OpenAI has announced a rollout of invisible watermarks for text outputs. API customers can enable them now, while the company prepares a phased rollout for eligible ChatGPT and Codex users in the European Union. According to OpenAI, the technology lets a detector recognize a statistical signal embedded in a model’s word choices.
API watermarks are optional, with ChatGPT and Codex rollout planned for the EU
From October 5, 2026, API customers worldwide can activate text watermarks for select models. The API’s default setting remains unchanged: watermarking stays off unless a customer enables it. OpenAI is also working with cloud partners to make watermarking available for outputs from its models through their services in the coming weeks.
For ChatGPT and Codex, the company plans to introduce watermarking over the coming weeks for eligible users across all plans in the European Union. It will not be a worldwide default at launch. OpenAI links the rollout to the EU AI Act’s requirement for machine-readable identification of text generated by generative AI.
How textGrain works and who can access the detector
According to OpenAI, textGrain influences word choices during generation so that the resulting text carries an invisible statistical signal. The detector searches a submitted passage for that signal to assess whether it contains an OpenAI watermark. The company plans to release textGrain as open source so other developers can use and build on the technology.
Researchers and expert organizations can already apply for detector access. OpenAI will initially approve applications individually to support evaluation and improvement of the technology. The detector will not be publicly available at launch because it can miss a watermark or incorrectly report one where none exists.
According to OpenAI, the tool is designed to report only whether it finds the company’s watermark. The watermark does not link text to a particular person, organization or account. The detector will not reveal a user’s prompt or the contents of their conversations.
Detection depends on text length, subject matter and editing
In OpenAI’s own evaluation, textGrain matched or outperformed the other approaches tested, including SynthID for text. The company also cautions that results under ideal conditions do not guarantee reliable detection in everyday use.
OpenAI’s tests of responses to questions from the ELI5 dataset found that detection rates varied with passage length and subject matter. At a target false positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages for content such as psychology. For 400-token passages, the rate was about 95%. Detection was substantially lower for mathematical content, where word choices are more constrained.
Another OpenAI test examined synonym substitutions in watermarked English responses to ELI5 questions. For 400-token passages, replacing 10% of the words reduced detection from about 92% to 66%. When 25% of the words were replaced, the detector found the watermark in 17% of cases.
Astra tests compare watermarked and unwatermarked outputs
OpenAI says its Astra benchmarks show no meaningful performance differences between watermarked and unwatermarked outputs. The evaluation compares two watermark settings for the same Astra, max configuration, rather than the capabilities of different models.
Individual results in the company’s tests moved in both directions. On DeepSWE v1.1, the configuration scored 72.80% without watermarking and 71.68% with it. On Terminal-Bench 4.0, it scored 53.90% without watermarking and 56.06% with it.
What a detected watermark says about a passage
According to OpenAI, a detected signal can indicate that its system generated or processed part of a passage. It does not measure the human contribution, including the judgment, editing or creativity involved. The company also notes that a watermark does not establish ownership, lawful use, any obligation to disclose AI involvement or responsibility for the text.
OpenAI says the check does not establish whether the text is accurate. A watermark does not indicate whether a passage is misleading, harmful or presented in the appropriate context.
The absence of a detected watermark does not prove human authorship, according to OpenAI. Short passages, editing or translation can hinder detection. An output may also come from an unsupported model, predate the introduction of watermarking or have been generated by another company’s tools.
