OpenAI rolls out text watermarking under the EU AI Act: on by default for EU ChatGPT and Codex, opt-in for the API, detector limited to researchers

Why it matters

Key points

On October 5, 2026 (15:00 GMT), OpenAI published "Our approach to EU text provenance rules," explaining how it will make generated text machine-identifiable under the EU AI Act.

The rollout has three parts. Starting today, API customers worldwide can opt in to text watermarking for select models; it stays off by default in the API. Over the coming weeks, eligible ChatGPT and Codex text output in the EU will carry an invisible watermark across all plans, but this will not become a global default. Applications for the watermark detector open today, with access initially limited to approved researchers and expert organizations. Image and audio verification tools (openai.com/verify and the Content Provenance API) remain publicly available.

The technology, called textGrain, adds an invisible statistical signal to the model's word choices, and a detector checks for that signal. OpenAI says it matched or exceeded other approaches it tested, including SynthID for text, plans to open-source it, and saw no meaningful benchmark differences on Astra with and without watermarking.

Verifiable

Primary page: https://openai.com/index/eu-text-provenance (RSS pubDate Mon, 05 Oct 2026 15:00:00 GMT, category Safety). OpenAI's self-reported numbers: at a 1% target false positive rate, detection was about 80% for 200-token passages and about 95% for 400-token passages on psychology-style content, and substantially lower for mathematics. Replacing 10% of words with synonyms cut detection on 400-token passages from about 92% to 66%; replacing 25% cut it to 17%. These are internal evaluations, not yet independently reproduced.

Our take

This is the first time a frontier lab has put text watermarking into a large-scale production product, driven by regulation rather than technical maturity, so it goes to Latest. The key number is OpenAI's own: rewrite a quarter of the words and detection falls to 17%. That makes a text watermark a weak signal and a compliance artifact, not a tool for deciding who wrote something or how much, and certainly not for AI-cheating verdicts in schools or hiring. Gating the detector to researchers and spelling out what a watermark cannot tell you is OpenAI actively heading off that misuse.

It also splits the product by region: EU ChatGPT output is watermarked, other regions' is not, and with the API off by default most third-party apps will not carry a watermark unless local rules force it. Two things to watch: whether outsiders can reproduce these detection figures once textGrain is open-sourced, and whether Google, Anthropic and others adopt similar measures in the EU.