Both models report significant performance gains in complex tasks such as agentic coding—where the AI operates as an independent assistant to complete multi-step goals—and scientific research.
The introduction of watermarking is a direct response to the requirements of the EU AI Act, marking a shift toward greater accountability in AI-generated content.
Anthropic achieves this by "nudging" the model’s word choices during the inference process, which is the stage where the AI generates a response to a prompt.
By subtly selecting specific synonyms when multiple options are available, the system creates a detectable pattern that is designed to remain identifiable even if the text is copied, pasted, or lightly edited.
While the watermarks do not contain personal user data or impact the quality of the writing, they can be identified using Anthropic’s detection API.
This tool is currently available to specific groups, including regulators, law enforcement, and fact-checkers, with plans to expand access over time.
Anthropic intends to implement this technology across its older models eventually, establishing a standard where AI-generated material can be verified by eligible organizations.