Anthropic expands Claude text watermarking to more models

Anthropic is expanding its invisible text watermarking technology to more Claude models to comply with the EU AI Act. Starting September 30, the company will apply watermarks to three additional models: Fable 5, Sonnet 5, and Opus 4, bringing the total number of watermarked models to six. Older models are expected to be updated by December 2.

The watermarking method, inspired by Google DeepMind’s SynthID-Text, works by influencing the statistical selection of words to create a pattern that is imperceptible to human readers but detectable with the correct key. Anthropic states that this process does not alter the quality, readability, or cost of the output.

This rollout is part of a broader commitment to AI transparency and the EU Code of Practice on Transparency of AI-Generated Content. While the watermark is applied globally across all Claude platforms and APIs, it is designed to assist in provenance verification rather than identifying specific users or organizations. Anthropic also plans to provide a private-preview detection API for authorities and entities with transparency obligations.