OpenAI to Add Invisible Watermark to ChatGPT Text in EU to Meet AI Act Transparency Rules

OpenAI will embed an invisible watermark in all text produced by ChatGPT and Codex for users in the European Union, aiming to meet the EU AI Act’s new transparency requirements. The company announced the move on Monday, October 6, in a blog post and said the feature will roll out over the next few weeks to eligible users on every plan, though it will initially be limited to the EU. The watermark is off by default, but developers using OpenAI’s API worldwide can enable it for selected models. OpenAI acknowledged that watermarking and detection are early‑stage technologies with notable limits. Tests showed that swapping about 10% of words with synonyms drops detection rates from roughly 92% to 66%, and short passages, math answers or translated text are harder to flag. The watermark does not identify the user, does not imply ownership or responsibility, and its absence does not prove human authorship. The firm is also opening applications for access to its text‑watermark detector, restricting early access to approved researchers and expert organisations that can help evaluate reliability. A technical report named textGrain, co‑written with scholars from the University of Pennsylvania and Yale, explains the method of subtly nudging word choices so the pattern can be read by a detector. Anthropic made a similar announcement in early August, adding invisible watermarks to its Claude models to comply with the same EU rules. Both companies are among several AI firms pledging to follow the EU’s code of practice on AI‑generated content.

Context

The EU AI Act, effective from August 2, 2026, requires AI providers to embed machine‑readable marks in generated content, aiming to boost trust and curb misuse of synthetic media.