Anthropic is taking a scorched-earth approach to AI content labeling. The company confirmed this week that all new Claude models will embed machine-readable watermarks into text outputs from day one, with digitally signed provenance metadata attached to other generated files where supported. The rollout is global, not just European, and it’s driven by the EU’s AI Act, which mandates watermarking for AI-generated or manipulated content for any model released after August 2, with a grace period stretching to December 2026 for older systems.
The catch? Anthropic isn’t distinguishing between content Claude generates wholesale and content it merely touches. The EU’s own guidance explicitly exempts assistive functions like grammar correction or edits that don’t substantially alter meaning. But a watermark applied at the model level can’t tell the difference between a novel written from scratch and a comma splice fixed in an email draft. As Anthropic put it in their support article, text outputs will “carry embedded watermarks” invisible to users, and the company is applying this to all processed content where the technology supports it.
Nobody outside Anthropic can verify how aggressively these watermarks are actually being applied yet. The company says it plans to release a detection tool and share technical details on how to spot the marks, which the EU law requires providers to offer. Until that tool ships, we’re taking Anthropic’s word that a two-word typo correction gets the same invisible stamp as a 2,000-word essay.
For non-text outputs, Anthropic will lean on the C2PA standard, the same provenance framework that Adobe, Microsoft, and others have been pushing for images and video. The company also acknowledged the watermarks won’t work on every platform or feature that lacks support. That’s a meaningful hole — content pasted into an unsupported app loses its digital fingerprint, which undercuts the entire point of provenance tracking. The irony is hard to miss: a law designed to flag synthetic media may end up labeling your corrected typos as AI-manipulated content while letting genuinely problematic outputs slip through on unsupported platforms.
