Anthropic described on August 11, 2026, a machine-readable marking system for output from supported Claude models. Models launched in the EU on or after August 2, 2026, will embed invisible text watermarks and, for supported files, attach digitally signed provenance metadata. Anthropic links the rollout to commitments under the EU AI Act's Article 50(2) transparency code and says the marks apply worldwide across supported Claude surfaces.
Anthropic documented a machine-readable content-marking system for output from supported Claude models, combining imperceptible text watermarks with digitally signed provenance metadata for supported files. Its updated help documentation says Claude models launched in the EU on or after August 2, 2026, support marking at launch, while the company is working to add marking to models released earlier during the EU AI Act transition period.
Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. The company describes the marking rollout as part of putting those commitments into practice and says supported marks apply worldwide, not only in Europe.
Two marking mechanisms
For text, Anthropic says supported models weave an invisible watermark directly into generated output. The company says the mark does not change the text's meaning, quality, or readability, travels with copied-and-pasted text, and may persist through some editing.
For supported files, including SVG, PNG, and JPG formats, Anthropic uses signed provenance metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, standard. According to Anthropic, the metadata can indicate that Claude processed a file and can reveal tampering when the signed record remains intact.
The two mechanisms address different forms of generated material. The text watermark sits within written output, while C2PA provenance is metadata attached to a file. Downstream verification therefore depends on both the availability of detection tools and whether editing, format conversion, screenshots, or other transformations preserve the relevant signal.
Coverage and limitations
Anthropic lists Claude Platform API, Claude, Claude Code, Claude Cowork, and Claude Tag as covered surfaces for supported models. Its documentation also says embedded text watermarks apply through AWS, Google Cloud, and Microsoft Foundry, although signed file metadata may not be available on every platform or feature.
Anthropic has not yet published the detection mechanism or detailed technical documentation for the text watermark. The company says it will support users and third parties in detecting Claude marks and will publish further guidance. The Register noted that, without technical examples, the watermark's resistance to removal cannot yet be independently assessed.
Anthropic also states that a detected mark is not proof that Claude originated every idea in the material: user-written text processed through Claude could carry a mark. Conversely, an absent mark does not prove human authorship because older models, heavy editing, translation, short passages, metadata stripping, or unsupported platforms can remove or prevent a detectable signal.
For ML platform, trust, and moderation teams, the immediate distinction is between a provenance signal and a universal authenticity guarantee. The documentation describes machine-readable indicators for supported output, not a general detector for all AI-generated content.
Key Points #
- 1Anthropic uses invisible text watermarks and C2PA-based signed metadata to mark supported Claude output across text and files.
- 2The rollout covers Claude models launched in the EU on or after August 2, 2026, with worldwide marking across supported products and cloud channels.
- 3Detection details remain unpublished, and Anthropic warns that a present or absent mark is not a universal test of AI authorship.
Scoring Rationale #
The change affects Claude output across APIs, first-party products, and selected cloud distribution channels, making it relevant to teams building content-generation and moderation workflows. It is also an early operational response to EU AI Act transparency commitments. The absence of public detection specifications limits immediate implementation guidance.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.