OpenAI adds invisible text watermarks to ChatGPT in the EU — API users can opt in worldwide

#What happened

The EU AI Act requires machine-readable labeling of generated text. OpenAI's answer is textGrain, which embeds an invisible statistical signal in word choices, similar to Anthropic's SynthID-based watermark. The watermark rolls out to ChatGPT and Codex users in the EU "over the coming weeks." API customers can opt in worldwide starting today, including via cloud partners such as Microsoft Azure. OpenAI says textGrain matched or beat other approaches in internal tests and plans to open-source the technology.

#What the tests show

At a 1% false-positive target, detection hit ~95% on 400-token psychology passages and ~80% at 200 tokens, but was "substantially lower" for math content. Replacing just 10% of words with synonyms cut 400-token detection from ~92% to ~66%; replacing 25% dropped it to ~17%, making the watermark easy to evade. OpenAI says watermarking doesn't hurt output quality on frontier model Astra across eight benchmarks (including GPQA Diamond, BrowseComp, and DeepSWE), though writing-quality impact is not established.

#Why it matters

Access to the detector is restricted to selected researchers via an application form; it reports only whether a watermark was detected, not who wrote it. A detected watermark reveals nothing about authorship or user identity. This puts OpenAI on a different path from Anthropic, whose watermarking for Claude is mandatory globally.

Verified October 5, 2026.

#SOURCES

• the-decoder.com: OpenAI will watermark ChatGPT text in the EU
• Bloomberg Tax: OpenAI to add text watermarks for EU ChatGPT users