Text watermarking is fragile by design; provenance may be the real fix, argues Tansu Yegen

TansuYegen · x · 2026-10-06

Tansu Yegen weighs in on OpenAI's new EU text watermarking scheme: while it can detect patterns in generated text, OpenAI's own testing shows detection weakens as users edit the output — words are simply too easy to change. He argues the long-term answer may not be watermarking content at all, but proving the origin and editing history of information before it reaches us.

Related event: OpenAI launches textGrain text watermarking, starting in the EU(25 posts)→

Original post →

More from Safety

Safety channel →