AI Watermarks Need a Verification Process

AI watermarks are becoming a topic for communications, compliance, and business teams. Anthropic has announced text marking for future Claude models and cites EU AI Act requirements. For enterprises, this does not create automatic truth checking; it creates a new task in the approval process.

What a marker can do — and what it cannot

According to Anthropic, the method leaves a statistically detectable pattern in generated text without adding visible characters or generating extra tokens. A later test can assess the likelihood that Claude was involved in a text. That is useful for transparency, but it is not comprehensive proof of origin.

The marker does not say whether a text is factually correct or who is accountable for it. It also cannot reliably distinguish fully generated text from a heavily edited draft. Its evidential value is limited for short passages, proofreading, or code. A complete rewrite can also remove the signal.

For files, Anthropic also refers to C2PA Content Credentials. The open standard describes signed metadata and validation to make an asset’s origin and processing traceable. This information does not replace subject-matter approval either: it provides context about what happened to an asset.

Technology becomes a review path

The key management question is therefore not, “Is the text marked?” It is, “What decision follows from a verification result?” A practical process separates four layers:

Origin: Was an AI signal or Content Credential found?
Content: Have facts, figures, and sources been independently checked?
Accountability: Which person or role approves external content?
Archive: Which version, check, and approval are stored for later review?

This prevents two wrong reactions: treating AI markers as blanket proof or ignoring them completely. Especially for proposals, contract drafts, customer communications, and published expert content, a detection should trigger review rather than automatically block publication.

What DACH enterprises should do now

Start with a small, clearly defined area, such as external marketing or service responses. Define where AI use must be disclosed, which checks apply before publication, and who decides exceptions. Add version storage and a source check to the workflow — regardless of whether a provider makes its watermark or C2PA metadata technically available.

This turns transparency from another checkbox into a dependable part of quality and accountability.