OpenAI to Watermark ChatGPT Text in EU for AI Act Compliance

OpenAI will introduce an invisible watermarking system to text generated by ChatGPT and Codex within the European Union, aligning with transparency mandates under the EU AI Act. The regulatory framework, which took effect on August 2, requires AI providers to embed machine-readable markers in synthetic content. OpenAI announced the initiative on Monday, stating that the feature will roll out gradually over the coming weeks to eligible EU users across all subscription tiers. While the same capability is now available to global API developers, it remains disabled by default outside the bloc. The watermarking mechanism, detailed in a co-authored technical paper developed with researchers from the University of Pennsylvania and Yale University, operates through a method termed textGrain. Rather than inserting visible symbols, the system subtly biases the model's next-word predictions using a cryptographic key. This creates a statistical pattern embedded directly within the text, allowing authorized detectors to identify synthetic origin without altering readability. OpenAI confirmed that enabling the feature produces no measurable degradation in model performance. Importantly, the system is designed to protect privacy by not tracking individual users. Despite its technical sophistication, OpenAI has acknowledged inherent limitations. Edited passages, short excerpts, mathematical solutions, and translated content show reduced detection accuracy. Swapping merely ten percent of vocabulary with synonyms can cause detection rates to fall from approximately ninety-two percent to sixty-six percent. Given these constraints, the company will initially restrict detector access to vetted researchers and authorized organizations to ensure responsible evaluation and application. OpenAI also clarified that the absence of a watermark does not verify human authorship, as content may originate from other AI providers or undergo extensive manual revision. This deployment follows industry-wide shifts toward AI transparency. Anthropic recently implemented global watermarking for its Claude models, prompting debate among users regarding attribution and tool versus authorship dynamics. OpenAI's current rollout represents a measured approach, having previously delayed global implementation due to competitive concerns. Leading technology firms, including Google, Meta, Microsoft, and Anthropic, have already pledged adherence to the EU's voluntary code of practice on synthetic content disclosure. The initiative underscores a broader industry transition toward standardized, regulation-compliant transparency measures as AI integration accelerates across digital platforms.