OpenAI is rolling out an invisible, machine-readable watermark in text output for ChatGPT and Codex, starting with users in the European Union. The Verge reported the launch on 5 October 2026, and the change is driven by the EU AI Act rather than by a product decision. If you write with these tools, the practical question is not whether the text reads differently, because it does not, but what a watermark can and cannot prove.
Key facts
| Fact | Detail |
|---|---|
| Question answered | OpenAI Is Adding Text Watermarking to ChatGPT and Codex |
| Short answer | OpenAI is rolling out textGrain, an invisible machine-readable watermark, for ChatGPT and Codex users in the European Union, with worldwide opt-in for API customers and detector access for approved researchers. |
| Reading time | ~3 min |
| Last updated | 2026-10-06 |
What is actually shipping
- EU-first rollout. Over the coming weeks OpenAI will introduce text watermarking to eligible ChatGPT and Codex users across all plans, in the EU only. OpenAI states plainly that watermarking will not be the global default at launch and describes the regional approach as room to learn from real-world use.
- Opt-in API watermarking. From now, API customers worldwide can opt in to watermarked text outputs for select models. It is a switch, not a mandate.
- Cloud partners later. OpenAI says it is working with cloud partners to make watermarking available for model outputs accessed through their services in the coming weeks.
- Detector access by application. Approved researchers and expert organisations can apply from now, granted initially case by case. The detector reports whether it finds an OpenAI watermark without identifying the user or revealing prompts or conversations.
What the watermark claims, and what it does not
OpenAI says its textGrain approach matched or exceeded other schemes, including Google DeepMind’s SynthID for text, which is also the basis for the watermarking Anthropic announced in August. It published benchmark scores showing similar performance for watermarked and unwatermarked text, which is the part that should matter most to anyone worried the mark degrades output quality.
The disclaimer is the more important half. OpenAI says textGrain does not guarantee reliable detection, and that a text watermark does not verify accuracy, does not establish who owns the text, does not measure how much a human contributed, and does not prove human authorship. OpenAI’s own framing acknowledges the risk of missed watermarks and false positives.
What it means if you publish AI-assisted writing
- If you are in the EU and draft with ChatGPT or Codex, some output may carry a watermark. It is invisible and machine-readable, so readers see no difference.
- Watermarking is not a disclosure regime. It does not tell a reader that AI was involved, and it does not certify authorship in either direction.
- If your compliance obligations call for labelled or watermarked output, the API opt-in is the lever you control today.
- Do not read no watermark detected as a clean bill of health. Detection is probabilistic by OpenAI’s own description.
For teams running a content pipeline with a desktop agent, the practical consequence is that provenance tooling is arriving per region and per plan, so the same workflow can produce differently marked output depending on where the account sits. Keep human review where it already mattered, which is accuracy and attribution, and treat watermark status as a separate compliance checkbox rather than a quality signal.
MOOGH runs agent tasks on your own Windows machine with the plan shown and approved before execution, which keeps a fixed, inspectable path from draft to file. If your publishing workflow needs that, you can download the client.
FAQ
- Does the watermark change the text? OpenAI published benchmark scores indicating comparable performance between watermarked and unwatermarked output, so the visible behaviour should be unchanged.
- Will this reach other regions? OpenAI has not committed to a global default. It describes the EU launch as a regional step and says cloud partners will be able to offer watermarked outputs in the coming weeks.
- Who can use the detector? Approved researchers and expert organisations, on a case-by-case basis, with results that do not identify the user or reveal prompts.
- Should publishers label AI-assisted text because of this? Not because of the watermark itself. A watermark is provenance machinery, and OpenAI explicitly says it is not proof of authorship or accuracy.
Related reading: