The Label Is a Clue

Claude can now leave a machine-readable trace in supported output. The trace may survive copying or travel inside a file. It still cannot tell a reader who supplied an idea, whether a claim is true, or who accepted responsibility before publication.

That distinction moved from a technical argument into an operating requirement this month. Transparency duties under Article 50 of the EU AI Act began applying on August 2. The rules require providers to mark certain synthetic outputs in machine-readable form. They also require deployers to disclose certain deepfakes and public-interest text produced without human review or editorial control. The European Commission’s guidance says substantive human review must involve deliberate examination by a qualified person. A superficial grammar pass does not meet that standard.

Anthropic has now explained how it is implementing the provider side. According to its Help Center, supported Claude models launched in the EU on or after August 2 attach imperceptible watermarks to text. Supported SVG, PNG, and JPG files use signed C2PA metadata. Anthropic says the approach applies globally across supported Claude products, although coverage varies by surface, feature, and file type. Its detector documentation is still forthcoming.

The most useful part of Anthropic’s explanation is its restraint. A detected mark indicates that material may have been processed by Claude. It does not establish that Claude originated the work or that the material is accurate. An absent mark proves little too. Older models, substantial edits, translation, screenshots, file conversion, unsupported surfaces, and stripped metadata can all break the trail.

OpenAI describes similar limits in its Article 50 implementation note: uploads, downloads, resizing, screenshots, and format changes can remove C2PA metadata. C2PA’s own documentation describes watermarking and fingerprinting as ways to help rediscover credentials after metadata disappears. That is a design objective rather than a universal guarantee.

The emerging system therefore offers evidence, not a verdict. A publisher may receive a document drafted by a person, extended by a model, rewritten by another person, formatted by a second tool, and uploaded through a platform that drops its metadata. A binary “AI or human” label flattens that history into an answer the evidence cannot support.

Visible labels have limits as well. The Commission’s optional icon guidance, updated August 10, says an icon alone does not establish compliance. Its user testing found that an icon worked better when paired with explanatory text. The guidance also calls for labels to remain visible when material is reshared or downloaded. That is a practical admission that disclosure has to survive distribution, not merely exist at the moment of creation.

Verification bottleneck

Verification is becoming the scarce institutional function.

Generation and editing moved faster than provenance systems can preserve the full chain.
Publishers, platforms, compliance teams, and readers now have to distinguish evidence of model processing from claims about authorship, accuracy, and responsibility.
Watch for Anthropic’s detector documentation, independent robustness tests, and whether content platforms preserve credentials through ordinary upload and export workflows.

A durable record needs more than a detector result. It should connect the source material, model or tool used, version history, disclosure applied, reviewer decision, and published artifact. Machine-readable marks can strengthen that chain. They should occupy one place inside it.

Opportunities

Where value may appear is in practical provenance work. A small publisher could add a review ledger to its CMS. A verification service could test whether credentials survive common editing and distribution paths. A toolmaker could pair C2PA checks with source attachments, version history, and named human approval. A consultant could help organizations define what substantive review means for their own high-consequence material.

The useful product will resist the temptation to declare content clean or contaminated. It will show what evidence exists, what disappeared, who checked the work, and which questions remain open.

The Label Is a Clue

Verification bottleneck

Opportunities

Sources