OpenAI announced on October 5 that it will embed invisible watermarks into text generated by ChatGPT and Codex models for users in the European Union, beginning in the coming weeks. The initiative represents the company's response to the EU AI Act, which mandates that generative AI systems mark their outputs in a machine-readable format to flag content as artificially generated.
The watermarking system, which OpenAI calls textGrain, works by adjusting the statistical patterns of word choices during text generation. The watermark remains imperceptible to human readers but can be detected by specialized software. According to OpenAI's testing, the detector identified watermarks in approximately 80 percent of 200-token passages and about 95 percent of 400-token passages in subjects like psychology. Detection rates were substantially lower for mathematics and other fields where word choice is more constrained.
Starting immediately, OpenAI API customers worldwide can opt into watermarking for select models, though the feature remains disabled by default globally. The mandatory watermarking in the EU reflects Article 50 of the EU AI Act, which began applying on August 2, 2026. The European Commission previously set a December 2, 2026 deadline for companies to meet the transparency and detection obligations for AI systems placed on the market before August 2. OpenAI stated that the EU-limited rollout allows the company to gather real-world feedback before considering broader deployment.
Access to OpenAI's detection tool will initially be restricted to approved researchers and expert organizations on a case-by-case basis. The company declined to make the detector publicly available at launch, citing risks of false positives when applied to large volumes of text. According to OpenAI, the watermark does not identify individual users or reveal the prompts provided to generate the content—it only reports whether a watermark is detected.
OpenAI's approach follows similar moves by competitors. Anthropic launched watermarking for its Claude model in August 2026, and Google DeepMind has been watermarking its outputs for roughly two years. However, neither of those companies' watermarking systems currently apply to United States users. OpenAI's watermarking also does not extend to U.S. users despite domestic support for such measures; a recent Rainey Center poll found that nearly 80 percent of Americans support AI watermarking requirements.
Testing revealed vulnerabilities in the watermarking system. When researchers replaced 10 percent of words in 400-token passages with synonyms, detection rates dropped from approximately 92 percent to 66 percent. Replacing 25 percent of words further reduced detection to 17 percent. OpenAI noted that translated text also proved harder to detect reliably. The company emphasized that a missing watermark does not prove human authorship, as editing, translation, use of older models, or other AI tools could all result in undetected watermarks.
OpenAI stressed that watermarking does not measure human contribution to content and carries no implications for text accuracy. Performance benchmarks showed minimal to no difference in model outputs whether watermarking was enabled or disabled. The company acknowledged that sophisticated adversaries could attempt to remove or obscure watermarks through paraphrasing tools, manual editing, or translation, though it indicated ongoing efforts to strengthen watermark resilience.
