adtestbench

AI news · OpenAI · regulation

OpenAI adds a hidden watermark to ChatGPT text in the EU

ChatGPT and Codex output in the EU gets an invisible statistical mark to meet the AI Act; API customers anywhere can switch it on.

By Alexandre S. , 02:30 UTC

OpenAI will mark text written by ChatGPT and Codex for users in the European Union, the company said on Monday 5 October 2026. The mark rolls out over the coming weeks on every plan, in the EU only. API customers anywhere can turn it on for some models from the same day; in the API it stays off unless they do.

Retro-futurist illustration: a striped sunset over a grid horizon under a starry sky, with a framed window standing on the horizon.
Drawn by adtestbench from “Our approach to EU text provenance rules”,

The system, called textGrain, nudges the model’s word choices to leave a statistical signal that readers cannot see. OpenAI’s own figures show how far it reaches: it found the mark in about 80% of 200-token passages and about 95% of 400-token ones. Swapping a quarter of the words for synonyms cut detection to 17%. The detector goes first to approved researchers and organisations, and OpenAI says the mark does not identify the user.

The reason is legal. The AI Act’s transparency rules, in force since 2 August, require providers to make generated text machine-readable, TechCrunch notes. It adds that Anthropic said two months ago it would watermark Claude’s text worldwide, and that OpenAI had held back a watermark it built earlier.

For teams that publish AI-drafted copy in Europe, ChatGPT text now carries a mark that a detector can find, until heavy editing wears it away.