Invisible markers in AI-written text are becoming standard as Europe's 2 December deadline draws near

Since Saturday, text generated by seven Grok models through xAI's API carries an invisible watermark by default. The mark is embedded in the model's own choice of words and adds no visible characters, so copying and pasting does not remove it, DataStudios reports. Developers can see in a separate header whether a response has been marked. Only a team administrator can switch it off, and must confirm a legal exemption to do so.

xAI is not alone. According to an overview by Eden AI, Anthropic extended marking on its own API to four older Claude models on 4 November, with no option to turn it off. Since October, OpenAI has offered a watermark that developers have to switch on themselves, while text from ChatGPT and Codex in the EU is marked by default. For Mistral, Meta and most other providers, Eden AI found no public information about a watermark.

The trigger is Article 50 of the EU AI Act, which requires providers of generative AI to mark their output in a machine-readable way. That obligation has applied since 2 August, but systems already on the market before then have until 2 December. Breaches of the transparency rules can cost up to 15 million euros or 3 percent of worldwide annual turnover, according to DataStudios. No figures on how reliable detection is have been published yet.

Transparency: this edition was compiled by Claude, made by Anthropic.