Rolling Out Invisible Watermarks Across the EU
To meet the European Union’s new transparency regulations for generative AI, OpenAI will introduce invisible watermarks in text and code generated within the EU. This initiative relies on the textGrain technology, which has been tested to deliver detection rates comparable to or better than the SynthID alternative for textual content. Detection success is around 80% for short texts, though it may decline when handling mathematical expressions or after content edits.
Scope and User Control Over Watermarking
The watermarking will be applied to outputs from ChatGPT and Codex platforms operating in the EU. However, this feature won’t be enabled automatically for all results; users of specific models will have the option to consent to watermark insertion. Only authorized researchers and specialized entities will have access to the tools needed to detect these watermarks. Importantly, the detection process will not reveal user identities, search queries, or conversation histories.
This move aligns with Article 50 of the EU Artificial Intelligence Act, which mandates that providers of generative AI systems label AI-produced content in a way that machines can recognize. New market entrants are already required to comply, while established companies such as OpenAI, Anthropic, Microsoft, Google, and Meta must implement these measures by December 2.
Embedding watermarks in AI-generated content marks a crucial advancement toward greater transparency in generative technology usage.
Such measures aim to help users clearly identify AI-created materials and reduce risks of misinformation. Overall, the EU’s new requirements are expected to drive significant changes in the AI industry, encouraging companies to adopt higher ethical standards in their development and deployment of AI tools.
As the EU implements stricter regulations on generative AI, the challenge of identifying AI-generated text is becoming increasingly complex. This evolution is largely due to advancements in language models and the limitations of current detection tools. To explore how these developments impact the effectiveness of AI content identification, read more about the ongoing challenges in AI text detection technologies.