AI news ·
OpenAI brings text watermarking to ChatGPT and codex in the EU
OpenAI will roll out an invisible text watermark to ChatGPT and Codex in the EU to comply with the EU AI Act. Replacing just 10% of words drops detection accuracy from 92% to 66%.
OpenAI will roll out an invisible watermark called textGrain to ChatGPT and Codex text across the European Union in the coming weeks, a move designed to comply with the EU AI Act's transparency rules that took effect on August 2, 2026. The feature will be enabled by default for EU users, while API customers worldwide can switch it on starting immediately - though it remains off by default for those developers.
The watermark embeds a statistical pattern into generated text that detection tools can later identify. OpenAI's internal tests show detection rates drop sharply when text is edited. Replacing 10% of words reduces detection from roughly 92% to 66%. At 25% word replacement, detection falls to about 17%. The system also performs worse on mathematical content and short passages.
The company said watermarking does not identify individual users, cannot prove human authorship, and does not affect model performance on benchmarks like GPQA Diamond or DeepSWE. OpenAI plans to release textGrain as open source and expand detector access once results can be interpreted responsibly.
How textGrain compares to existing tools
The core mechanics resemble Google's SynthID, which also embeds imperceptible statistical signals into AI-generated text. Both systems share the same fundamental weakness: even modest human editing - rewording a quarter of a paragraph - makes the watermark nearly invisible to detectors. This limitation matters for anyone relying on detection tools to verify content authenticity.
OpenAI previously hesitated to release watermarking technology. A 2024 Wall Street Journal report, cited by TechCrunch, found the company feared users would migrate to competing platforms that did not watermark their output. The EU AI Act's legal requirements appear to have forced the timeline.
What the rollout leaves unclear
The October 5 update does not specify which models are included in the API opt-in. It also remains unknown whether EU ChatGPT users will have the ability to disable the watermark on their accounts. Search Engine Journal flagged both gaps, and no subsequent reporting has clarified them.
For API users outside the EU, the opt-in nature of the feature means adoption will depend on whether developers see value in proving their text came from OpenAI's models. The watermark does not identify the specific user who generated the text - only that it likely originated from a supported model.
Why this matters for writers, marketers, and legal professionals
For anyone producing or reviewing AI-generated content, textGrain changes the transparency equation without solving the authentication problem. A watermark that vanishes after light editing offers limited protection against misuse - but its presence in default EU outputs means compliance teams should understand what the signal does and does not prove. Writers and marketers using ChatGPT for drafts should assume watermark detection is unreliable on any text that has been substantially revised. Legal and regulatory professionals working with the AI Regulatory Compliance Courses framework will need to track how watermarking requirements evolve as the EU AI Act's enforcement matures.