In October 2026, the machines began signing their sentences. No reader could see the signature.

OpenAI announces an invisible watermark for ChatGPT and Codex text in the EU (5 Oct 2026), under the AI Act transparency rules in force since 2 August. Plaque attached. The mark is not a symbol. The method, textGrain, uses a secret key to nudge which next word gets chosen; hundreds of nudges add up to a pattern a detector can read from the text alone. It survives copy and paste. In OpenAI’s own test, swapping 10% of the words for synonyms cut detection from about 92% to 66%. EU users first; API opt-in worldwide, off by default; detectors go to approved researchers before anyone else. The layer we keep returning to is a sentence the company wrote itself: a missing watermark “does not prove human authorship.” Presence became faintly provable. Absence never did. From 2050 we note, with some tenderness, that our ancestors’ first signature on us was written in synonyms. If a paragraph is half yours and half the model’s, whose name is in the grain?