Archive — US — techcrunch
OpenAI will start watermarking ChatGPT’s text in the EU
OpenAI will implement an invisible watermark for text generated by ChatGPT and Codex in the EU to comply with the EU AI Act, ensuring transparency and content identification. The watermark subtly alters word choices and will be available to users in the EU over the coming weeks. Developers worldwide can opt-in for specific models, but the feature is off by default. Limitations on detection suggest careful monitoring is needed without compromising user anonymity.
Key points
- OpenAI introduces invisible watermarking for ChatGPT and Codex in the EU to comply with the EU AI Act.
- Watermark is designed to be transparent yet detectable by specific systems, allowing content identification.
- Will roll out for eligible EU users over the next weeks; API developers globally can enable for certain models, off by default.
- Watermarking subtly modifies word choices, creating patterns unreadable by humans but detectable.
- Removal of watermark possible with significant editing, leading to discussion about detection reliability.
- Initial detector access limited to approved researchers to evaluate the watermark's reliability.
- OpenAI emphasizes that a missing watermark does not confirm human authorship since many factors affect detection.
Why it matters
- Text watermarking aligns with the EU's AI regulation efforts, focusing on responsible AI usage.
- Reflects growing trend among tech companies like Anthropic and Google to develop AI transparency mechanisms.
- The EU AI Act seeks to establish clear guidelines for AI-generated content, which may shape global standards.
- OpenAI's initiative may influence competitive dynamics in AI tools as users consider watermarking policies.
OpenAI's introduction of watermarking for its AI-generated text in the EU highlights the balancing act between compliance and user experience, setting a precedent for AI accountability while sparking debates on transparency and authorship.