OpenAI Begins EU-Specific Watermarking for ChatGPT Text

The watermark operates by subtly adjusting word choices during generation, creating a pattern undetectable to humans but identifiable by specialized software. Because this signature is embedded directly into the text, it persists even when content is copied and pasted elsewhere. OpenAI maintains that the process does not compromise model performance or identify individual users.

Technically described in the company's new "textGrain" report, the method uses a secret key to guide next-word predictions. However, the system is not foolproof. Replacing just 10% of words with synonyms can cause detection rates to plummet from 92% to 66%. Furthermore, short passages, mathematical solutions, and translations remain difficult to verify. Due to these reliability concerns, OpenAI is restricting detector access to approved researchers and expert organizations for the time being.

This rollout follows a similar move by Anthropic, which implemented worldwide watermarking for its Claude model earlier this year. OpenAI had previously developed a similar tool but hesitated to release it, citing fears that users might migrate to competitors lacking such restrictions. The company now stresses that the absence of a watermark does not confirm human authorship, as content may be heavily edited or generated by alternative AI systems.