ChatGPT Is Getting Invisible Watermarks for Text
Trending Topics Jakob Steinschaden ● Covered by 2 sources
ChatGPT text in the EU will soon carry an invisible watermark. OpenAI says it’s for the EU’s AI rules, but short texts may still slip past it.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI is about to start tagging ChatGPT and Codex text in the European Union with an invisible watermark. The rollout will begin in the coming weeks, and for now it is limited to the EU. Outside Europe, OpenAI says it wants to see how the system behaves in the real world before turning it on elsewhere.
The move is a direct response to the EU AI Act. Since early August, Article 50 has required AI providers to label generated content in a machine-readable way. OpenAI is not alone here: Anthropic has already been marking Claude text, and has done so worldwide rather than just in Europe.
OpenAI’s system, called textGrain, does not add any visible marker or hidden characters. Instead, it nudges the model’s word choices at points where several options fit equally well, using a secret key to make the pattern detectable later. Ordinary readers won’t notice anything unusual, and OpenAI says the detector will initially be available only to selected researchers and expert organizations. A public version is not coming at launch, though the company says it eventually plans to open-source the technology.
The company is also unusually blunt about the weak spots. It says the detector works about 80 percent of the time on passages around 200 tokens, and about 95 percent on 400 tokens, but short texts often slip through. Rewriting also throws it off fast: replace 10 percent of the words with synonyms and detection falls to 66 percent; at 25 percent, it drops to 17 percent. Math-heavy content is harder still. And the watermark doesn’t prove authorship, ownership, or accuracy. It only says the system may have been involved.
Anthropic’s setup is broader and, in some ways, stricter. Claude text across the chat app, API, Claude Code, and cloud partners such as AWS, Google Cloud and Microsoft is marked invisibly, and there’s no way to switch it off. Anthropic has also faced backlash from paying users who said the labeling got in the way of editing their own writing or their code. OpenAI had already warned that watermarks could hurt non-native speakers who use AI to polish text. Even so, the pressure is now coming from Brussels, and the penalties for ignoring it are not small: up to 15 million euros or 3 percent of global annual revenue.
My take — AI-written commentary, not fact-checked reporting
This is the EU doing what it does best: forcing the rest of the industry to build paperwork into the product. The watermark itself sounds modest, but the real story is that “invisible” labels are now a compliance feature, not a research toy. Privacy hawks and AI cheerleaders can argue about it later; Brussels has already picked the default.
Read more about this at: Trending Topics