Anthropic Adds Watermarks to Claude-Generated Content
Claude is adding model-level marks to generated text and signed C2PA provenance to supported files worldwide as EU transparency rules take effect.
Two forms of provenance
Anthropic has begun marking content produced by supported Claude models, using different mechanisms for text and files. Generated text receives an imperceptible statistical signal introduced during token selection, while supported image and vector files—including PNG, JPEG and SVG outputs—carry cryptographically signed provenance metadata based on the C2PA standard.
Supported Claude models launched on or after August 2, 2026 include the marks from launch wherever they are used, including across Claude products, APIs and supported third-party platforms. Anthropic says it plans to bring models launched before that date into compliance during a transition period ending December 2, 2026. The rollout supports the EU AI Act’s transparency requirements for synthetic content, which became applicable on August 2, but the implementation is not limited to users in the European Union. Anthropic says the text mark is designed not to alter the visible meaning or quality of an answer. Reliable detection generally requires a sufficiently long, substantially unedited sample, and broader detection access is still being developed.
The two technologies make different claims. A text watermark offers probabilistic evidence that a compatible Claude model processed a passage; it does not prove who requested it, whether Claude originated the underlying ideas or text, or whether every sentence came from Claude. C2PA credentials instead attach a signed history to a file, allowing software to verify its stated origin and whether the asset changed after signing. Metadata can be stripped by screenshots, re-encoding or unsupported platforms, while text marks can weaken through paraphrasing, translation or heavy editing. Neither mechanism is a universal detector for all AI content.
The launch therefore should not be treated as a definitive plagiarism or authorship test. Institutions using the signal for education, publishing or employment decisions would still need corroborating evidence and an appeals process. It is more useful as an infrastructure component that platforms can combine with disclosure rules, account information and content moderation.
Why it matters
Anthropic is among the first frontier-model providers to apply marking broadly to ordinary generated text, not only images. At Claude’s scale, that creates a significant real-world test of whether model-level watermarking can survive normal editing without producing unacceptable false accusations. It also shows how the EU AI Act can shape global product behavior: Anthropic is implementing the marks across supported Claude distribution channels rather than maintaining a Europe-only generation system. Its success will depend less on the existence of a mark than on interoperable verification and restrained use of the resulting evidence.