OpenAI is rolling out a watermarking system for AI-generated text to meet the EU's incoming provenance rules.
The company laid out how the system works. Certain text outputs get an invisible pattern embedded in them. A separate detection tool can then check whether a passage carries that pattern. Access to that detection tool is rolling out in stages, starting with outside researchers rather than the general public or media outlets.
That staged rollout matters more than the watermark itself. Text watermarking is notoriously fragile: unlike marks on images or video, a single paraphrase, translation, or copy-paste edit can strip the pattern entirely. Limiting early access to researchers lets OpenAI stress-test the detector's accuracy before anyone can use it, or abuse it, at scale.
Still, it's a strange look for a tool meant to build public trust: the people most likely to want to check a suspicious bit of text (journalists, teachers, fact-checkers) are exactly the group waiting longest for access.