Claude’s new invisible watermark is here – for now
▼ Summary
– Anthropic will watermark content processed by its AI models, not just generated content, to comply with the EU AI Act.
– New models released globally will include watermarks from day one, with text outputs carrying invisible embedded watermarks and other files using digitally signed C2PA provenance metadata.
– The watermarks apply to all processed content where supported, even though the EU exempts cases like standard editing or grammar correction that don’t substantially alter text.
– The effectiveness of the watermarking is uncertain until Anthropic releases a detection tool, which it plans to do to meet EU technical support requirements.
– Watermarks won’t function on some platforms or features that lack support for them.
Anthropic is taking a sweeping approach to content attribution, announcing that it will embed invisible watermarks into all content processed by its AI models, not just content they generate from scratch. The company detailed the rollout in a support document, framing it as a direct response to the European Union’s AI Act, which mandates that providers mark AI-generated or manipulated media across text, audio, image, and video. The regulation applies to models released after August 2, with a grace period for older systems extending until December 2026.
The company confirmed that every new model released globally, regardless of the user’s location, will carry these markers from launch. Text produced by Claude will include hidden watermarks that remain imperceptible to the reader, while other file types will incorporate digitally signed provenance metadata where the technology permits. Anthropic’s stated goal is to make AI-assisted content traceable at scale.
What stands out here is the strategy’s breadth. The EU’s rules carve out exceptions for cases where AI merely assists with standard editing, such as grammar fixes, or when it does not substantially alter the user’s original text. Anthropic, however, is choosing to apply watermarks to all processed content where technically feasible, a policy that effectively ignores those exemptions. Because a model-level watermark cannot distinguish between wholesale generation and a minor punctuation tweak, Claude may end up tagging content the law was designed to leave untouched. The true scope of this coverage will remain unclear until Anthropic releases a detection tool for independent testing. The company has indicated it will eventually publish details on how to identify the marks, a step required to provide the technical support the EU legislation demands.
There are practical limits to the system. Anthropic acknowledged that some platforms and features will not support the watermarking, and for non-text outputs, it will rely on the C2PA metadata standard to log provenance.
(Source: Ars Technica)
