Anthropic to add watermark to AI-generated text to meet EU regulations

Anthropic announced that it will modify the way its chatbot makes small, random choices in order to meet the requirements of the European Union’s upcoming AI regulation. The company said the changes will be implemented in the next software update and are designed to increase transparency and reduce the risk of unintended bias in the model’s responses. The move comes as the EU’s AI Act moves closer to enforcement, mandating that high‑risk systems provide clear documentation of their decision‑making processes.

Key to the update is a new algorithm that limits the extent of random variation in word choice and sentence structure. Anthropic noted that typical machine‑generated tropes—such as the overuse of words like “delve,” an excess of em dashes, and the repetitive “it’s not X but Y” construction—are often a byproduct of these random choices. By tightening the randomness, the company aims to produce more consistent and less stylized output while still preserving the model’s flexibility.

With the update, Anthropic intends to demonstrate compliance with the EU’s transparency and accountability obligations, potentially setting a standard for other AI developers. The company’s statement suggests that the changes will not affect the core capabilities of the chatbot, but will instead refine its linguistic patterns to better align with regulatory expectations and user expectations for natural, varied language.