OpenAI is adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act. The watermark, which subtly shapes word choices, will roll out over the next few weeks to EU users. While it won’t identify users, it can be removed by editing 10% of the text, making detection less reliable.

The watermark doesn’t impact the model’s performance, according to OpenAI, but researchers will need approval to access detectors. Anthropic has also begun watermarking its text globally, drawing criticism from some users who feel they provided the creative input.

OpenAI had previously built a text watermark but held off on release due to concerns about users switching to rivals that didn’t watermark. Now, as a compliance measure, the watermark will help ensure transparency in AI-generated content.

The move comes as other tech giants, including Anthropic, Google, Meta, and Microsoft, commit to following the EU’s code of practice on AI-generated content, aiming for greater accountability and transparency.