What happened
- Anthropic published a free tool at claude.com/check-content that reads the C2PA content credential Claude attaches when it generates or edits a file.
- Verification happens entirely in the browser. The file isn’t uploaded to the company’s servers.
- It accepts images, video and audio of up to 100 MB. Marking with provenance metadata covers a narrower set of formats: .svg, .png and .jpg, according to the help center.
- The company clarifies on the page itself that the tool identifies only the credential and cannot determine whether Claude was involved in creating the content.
Why it matters
- A positive result proves something. A negative one proves nothing. The credential is lost when editing, converting formats or taking a screenshot. Any editorial policy that relies on the absence of a mark is relying on thin air.
- For anyone commissioning content from third parties, this changes how the contract is worded. Requiring “no AI” is useless, because it can’t be checked. What works is requiring the original file with its metadata intact, which is the only thing that can be verified. The debate over what material those models were trained on runs on a separate track, in the courts.
- Marking applies to Claude models released since August 2, 2026. Everything produced before has nothing to verify, and retroactive application to earlier models has no schedule.
The number
100 MB. Maximum size per file. The real restriction, however, isn’t the size: an intermediate save in almost any program can strip the manifest without warning.
Context
The system combines two signals of different kinds. Files carry signed C2PA metadata, which any reader compatible with the standard can inspect, including those from Adobe Firefly or an enabled camera. Text carries a statistical watermark embedded in word choice, and for that there’s no public detector. Anthropic confirmed on August 11 that it’s preparing a detection interface, with no date or access terms published.
What’s next
- Detection of the text watermark was announced as a product in development. No date.
- No schedule for extending marking to models released before August 2, 2026.
- Article 50 of the European Union’s AI Act and its code of practice require signatories to support third-party detection. That’s the engine of the rollout, and it explains why Google and OpenAI committed to the same thing in May.
Bottom line
The standard works as long as the file travels intact, and files never travel intact. Just this week Adobe moved its production into a chat, where every download and every re-upload is a chance to lose the manifest.
