OpenAI adds textGrain watermarks to ChatGPT in the EU

OpenAI is rolling out an invisible text watermarking system called textGrain to comply with the European Union's AI Act. Over the coming weeks, the company will enable this technology for ChatGPT and Codex users within the EU. Unlike competitor Anthropic, which mandates watermarks globally for Claude, OpenAI is making textGrain an optional opt-in feature for API users worldwide. The technology will also become available through cloud partners like Microsoft Azure in the near future, and OpenAI plans to eventually release it as open-source software.

The textGrain system embeds a statistical signal directly into the model's word choices, similar to Google's SynthID technology. OpenAI tested the system on its frontier model, Astra, and reported no significant performance degradation across eight benchmarks, including GPQA Diamond, BrowseComp, and DeepSWE. However, the effectiveness of the watermark depends heavily on the length and subject matter of the generated text.

With a target false-positive rate of 1 percent, the detector identified watermarks in 95 percent of 400-token psychology passages, but that rate fell to 80 percent at 200 tokens. For math content, where word choices are more constrained, detection dropped to 60 percent for 400-token texts. Furthermore, simple edits easily disrupt the signal. Replacing 10 percent of the words in a 400-token passage with synonyms reduced detection from 92 percent to 66 percent, while replacing 25 percent of the words caused detection to plummet below 20 percent.

For developers and enterprise practitioners, the optional nature of the API watermark provides crucial flexibility, preventing forced compliance that might drive users toward open-weight alternatives. Currently, OpenAI is restricting access to the textGrain detector to selected researchers and specialist organizations under the EU's Code of Practice. The tool will only confirm if an OpenAI watermark is present, without exposing user identities, prompts, or conversation histories. Meanwhile, public verification tools like openai.com/verify and the Content Provenance API remain available for audio and images.