OpenAI to Default-Watermark ChatGPT Text Outputs in the EU
OpenAI is enabling text output watermarking by default for ChatGPT users located within the European Union. This setting specifically targets the EU market to align with regional transparency requirements. This move underscores how European regulatory mandates, such as the EU AI Act, are pressing AI companies to implement technical compliance safeguards even when the underlying tools face practical limitations. It highlights a growing rift in default product experiences between the EU and the rest of the world. Despite being deployed to fulfill governance obligations, current LLM text watermarking techniques remain technically unreliable and can be easily bypassed by paraphrasing or editing the output. As a result, OpenAI is confining this default watermarking policy strictly to the EU rather than deploying it globally.
## BACKGROUND
Article 50 of the EU AI Act mandates transparency obligations for providers of AI systems, requiring synthetic content to be marked so it can be identified as AI-generated. Large language model (LLM) text watermarking works by subtly modifying the probability distribution of generated words (tokens), creating a hidden statistical signature that detection algorithms can trace without disrupting text readability.