OpenAI Details Text Watermarking Approach for EU AI Act
OpenAI is rolling out a phased text watermarking approach to meet EU AI Act requirements, introducing its textGrain technology, API opt-in capabilities, and initial detector access for approved researchers.
Radwan AbdulhadiEditor and technology writer
Responding to EU AI Act Requirements
OpenAI has shared its approach to text watermarking in response to the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine-readable way. As detailed by [OpenAI News](https://openai.com/index/eu-text-provenance/), text watermarking and detection remain early technologies with significant limitations, and views about their benefits and responsible uses are still developing.
The company's phased approach reflects both the regulatory requirements and the inherent limitations of the technology. OpenAI emphasizes the importance of transparency regarding what a text watermark can and cannot reveal to users.
Phased Rollout for APIs and European Users
Starting immediately, API customers globally can choose to opt in to text watermarking for select models, though the feature will remain off by default within the API. Over the coming weeks, OpenAI will also begin adding an invisible watermark to eligible ChatGPT and Codex text output generated within the European Union.
Concurrently, applications are opening for access to the text watermark detector. This access will initially be limited to approved researchers and expert organizations who can help evaluate and improve the technology.
It is important to note that these new measures apply strictly to text provenance. Existing verification tools for audio and images—such as the [openai.com/verify ](https://openai.com/verify) web tool and the [Content Provenance API (opens in a new window)](https://developers.openai.com/api/docs/guides/content-provenance)—will continue to remain publicly accessible for organizations evaluating media files.
Introducing the textGrain Technology
The underlying text watermarking technology, known as textGrain, embeds an invisible statistical signal into a model's word choices. The system's detector scans for this signal to determine whether a given passage contains an OpenAI watermark. Further technical insights can be found in the official [technical report (opens in a new window)](https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf), which is slated for updates with additional details in the coming weeks. OpenAI also plans to release the technology as open source to allow external developers to build upon it.
According to internal evaluations, textGrain matched or exceeded the performance of other tested approaches, including SynthID for text. However, strong performance under ideal conditions does not guarantee reliable detection during everyday use.
Detection Challenges and Limitations
OpenAI's evaluations highlight that detectors can encounter false positives—reporting a watermark where none exists—or false negatives—missing a watermark that is present. Shorter or more constrained text proves particularly challenging to detect reliably.
At a target false positive rate of 1%, the detector successfully identified watermarks in roughly 80% of 200-token passages, compared to about 95% of 400-token passages for content categories like psychology. Conversely, detection rates dropped substantially for subjects such as mathematics, where word choices offer less flexibility.
Furthermore, text editing can weaken the watermark signal significantly. In evaluations involving 400-token passages, replacing 10% of words with synonyms reduced the detection rate from approximately 92% to 66%, while replacing 25% of words drove the rate down to 17%.
Interpreting Watermark Results
OpenAI stresses that text watermarks offer only a limited signal regarding the role its systems played in generating a passage. A positive detection result comes with several distinct boundaries on what can be inferred.
A watermark does not measure human contribution, meaning it cannot indicate how much human judgment, editing, or creativity went into a text. It also fails to establish ownership or legal responsibility, does not determine whether text use was lawful or disclosure was required, and does not identify the underlying user, account, or prompt associated with the passage.
Sources
- OpenAI NewsOur approach to EU text provenance rules
Continue chronologically
Related entity coverage
Related stories
OpenAI Hosts DevDay 2026 With Over 20 Major Announcements
OpenAI highlights more than 20 major announcements at DevDay 2026, introducing new agents, Codex updates, models, and builder tools.
OpenAI Expands Academy with Role-Specific AI Learning Paths
OpenAI has expanded its OpenAI Academy to include specialized learning paths for developers, leaders, educators, and students to build practical AI skills.
OpenAI Is Testing More Visual Ads in ChatGPT
OpenAI is rolling out a new visual ad format in ChatGPT for US users, beginning with image generation features later this month.
OpenAI to Show Visual Ads in ChatGPT During Image Generation
OpenAI is expanding ChatGPT ads by testing a new visual format during image generation in the U.S. and introducing new measurement tools.
