OpenAI launches EU-only text watermarking for ChatGPT

OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act. The feature, called textGrain, will roll out over the coming weeks to eligible users on all plans within the EU, while API developers worldwide can opt to activate it.

OpenAI is implementing the change in response to the EU AI Act's transparency rules, which took effect in August. The regulation requires generative AI providers to make their text output machine-readable. Providers already on the market have until 2 December to comply. The watermark rollout is currently limited to the EU.

Technical approach and detection challenges

The watermarking system, textGrain, uses a secret key to subtly shape the model's word choices, creating a statistical pattern. A detector can later spot this pattern using only the text and the key. Because the signal lives in the words themselves, it travels with the text when copied and pasted.

OpenAI published a technical report on the method with researchers from the University of Pennsylvania and Yale. The company said textGrain matched or beat other approaches it tested, including Google's SynthID for text. Benchmark scores for its GPT-6 Astra model showed no meaningful performance difference with the watermark enabled.

Detection effectiveness varies significantly with text length and editing. At a target false positive rate of 1%, the detector found the watermark in about 80% of 200-token passages. For 400-token passages, the rate was about 95%. Editing drastically reduces detection.

Detection is also harder for short passages, math answers, and translated text. OpenAI cautions that a detected watermark does not prove human authorship, reveal the generator's identity, or indicate how much of the text was edited by a human. Conversely, a missing watermark does not prove a human wrote the text.

Compliance and regulatory context

The initiative is a direct response to Article 50 of the EU AI Act. OpenAI announced the watermarking in a blog post. The regional approach gives the company room to learn from real-world use and feedback before considering wider deployment.

Applications for the watermark detector are open but initially limited to approved researchers and expert organizations. OpenAI said it will grant access case by case, in line with the EU's Code of Practice. The company plans to widen access once results can be interpreted responsibly and is also working with cloud partners to offer the technology through their services.

Industry context and limitations

OpenAI's move follows broader industry efforts. Anthropic began watermarking text generated by Claude worldwide in August, a decision that drew some user backlash. Tools to strip AI watermarks have since appeared. Google DeepMind has also used its SynthID technology to watermark AI-designed proteins.

The text detection tool is not publicly accessible at launch. OpenAI stated it will also release the textGrain technology as open source. Its separate tools for checking images and audio remain publicly available.

OpenAI will grant watermarking access case by case to approved researchers and expert organizations, with a wider rollout planned after responsible interpretation of results.