2026-10-06 / Signal #6
ChatGPT’s sentences are getting an invisible EU watermark — rewriting kills it
“Europe just put a watermark in ChatGPT’s word choices. Change a quarter of the words and it disappears.”
The Story
To meet the EU AI Act, OpenAI will invisibly watermark eligible ChatGPT and Codex text in the EU over the coming weeks via textGrain, a statistical fingerprint in word choice. API watermarking is opt-in and off by default worldwide; there is no global ChatGPT default. OpenAI’s own tests: swap 10% of words and detection falls from ~92% to 66%; swap 25% and it falls to 17%. The detector will not be public. A watermark does not prove who wrote it, how much a human edited, or that unmarked text is human.
Why It Matters
Law demanded a serial number on language. The serial number lives in synonyms, and a light rewrite launders it. Authorship becomes a regional setting.
Evidence
OpenAI (5 Oct 2026); TechCrunch; The Verge
Sources
Daily scan: 2026-10-06