OpenAI announced it will begin adding invisible watermarks to ChatGPT text outputs by default for users in the European Union, a move designed to comply with the EU AI Act's transparency requirements. The watermarks, which are embedded in the text at the token level, are intended to help identify AI-generated content and provide a provenance layer for downstream consumers.
However, the company confirmed the feature will only be enabled for EU users, not globally. The decision highlights the growing fragmentation of AI regulation, with different regions imposing different requirements on the same products. Users outside the EU will not receive watermarked outputs unless they specifically opt in, creating a two-tier system that critics say undermines the universal application of transparency standards.
Critics have pointed out that text watermarks are relatively easy to circumvent — paraphrasing, translation, or simple post-editing can strip them. The technology is more reliable for images and video, where alterations are more visible, but text remains a challenge. OpenAI acknowledges the limitations but says the watermarks are one layer of a broader provenance system that includes metadata and content credentials.
The EU AI Act, which began phasing in earlier this year, requires providers of general-purpose AI models to disclose when content is AI-generated. OpenAI's move is among the first major implementations of this requirement by a frontier lab, setting a precedent that other providers will likely follow in the coming months.
