Anthropic has expanded its text watermarking rollout to several older Claude models to meet the upcoming EU AI Act deadline in December 2026. However, API users can easily bypass the watermarking system on models like Claude Opus 4.6 and Claude Sonnet 4.6 by simply tweaking sampling parameters such as temperature, top_p, and top_k away from their default values. The watermark relies on an imperceptible word choice pattern using a variant of Google DeepMind's SynthID method, which leaves output quality, pricing, and formatting unaffected.
The staged rollout includes updates for models such as Claude Opus 4.5, Claude Sonnet 4.5, and Claude Haiku 4.5, with Amazon Bedrock deployments experiencing minor delays. While newer models enforce the watermark unconditionally, the loophole on older versions exists because the detection pattern only activates when sampling parameters remain at default settings. Anthropic has warned developers relying on compliance mechanisms to avoid changing these specific configuration values.
This watermarking initiative highlights the technical challenges AI companies face when retrofitting legacy generative models to comply with evolving regional regulations.