OpenAI Implements Text Watermarking for ChatGPT in the EU
OpenAI introduces textGrain technology to comply with the European Union's AI Act transparency mandates.
1EU AI Act Compliance
OpenAI is introducing invisible watermarking for ChatGPT and Codex in the European Union to meet Article 50 requirements. This mandate requires generative AI providers to ensure machine-readable detection of artificial content. While the rollout begins in the EU, global API customers can now opt in to use these watermarking features.
2The textGrain Mechanism
The textGrain technology functions by embedding an invisible statistical signal directly into the model's word choices during generation. Rather than inserting hidden characters or specific tokens, it influences token selection using a secret key. This creates a detectable pattern that can be analyzed by specialized software to verify the content's origin.
3Detection Accuracy Metrics
OpenAI reports that its detector identifies watermarks in 80% of 200-token passages and 95% of 400-token passages within psychology-related content. However, performance varies significantly by domain; detection is substantially lower for mathematics due to limited flexibility in word choice. These figures maintain a target false-positive rate of approximately one percent.
4Vulnerability to Editing
The watermark signal is highly sensitive to text manipulation. Replacing just 10% of words with synonyms drops detection rates from 92% to 66%. If 25% of the text is altered, detection falls to 17%. Consequently, OpenAI is restricting access to the detector to approved researchers and expert organizations to prevent misuse.
5Provenance vs Authorship
OpenAI clarifies that textGrain is a provenance signal, not a definitive test for AI authorship. It cannot measure human judgment, creativity, or editing effort. Furthermore, the absence of a detected watermark does not prove human creation, as detection can fail due to translation, short text length, or use of unsupported models.
6Broader Provenance Strategy
OpenAI views text watermarking as only one component of a larger provenance ecosystem. For audio and image content, the company utilizes C2PA-based Content Credentials and SynthID watermarks. They maintain that no single technique is sufficient on its own, necessitating a multi-layered approach to verify the authenticity of various digital media formats.
7Open Source Commitment
To foster industry-wide improvements, OpenAI plans to release textGrain as open-source technology. This move aims to allow external developers and researchers to build upon their approach. By sharing the methodology, the company hopes to refine detection capabilities and establish more robust standards for identifying AI-generated content across the global digital landscape.
8Regulatory Context and Impact
The transparency obligations under the EU AI Act have been in effect since August 2, 2026. Major industry players including Google, Meta, and Anthropic have signed the voluntary Code of Practice regarding these requirements. This shift signals a transition toward standardized transparency measures for generative AI systems operating within European markets.