Anthropic has formalized its commitment to the EU AI Act’s Article 50(2) Code of Practice, outlining a comprehensive framework for embedding transparency into its generative AI models. The policy mandates that all Claude models released on or after August 2, 2026, must integrate machine-readable marking by default, with ongoing efforts to retrofit support for legacy models.

Core Methodology

Anthropic employs a dual-layered technical approach to marking, categorized by the modality of the output:

1. Embedded Text Watermarking:

The system utilizes an imperceptible watermarking mechanism applied directly at the model level. This process weaves a statistical signal into the generation sequence without altering the linguistic quality, semantic integrity, or readability of the output. Because the signal is intrinsic to the token distribution, it remains persistent during copy-paste operations and minor textual modifications, facilitating cross-platform provenance tracking.

2. C2PA Provenance Metadata:

For non-textual outputs—specifically supported file types like .svg, .png, and .jpg—Anthropic utilizes the Coalition for Content Provenance and Authenticity (C2PA) open standard. This involves appending cryptographically signed metadata to the file headers. This methodology allows third parties to verify the origin of a file and detect unauthorized post-generation tampering or alterations by validating the digital signature against the original provenance records.

Implementation and Scope

The marking infrastructure is designed for universal application across the Claude ecosystem, including the Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. Integration extends to cloud partnerships (AWS, Google Cloud, Microsoft Foundry), ensuring that machine-readable signals remain embedded regardless of the hosting environment.

Limitations and Detection

Anthropic acknowledges inherent constraints in the efficacy of these signals:

• Non-Conclusivity: The presence of a mark confirms Claude’s involvement but does not equate to primary authorship, as models are often used for auxiliary tasks (e.g., summarization or translation of external content).
• Signal Degradation: Marks may become undetectable due to heavy post-processing, such as aggressive paraphrasing, re-saving files (stripping metadata), taking screenshots, or generating text sequences too brief to sustain a statistically significant signal.
• Absence of Marks: The lack of a detectable mark does not constitute a definitive "human-only" verification, as it may result from unsupported legacy models, format conversion, or specific platform-level technical limitations.

Anthropic intends to provide comprehensive technical documentation to enable third-party detection and support developers in meeting their own transparency obligations under the EU AI Act.