OpenAI will start watermarking ChatGPT's text in the EU
OpenAI will watermark ChatGPT and Codex text in the EU to comply with the AI Act. Editing can make the invisible marks harder to detect, it says.
TL;DR
- OpenAI will implement an invisible watermark on text generated by ChatGPT and Codex in the EU.
- This measure is to comply with the EU AI Act's transparency rules requiring AI-generated content to be identifiable.
- The watermark works by subtly shaping word choices, creating a pattern detectable by a system but invisible to humans.
- The watermark travels with the text when copied and pasted.
- OpenAI states the watermark does not identify the user and has no meaningful impact on model performance.
- Editing, such as word substitution, can reduce the watermark's detection rate.
- Short passages, math answers, and translated text are more difficult to watermark effectively.
- Initial detector access will be limited to approved researchers due to limitations and reliability concerns.
- A missing watermark does not definitively prove human authorship.
- OpenAI joins other major tech companies in committing to the EU's code of practice on AI-generated content.