OpenAI's Watermarks Are Only as Good as Who Can Check Them
OpenAI published how it's approaching text watermarking under EU rules, including where watermarks apply and how detection works. The detail I'd underline is that detection access starts with researchers. That's a sensible way to test a new technique, but it means most enterprises can't verify provenance themselves yet. In the privacy programs I've built, the control nobody could actually test was always the one that failed the audit. So if your AI policy says you label generated text, ask who checks it and with what tool. A watermark nobody can read is still just a promise. Who in your organization owns proving it?
Source: Our approach to EU text provenance rules (OpenAI)