Research note
Research note: OpenAI will add an invisible watermark to ChatGPT text in the EU
How we checked OpenAI’s EU watermark plan, its detection figures and the EU rule behind it, using OpenAI’s post, TechCrunch and the European Commission.
What we set out to check
Who gets the watermark, when, why, how well it can be detected, and who can run the detector.
What we found
- OpenAI’s post confirms EU-only for ChatGPT and Codex, opt-in for the API worldwide, and limited detector access.
- The European Commission confirms the transparency requirements apply from 2 August 2026. Its page lists delays to other parts of the AI Act, not to these rules.
Numbers that differ
None. TechCrunch’s figures match OpenAI’s post.
Labels we chose and why
- Confirmed for scope and access, from OpenAI.
- Company claim for detection rates, which come from OpenAI’s own tests.
What we could not verify
- Independent tests of the watermark.
- When the rollout will reach every EU account.
What we searched
- OpenAI watermark ChatGPT text EU
- textGrain watermark detection rate
- EU AI Act transparency obligations apply August 2026
Every source we read
| Source | Used? | Why |
|---|---|---|
| OpenAI: Text provenance in the EU | USED | Primary source for scope, API opt-in, detector access and test figures. |
| TechCrunch | USED | Independent report. Used for the stated reason (EU AI Act) and context. |
| European Commission: AI Act | USED | Official EU page. Confirms transparency requirements apply from 2 August 2026, with no delay listed for them. |