OpenAI Brings AI Text Watermarking to ChatGPT Users in the EU

OpenAI is rolling out an invisible watermarking feature called textGrain for ChatGPT and Codex users located exclusively in the European Union, starting in the coming weeks, as part of compliance measures with the EU AI Act.

How textGrain Alters AI Word Choices

To implement the watermarking mechanism, OpenAI uses an internal system designated as textGrain. The technology alters the output by introducing a subtle statistical signal directly into the model’s token selection process. According to documentation published on the OpenAI website, this changes the generated text’s word choice based on the algorithm rather than optimizing purely for the user’s prompt parameters. OpenAI noted in its announcement on October 5 that the update applies to eligible ChatGPT and Codex text outputs in the European Union, leaving users in the United States and the UK unaffected.

OpenAI will be rolling out text watermarking for ChatGPT and Codex in the EU in the coming weeks. Watermarking will remain off by default in the API. > "Across the benchmarks we use to assess Astra, our latest frontier model, we do not see meaningful performance differences… https://t.co/eO9j7hjXmn pic.twitter.com/4TnvH5qcAg”

— https://x.com/testingcatalog/status/2107141677060104617

The system relies on a specialized detector to assess whether a given text passage carries the OpenAI watermark. However, that detection tool is not being made publicly available at launch. Instead, OpenAI is restricting detector access exclusively to approved researchers and expert organizations that apply for access. The tool will report the presence of an OpenAI watermark without exposing user identities, prompts, or conversation histories.

Compliance Pressures and Regional Limitations

The decision by OpenAI follows similar moves by other generative AI providers. Earlier this year, Anthropic updated its systems to insert watermarks into generated text to comply with the EU AI Act, which requires all AI-generated text to be identifiable in a machine-readable format. While Anthropic chose to roll out its watermarking globally, OpenAI has limited its initial geographical rollout strictly to the EU.

OpenAI itself has acknowledged the technical limitations inherent in the software. Documentation specifies that shorter or more constrained text strings are significantly harder to detect, and basic text editing can easily weaken or strip out the statistical signal entirely. This dynamic creates a system that modifies the user’s requested output via an invisible signal, yet remains vulnerable to trivial workarounds like manual text editing, while the verification tool remains locked away from the general public.

Practical Limits of the EU Deployment

The restriction of the textGrain detector to select researchers leaves everyday users unable to independently verify whether a text carries the watermark. OpenAI cited the risk of false positives and missed watermarks as the primary reason for withholding public access to the detector at launch.

For European ChatGPT customers, the update means generation patterns will adjust to embed the statistical signal automatically over the coming weeks. Meanwhile, users outside the EU will continue receiving standard outputs without the algorithmic word-choice modifications.

More on this story: OpenAI textGrain Watermark: What It Means for Businesses in the EU · Google Makes SynthID Detector Publicly Available Globally