OpenAI will add an invisible watermark to text generated by ChatGPT and Codex in the European Union. The company announced it Monday in a blog post, and the move is meant to comply with the EU AI Act. TechCrunch AI reports that the rollout reaches eligible users on all plans over the coming weeks, but only inside the EU. Outside Europe, API developers can switch it on for select models starting now. It’s off by default.
📌 Key Points
- What: An invisible text watermark built into the words ChatGPT and Codex choose
- Where: Turned on in the EU only. OpenAI isn’t making it a global default at launch
- API: Developers anywhere can turn it on for select models, and it’s off by default
- Why now: The EU AI Act’s transparency rules took effect on August 2. They require AI companies to mark generated content in a way other systems can identify
- Privacy: OpenAI says the watermark doesn’t identify the user
- Performance: OpenAI says it saw no meaningful change in model quality with the watermark switched on
- Detection access: Only approved researchers and expert organizations get the detector for now
🔍 How the Watermark Works
You won’t find a symbol or hidden character here. The watermark works by subtly shaping the model’s word choices. That leaves a statistical pattern that readers can’t see but a detector can pick up.
OpenAI published a technical report on the method, called textGrain, written with researchers from the University of Pennsylvania and Yale. The basic idea is simple. A secret key sorts the model’s next-word predictions and nudges it toward certain options. One nudge means nothing. Add up hundreds of them, and a detector can flag AI-generated text using only the text itself and the key.
The pattern lives in the words, so it survives copy and paste. That’s the big advantage over metadata tags, which disappear the moment someone moves the text somewhere else.
⚠️ The Limits OpenAI Admits
What stands out is how openly OpenAI talks about the weak spots. Its own tests show the watermark can be worn down:
- Swapping 10% of words for synonyms dropped detection from about 92% to 66%
- Short passages are harder to detect
- Math answers are harder to detect
- Translated text is harder to detect
“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” the company said.
OpenAI also warned that a missing watermark “does not prove human authorship.” The text might be too short, too heavily edited, or written by another company’s model. Even when the watermark shows up, it only tells part of the story. It “can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it,” OpenAI said.
🧭 Context: OpenAI Waited, Regulation Didn’t
This isn’t new technology for OpenAI. The Wall Street Journal reported in 2024 that the company had already built a text watermark but held off on releasing it. Part of the reason was a worry that users would leave for rivals that didn’t watermark.
The EU AI Act took that worry off the table, at least in Europe. When every major provider has to mark its output, nobody loses users for doing it. Anthropic, Google, Meta, Microsoft, and OpenAI have all committed to following the EU’s code of practice on AI-generated content.
Anthropic went further two months ago. It said it would watermark text generated by Claude worldwide, not just in the EU. That drew backlash from some Claude users, who argued they had supplied “the instructions, context, decisions” while Claude was just “the tool.” OpenAI’s EU-only approach looks like a deliberate way to avoid that fight in its other markets for now.
🎯 What This Means for You
- EU users: Expect your ChatGPT and Codex output to carry a hidden signal in the coming weeks. It doesn’t tie text to your identity.
- Developers outside the EU: Nothing changes unless you turn the watermark on in the API.
- Educators and employers: Don’t treat this as an AI detector you can lean on. Access is restricted, light editing weakens it, and a negative result proves nothing.
- Content teams: If you publish AI-assisted copy in Europe, plan for disclosure questions. Regulators now have a technical hook.
🔮 What Comes Next
The real question is whether EU-only stays EU-only. Anthropic already went global, and regulators elsewhere are watching how the AI Act plays out. If detection gets more reliable and the backlash stays manageable, watermarking could quietly become standard across the industry. Full details are available in the original TechCrunch AI report.