OpenAI Starts Watermarking ChatGPT Text for EU Users
OpenAI has begun rolling out invisible text watermarking for eligible ChatGPT and Codex outputs in the European Union to comply with transparency requirements under the EU AI Act. The system, called textGrain, embeds a statistical pattern into AI-generated text by subtly influencing word choices rather than inserting hidden characters or formatting changes. According to OpenAI, the watermark becomes more reliable as text length increases, reaching detection rates of roughly 95% on 400-token passages under test conditions. The company warns that editing, paraphrasing, or translating content can significantly reduce detection accuracy, with even modest word substitutions weakening the signal. OpenAI also notes that code and shorter responses may not be consistently watermarked because of technical limitations and regulatory exceptions. Access to watermark detection tools will initially be restricted to approved researchers and expert organizations rather than the general public. The company emphasizes that a detected watermark does not identify the author, reveal prompts, verify accuracy, or prove ownership of the content.