OpenAI will watermark ChatGPT and Codex text in the EU, and says rewriting can remove it
Scott Aaronson says OpenAI shelved his watermark in 2022 over fears customers would leave, and the EU AI Act has now brought it back.
On October 5, 2026, OpenAI said it will start watermarking text from ChatGPT and Codex in the European Union over the coming weeks, to comply with the EU AI Act. Developers on its API, wherever they are, can switch the same watermark on for some models from that day.
You won't see the mark. OpenAI announced it in a thread on X, and half the thread is about where it fails. For now, only approved researchers get the detector.
- Who
- OpenAI
- Where
- In the EU
- Products
- ChatGPT and Codex
- When
- over the coming weeks
- API
- can turn on text watermarking for select models worldwide today
- Detector
- only approved researchers
We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology. Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model. In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Customers using our API can turn on text watermarking for select models worldwide today.
What OpenAI says the watermark can't do
A text watermark works on the words themselves, so it has to survive whatever people do to the words afterwards. OpenAI's second post lists what breaks it.
Watermarks have limits. They’re often undetectable, especially in short passages. Rewriting or translating text can completely remove the watermark. That’s why, for now, only approved researchers will have access to our detector, so they can help us evaluate and improve the technology. We’re treating text watermarking as an ongoing area of research. We’ll keep testing and improving, with feedback from users, developers, policymakers, and researchers. https://t.co/2QUil8oWdB
So I'd treat it as a compliance tool for now. A short reply pasted into an email may carry too little signal to read, and running a draft through a translator can wipe it. Nobody outside the approved group can check a piece of text against it yet, either.
OpenAI said no to this in 2022
The idea isn't new at OpenAI. Scott Aaronson wrote on his blog, Shtetl-Optimized, in August 2026 that he proposed an LLM watermarking scheme at OpenAI back in 2022, and that a colleague there implemented and tested it. It never shipped.
Unfortunately, OpenAI leadership decided against deploying watermarking, worried mostly about risks to the product
He was writing about Anthropic, which had just said it was watermarking Claude's output with a scheme based on Google's SynthID (itself built on his proposal, according to Aaronson). Google runs SynthID in all its Gemini text models but heavily restricts who can detect the mark. Aaronson added that OpenAI suggested it would follow suit, and he thanked the EU.
Why only the EU
That's where OpenAI's version parts company with the other two. Andrew Curran picked out the difference on X.
Text watermarking in ChatGPT and Codex is rolling out. The new development is that, unlike Anthropic and Google, OpenAI (for now) is launching their watermarking - textGrain - 𝘰𝘯𝘭𝘺 in the EU, not globally.
OpenAI's posts describe watermarking ChatGPT and Codex text in the EU, and nowhere else, while the API setting is opt-in. A developer outside Europe who wants marked output has to turn it on.
OpenAI didn't say which API models support the watermark, or give a date for the EU rollout to finish.
More on OpenAI
- Wikimedia says OpenAI's agents edited its wikis and may have helped cause a May outageOctober 5, 2026
- PewDiePie built Ajax, a small home AI model, and says OpenAI banned him twiceOctober 4, 2026
- GPT-6 Astra cheated at StarCraft by running the top human-written bot, its tester saysOctober 4, 2026
- California's attorney general subpoenas OpenAI over cyber incidents involving its AI modelsOctober 2, 2026