OpenAI will start watermarking ChatGPT’s text in the EU

AI-generated text in the EU now carries an invisible identifier, meaning its origin is traceable.

OpenAI will begin embedding invisible watermarks into text generated by ChatGPT and Codex within the European Union, effective immediately. This measure is a direct response to the EU AI Act's transparency mandates, which require AI-generated content to be identifiable by other systems. The watermark, a subtle pattern shaped by word choices rather than a visible symbol, will be applied to all eligible users across all plans in the EU over the coming weeks. Developers globally can opt into this feature for select API models starting now, though it is off by default. OpenAI has stated the watermark does not impact model performance and does not identify individual users. The company also released technical details of its textGrain method, developed with university researchers, which uses a secret key to subtly influence word predictions.

The EU AI Act's requirement for AI-generated content to be identifiable is driving a new wave of transparency tools, with OpenAI's watermarking serving as a prominent example. While this move aims to comply with regulations and foster trust, it also introduces complexities. The effectiveness of the watermark is not absolute, as OpenAI acknowledges that editing, especially replacing a significant portion of words, can degrade detection rates. Furthermore, short passages, mathematical solutions, and translated content are noted as being more challenging to watermark, suggesting that the current implementation may not cover all forms of AI output comprehensively.

This development signals a significant step towards regulatory compliance in the generative AI space, particularly within the EU. It establishes a precedent for how AI providers will need to integrate transparency mechanisms into their core offerings. The company's cautious rollout, offering initial detector access only to approved researchers, suggests an ongoing effort to refine reliability and address potential misuse. As other regions and further iterations of the EU AI Act evolve, we can anticipate more sophisticated methods of content provenance and detection becoming standard, impacting how AI-generated content is authenticated and utilized across various industries.