Key facts
- OpenAI will add an invisible watermark to text generated by ChatGPT and Codex in the EU.
- The watermark is required to comply with the EU AI Act's transparency rules.
- The feature will roll out to eligible EU users over the coming weeks.
- Developers can enable watermarking for select models via API globally starting immediately.
- The watermark is not visible to humans but detectable by machines using a secret key.
- OpenAI's tests show the watermark can be removed by editing text, with detection dropping significantly when 10% of words are replaced.
OpenAI announced Monday that it will begin implementing an invisible watermark on text generated by its ChatGPT and Codex models within the European Union. This move is a direct response to the EU AI Act's transparency requirements, which mandate that AI-generated content be identifiable by other systems. The rules took effect on August 2.
The watermarking feature will be gradually rolled out to eligible ChatGPT and Codex users across all plans in the EU over the coming weeks. For developers utilizing OpenAI's API globally, the option to enable watermarking for select models is available immediately, though it is off by default. OpenAI stated it is not making text watermarking a global default at this time.
The watermark functions by subtly influencing word choices, creating a pattern undetectable to the human eye but recognizable by specialized detectors. This method ensures the watermark travels with the text when copied and pasted. OpenAI asserts that the watermark does not identify individual users and has observed no significant impact on model performance when enabled.
Alongside the announcement, OpenAI released a technical report detailing its method, named textGrain. This report, co-authored with researchers from the University of Pennsylvania and Yale, explains how a secret key can be used to sort next-word predictions, enabling detectors to identify AI-generated content. However, OpenAI's tests indicate that the watermark's effectiveness can be diminished by editing; replacing 10% of words with synonyms reduced detection rates from approximately 92% to 66%. The company also noted that short passages, translated text, and mathematical answers are more difficult to watermark effectively.
Due to these limitations, OpenAI plans to grant initial access to its detector only to approved researchers and expert organizations to assess its reliability and responsible use. The company cautioned that a missing watermark does not definitively prove human authorship, as text could be too short, heavily edited, or generated by another AI system. OpenAI clarified that watermarks can indicate processing by its systems but not the extent of human involvement.
This development follows a similar announcement from Anthropic two months prior, which stated it would watermark text from its Claude model globally. That decision faced criticism from some users who felt they were the primary contributors through their instructions and editing, with Claude merely serving as a tool.
The Wall Street Journal reported in 2024 that OpenAI had previously developed a text watermark but delayed its release due to concerns that users might migrate to competitors not employing such measures. OpenAI, along with other major AI companies including Anthropic, Google, Meta, and Microsoft, has committed to adhering to the EU's code of practice concerning AI-generated content.