Claude to introduce machine-readable marks for AI-generated content: implications of the EU AI Act and impact on Japanese companies

Anthropic is reported to have announced through its Help Center a policy of adding machine-readable marks to Claude outputs. The policy is intended to improve the identifiability of AI-generated content by embedding watermarks in text and adding signed provenance metadata to files. It is reportedly based on the transparency obligations under Article 50 of the EU AI Act and is expected to be implemented first in new models made available in the EU on or after 2 August 2026.

The policy on Claude’s marks is said to apply not only to usage formats such as Claude Platform, Claude itself, Code, Cowork and Tag, but across all regions where Claude is available. Measures for existing models are also expected to be advanced in line with the transition period. Japanese companies using Claude internally will not immediately become subject to the obligations, but they may face requirements concerning labeling and verification when providing services for the EU market or publishing content externally.

The impact on Japanese companies is expected to include the need to organize how AI-generated material is labeled, documented, assigned editorial responsibility and handled in terms of metadata when providing services for the EU market or communicating externally. Considerations related to internal policy development and operations are also said to include labeling policies, log management and the use of contractors and other outsourced providers.

The source article states that the marks are not a “complete substitute for proof” and should be treated as supplementary information for detection. It also reports that details of the detection capabilities are expected to be disclosed in the future. Specific APIs and numerical detection thresholds are currently unknown.