OPENAI has announced steps to implement text watermarking in response to the EU AI Act, introducing an invisible signal, called textGrain, to eligible ChatGPT and Codex outputs in the European Union. The feature will also be available to developers via the API, marking a transition from lab tests to real-world deployment.
The rollout is being conducted in phases: API customers can enable watermarking for select models now (off by default), EU users of ChatGPT and Codex will begin seeing watermarks in the coming weeks, and approved researchers and organisations will be able to apply for detector access. OpenAI emphasises a regional approach to gather feedback and is coordinating with cloud partners to broaden platform availability.
textGrain works by embedding an invisible statistical signal in the model’s word choices, which can be detected by a specialised detector. OpenAI claims the method matches or surpasses previously tested techniques like SynthID for text and notes that watermarking does not degrade output quality; benchmarks on Astra, the company’s frontier model, showed minimal impact (e.g., GPQA Diamond scoring 94.44% without watermarks and 93.94% with them).
However, detection has clear limits: at a 1% false-positive rate, about 80% of 200-token passages were detected, with higher success on longer passages and on some subject areas; editing, such as substituting 10–25% of words with synonyms, significantly reduces detectability. OpenAI will not publicise the detector at launch and warns that watermarking cannot prove authorship, verify accuracy, or identify a user.
The watermark is positioned as one layer within a broader provenance strategy, alongside Content Credentials for images and SynthID for multimedia. OpenAI also maintains an open verify tool for image and audio checks, and the company cautions that no single provenance technique is sufficient on its own. In practice, organisations should treat watermark results as one signal among several when assessing content provenance.