ChatGPT will start hiding a watermark in its writing for users in the EU
OpenAI says it will add an invisible watermark to text written by ChatGPT and Codex for users in the European Union over the coming weeks, to meet the EU's AI Act. The company is not making it the default elsewhere, and the tool that checks for the watermark is not open to the public.
AI-assisted · US Brief desk · Sources listed below
What happened
OpenAI said Monday it will add an invisible watermark to text written by ChatGPT and its coding tool Codex for users in the European Union. The rollout covers eligible users on all plans and will happen over the coming weeks. OpenAI said it is not making text watermarking a global default for now.
The move is meant to meet the EU's AI Act. Its rules on marking AI-made content took effect Aug. 2, 2026, according to the European Commission. OpenAI is one of about 190 signers of an EU code of practice on labeling AI content.
The watermark, called textGrain, does not add hidden characters or change how text looks. Instead, it nudges which words the model picks, leaving a statistical pattern a detector can find. OpenAI says the pattern can survive light edits and copy-and-paste, but heavy rewriting, paraphrasing or translation can erase it.
More context
There are limits. OpenAI says short answers, code and very precise answers, such as math, are hard to watermark. The EU code does not require marks on outputs under about 150 words. In OpenAI's tests across EU languages, detection rates at a 1% false-positive rate ranged from 69% in Spanish to 42% in Romanian before tuning.
There is no public checker for text. OpenAI says only approved researchers and expert groups can apply to use the detector, case by case, because of the risk of missed watermarks and false positives. It also says a watermark does not prove who wrote something or how much a person contributed.
Outside the EU, businesses that use OpenAI's API can now opt in to watermarking for select models. OpenAI says it tested watermarked and plain text and found no drop in quality. The Verge reported that Anthropic added similar watermarking in August to meet the same EU rules.
4 listed sources
- OpenAI Help Center, updated Oct 5, 2026: Provenance signals in OpenAI-generated content
- OpenAI, Oct 5, 2026: Our approach to EU text provenance rules
- The Verge (Stevie Bonifield), Oct 5, 2026: OpenAI is adding text watermarking in ChatGPT and Codex
- European Commission: Code of Practice on Transparency of AI-generated Content
References listed by US Brief; a source count is not a verification score.
Editorial sourcing notes
This story is based on OpenAI's own help-center explainer and the European Commission's Code of Practice page, with rollout details from The Verge. Detection rates are from OpenAI's own tests. OpenAI says watermarks do not prove who wrote a text or how much a person contributed.
Like US Brief? You can support it with a tip.