OpenAI Begins Text Watermarking for EU ChatGPT Users

OpenAI is initiating a rollout of invisible watermarking for text produced by ChatGPT and Codex within the European Union. According to an official blog post, the company is moving to satisfy the transparency mandates established by the EU AI Act, which officially came into effect on August 2.

Technical Implementation of textGrain

The Scale Report understands that the mechanism, dubbed textGrain, functions by subtly influencing the model's sequence of word predictions. By applying a secret key during generation, the system embeds a detectable pattern that persists even if the text is copied or moved. OpenAI released a technical report co-authored by researchers from the University of Pennsylvania and Yale to explain the process.

Reliability and Market Strategy

Despite the deployment, the efficacy of the watermark is vulnerable to simple edits. Testing revealed that replacing 10% of the content with synonyms can drastically reduce detection rates, while shorter passages and math-heavy responses remain difficult for the detector to identify. Consequently, OpenAI is restricting access to detection tools to approved research organizations.

This shift highlights the tension between regulatory compliance and user retention. Reports from 2024 indicated that OpenAI previously avoided releasing such features due to fears of losing users to competitors. By contrast, Anthropic has already implemented global watermarking for its Claude model, a move that faced criticism from users who felt the tool did not account for their own editorial contributions.