AI watermarks do not replace tenant activity records
Reviewed by Helen Jones
The European Commission’s Article 50 transparency obligations began applying on 2 August 2026. Anthropic says supported Claude models launched on or after that date use machine-readable marking: embedded watermarks in generated text and signed provenance metadata on supported files. The guidance and Anthropic’s own explanation both make clear that marks are signals with limits, not a complete record of how content was used.
A provider mark can tell you that a supported model processed some content. It cannot tell an adoption lead which tenant agent, extension, or workflow produced it, what data that process could reach, or whether a person approved the result before it was sent on. That is the gap an audit or incident review will expose. The mark itself is not conclusive proof of full provenance, and an absent mark does not prove that no AI was involved. Files can be converted or changed, text can be mixed with human work, and older models have a transition path. Treat provider marking as one useful signal inside an activity and approval record you control.
Analysis
Choose one customer-facing or decision-sensitive AI workflow and record its owner, tool or agent, source data, output destination, and required human sign-off. Add any available provider provenance mark as supporting evidence, not the whole audit trail.
Source note
Pulse published by Collab365 Spaces, reviewed by Helen Jones on . Cite as "AI watermarks do not replace a tenant activity record", Collab365 Spaces.