text watermarking to API customers worldwide while preparing to add it to eligible ChatGPT and Codex text output in the European Union. The rollout uses textGrain, a technology that embeds an invisible statistical pattern in generated wording. It also comes with a constraint for developers: enabling watermarking does not automatically give them access to the detector.
In its announcement, OpenAI presents the change as a response to the EU AI Act and its commitments under the EU Code of Practice on Transparency of AI-Generated Content. API watermarking is optional and remains off by default. The company says eligible ChatGPT and Codex output in the EU will receive watermarks over the coming weeks.
The distinction matters for application builders. The API rollout offers a choice about adding a provenance signal to model output; the EU product rollout introduces that signal within OpenAI’s own services. Neither makes every AI-generated passage reliably identifiable.
What is available, and where
API customers globally can opt in for select models, with watermarking remaining off by default, according to OpenAI’s announcement.
For ChatGPT and Codex, OpenAI’s announced scope is eligible text output in the EU. The company describes a rollout over the coming weeks, so the announcement should not be read as confirmation that every eligible output already carries the signal.
Cloud distribution is also still developing. The Decoder reports that API watermarking will become available through cloud partners such as Microsoft Azure in the coming weeks. Availability through those services is therefore a separate rollout consideration from OpenAI’s direct API opt-in.
Detection has a narrower initial audience. OpenAI is accepting applications from researchers and expert organizations, with access initially limited to approved applicants who can help evaluate and improve the technology. Developers deciding to watermark their output should distinguish that generation setting from the ability to inspect passages for the resulting mark.
How textGrain puts a signal into words
textGrain subtly adjusts the model’s random choice between possible words or word pieces as it writes, according to OpenAI’s Help Center. Across a passage, those choices form a statistical pattern that a detector can look for, including after some edits.
According to the OpenAI Help Center, the watermark is part of the wording itself and is not visible to readers. The watermark is a property of the pattern of language choices across the text.
That mechanism helps explain why detection depends on the content. Where a model has more flexibility in how to phrase an answer, there are more opportunities to embed a pattern. OpenAI reports substantially lower detection rates for mathematics, where word choice is less flexible. A result measured on explanatory prose therefore cannot simply be assumed to apply to every kind of output an application generates.
OpenAI says textGrain matched or exceeded other approaches it tested, including SynthID for text. It also reports no meaningful performance differences with and without watermarking across the benchmarks it uses to assess its latest frontier model, Astra. Those are company evaluation claims; they do not establish identical behavior for every downstream task.
The company plans to make the technology available in open source. That remains a stated plan, rather than evidence that an open-source implementation is already available.
Longer passages help, but editing weakens detection
OpenAI’s reported results show how much the detector depends on the passage it receives. At a target false positive rate of 1%, it identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology.
Those percentages describe detection under the reported evaluation conditions. They are not a guarantee that a particular passage will be recognized, and the target false positive rate means detection is not framed as error-free. The lower results for mathematics add another reason to consider content type alongside passage length.
Editing can weaken the pattern sharply. In a separate evaluation of 400-token passages, OpenAI reports that replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% reduced it to 17%. The 92% starting point belongs to that editing evaluation; it should not be conflated with the roughly 95% result in the passage-length comparison.
For an application that rewrites or edits model output before presenting it, these results raise a practical question: how much of the generation-time signal survives in the final text? The evidence supports treating that as a workflow-dependent limitation. It does not supply a universal threshold at which edited text becomes undetectable.
What a detected watermark can establish
The detector’s scope is deliberately limited. According to The Decoder, the tool reports only whether it detected an OpenAI watermark. The publication also reports OpenAI’s explanation that restricted access reflects the risk of flagging unmarked text or missing existing watermarks.
Even a detected watermark has a narrow meaning. Unite.AI reports OpenAI’s stated limitations: the signal does not measure human contribution, establish ownership or responsibility, identify the user, or verify accuracy. It cannot settle how much a person contributed to a passage after generation.
These limits also shape how an application should interpret an absent signal. Given the reported missed detections, weaker performance on some content and sensitivity to editing, failure to detect a watermark cannot establish that a passage was written entirely by a human.
The announcement concerns text detection access. OpenAI says its verification tools for images and audio will continue to be publicly accessible. Their availability should not be taken as evidence that the text detector has the same access policy.
A provenance choice for developers
OpenAI describes text watermarking as supporting the EU AI Act requirement to mark generated text. That gives the rollout a regulatory purpose, but the company’s own limitations leave a distinction between adding a machine-readable signal and drawing conclusions from it.
For developers, the immediate decisions are concrete: whether to enable the available setting for supported API use, whether the relevant cloud service has rolled it out, and how subsequent editing affects the text ultimately delivered to users. Detector access is a separate dependency, while ownership, authorship contribution and factual accuracy remain outside what the watermark establishes.
textGrain adds a provenance signal at generation time. Its usefulness downstream depends on the passage, the processing it undergoes and access to a detector whose results still require careful interpretation.
