OpenAI announced on Monday that it will implement an invisible watermark for text generated by ChatGPT and Codex within the European Union. This new measure aims to comply with the EU AI Act, which has established transparency requirements for AI-generated content, as detailed in a blog post from the company.
The EU AI Act, effective from August 2, stipulates that AI-generated content must be identifiable by other systems. OpenAI stated that the watermarking feature will gradually roll out to eligible ChatGPT and Codex users in the EU, although developers using the company’s API globally can enable it for specific models starting today, with the default setting being off.
This watermark is designed not as a visible symbol but rather through subtle modifications in the model’s word choices, creating a pattern that remains undetectable by readers but identifiable by detection systems. The watermark propagates with text when copied and pasted and does not track user identity. OpenAI claims that incorporating the watermark does not significantly affect the performance of its models.
Along with this announcement, OpenAI released a technical report detailing their new watermarking methodology, known as textGrain. This report, co-written with experts from the University of Pennsylvania and Yale, explains how a secret key is used to influence word predictions in a passage.
While it may serve its purpose, some editing can remove the watermark. In testing, altering 10% of the words with synonyms reduced the detection rate from approximately 92% to 66%. Additionally, short texts, mathematical outputs, and translations prove more challenging to detect.
The company noted that a lack of watermarking does not necessarily indicate that a text is human-authored; various factors could contribute to this absence, including length, modifications, or the use of other AI systems. According to OpenAI, while watermarks can suggest that content has been generated or processed by their system, they cannot quantify the extent of human influence involved in its creation.
This announcement follows a similar commitment from Anthropic, which earlier this year revealed plans to watermark content generated by its AI model, Claude, raising concerns among some users about potential ethical implications.
OpenAI had previously developed a watermarking system but opted against implementation due to potential user backlash favoring rivals without similar features, according to a Wall Street Journal report in 2024. Notably, Anthropic, Google, Meta, Microsoft, and OpenAI are now signatories to the EU’s code of practice addressing AI-generated materials.
Editor’s Take
The introduction of watermarking for AI-generated content marks a significant step in the pursuit of transparency within the AI industry. This initiative not only aligns with regulatory compliance but also encourages a more responsible use of AI technologies. For users and companies alike, understanding the source and authenticity of content can aid in decision-making and ethical considerations in various applications.
Source: techcrunch.com
