OpenAI is preparing to mark AI-generated text in a way readers cannot see, but detectors may be able to identify. The company’s approach, called textGrain, is designed for ChatGPT and Codex users in the European Union, while API customers worldwide will have a choice over whether to use it.
What OpenAI is changing
The EU AI Act reportedly requires AI providers to label generated text in a machine-readable format. OpenAI’s answer is an invisible watermark that leaves a statistical signal in the model’s word choices rather than placing a visible label on the page.
OpenAI says watermarking will be turned on for ChatGPT and Codex users in the European Union over the coming weeks. API watermarking, by contrast, will be opt-in worldwide. The same API option is also expected to become available through cloud partners such as Microsoft Azure in the coming weeks.
That setup differs from Anthropic’s mandatory watermarking for Claude, which applies globally regardless of how users access the models. OpenAI is choosing a narrower default rollout for consumer-facing EU use, paired with optional adoption for developers and businesses using the API.
How textGrain is supposed to work
textGrain embeds an invisible statistical signal into generated writing. The source describes it as working much like Claude’s SynthID watermark, which uses Google’s open-source technology.
In practical terms, the watermark is not a visible stamp, byline, or disclosure notice. It is a pattern that can be detected only by a tool designed to look for it. OpenAI says the tool will report only whether it detected an OpenAI watermark.
The company also says it plans to release textGrain as open source so others can build on it. OpenAI says the system matched or beat other approaches in internal tests, including Google’s SynthID for text.
Detection depends on length, topic, and edits
The limits of watermark detection are central to understanding what this change can and cannot do. OpenAI’s own figures show that detection rates vary depending on the length and subject matter of the text.
With the detector set to a target false-positive rate of 1 percent, it identified watermarks in about 95 percent of 400-token passages about psychology. At 200 tokens, detection fell to about 80 percent.
OpenAI says detection rates were “substantially lower” for math content. The reason given is that the model has less freedom to choose words in that kind of output, which makes the statistical signal harder to preserve.
Longer passages might make the signal stronger and improve detection, but the source notes that results could still depend on the content. OpenAI provides no data to test that assumption.
Editing also weakens detection. OpenAI says replacing just 10 percent of words with synonyms cuts detection rates for 400-token passages from about 92 percent to 66 percent. Replacing a quarter of the words drops detection to 17 percent.
That means the watermark can be made much harder to find through relatively small changes. The source also notes that this weakness could work in OpenAI’s favor, because users who do not want ChatGPT use detected might switch to open-weight models if watermarks become harder to remove.
What a watermark does not prove
OpenAI’s framing is careful about what detection means. A detected watermark does not show how much human creativity or editing went into the text. It also does not establish ownership, assign responsibility, identify a user, or prove that the content is accurate.
The reverse is also true. If a detector does not find a watermark, that does not prove a human wrote the text. The passage may be too short, edited, translated, or produced by a model the detector does not support.
OpenAI says watermarking does not hurt output quality, citing tests of its frontier model Astra. The company reports no significant performance differences with watermarking on or off across eight benchmarks, including GPQA Diamond, BrowseComp, and DeepSWE.
Those benchmark results do not settle every quality question. The source notes that they do not establish whether watermarking affects writing quality, an issue critics have also raised about Claude’s watermark.
Detector access will start narrow
OpenAI will not immediately make the textGrain detector broadly available. Only selected researchers and specialist organizations will initially get access, and they can apply through an application form.
OpenAI says access will be granted case by case under the EU’s Code of Practice. Anthropic takes a similar approach with its detection API.
The company says it is restricting access because the detector can flag unmarked text or miss watermarks. OpenAI plans to expand access “when we believe results can be interpreted responsibly.” It has not said when that might happen.
For images and audio, existing verification tools remain publicly available. The source names openai.com/verify and the Content Provenance API as examples.