OpenAI Starts Watermarking ChatGPT and Codex Text in the EU
OpenAI will add an invisible watermark to text from ChatGPT and Codex in the EU over the coming weeks. The change affects all eligible users on all plans within the EU, but not globally, as per a blog post published on October 5, 2026. OpenAI acts in compliance with the EU AI Act, which mandates that AI-generated text be machine-identifiable.
How Text Watermarking Works
The system, named textGrain, incorporates a hidden statistical signal into the model's word choices. A detector then identifies this signal. OpenAI asserts that textGrain outperforms other methods it has tested, including Google’s SynthID for text, as detailed in their technical report. The company also intends to make the technology available as open source.
Limitations of Text Watermarking
OpenAI highlights several limitations:
- At a 1% false positive rate, the detector can identify watermarks in approximately 80% of 200-token passages on psychological topics. This drops to around 95% for 400-token passages.
- Replacing even 10% of words with synonyms significantly reduces detection accuracy from about 92% to 66%.
- Detection is less effective for mathematical content, where word choice is less flexible.
The watermark does not disclose user information, measure human contribution, establish ownership, or verify accuracy. A missing watermark doesn't definitively confirm human authorship either; it could be due to text length, editing, translation, or the use of alternative AI tools.
OpenAI will initially grant access to approved researchers and expert organizations on a case-by-case basis, in alignment with the EU’s Code of Practice. The detector reports watermark findings without revealing user details or prompts. Access will expand once results can be interpreted responsibly.
Compliance with the EU AI Act
Article 50 of the AI Act mandates that generative AI providers make their text output machine-readable by December 2, 2026, as reported by Engadget. Anthropic began watermarking Claude’s text worldwide in August, with certain exemptions, and tools to remove these watermarks have since emerged. Google DeepMind has also used SynthID to watermark AI-designed proteins.