Anthropic has released a free tool with the Content Checker at claude.com/check-content that checks files for embedded Claude watermarks. The application reads signed C2PA metadata from image, video, and audio files and indicates whether a file was created or edited with the AI assistant Claude. According to Anthropic, all checks run exclusively locally in the browser, without the file being uploaded.
Tool checks files directly in the browser
The Content Checker accepts images in JPG, PNG, GIF, WEBP, TIFF, HEIC, AVIF, SVG, DNG, and JXL formats, videos in MP4, MOV, and AVI formats, as well as audio files in WAV, MP3, M4A, or FLAC, each up to 100 megabytes. As Anthropic explains in its own support center, a checked file never leaves the user’s device; this assurance cannot currently be independently verified, as the source code of the tool is not open. A positive result thus only confirms that Claude has processed the file – not how large the AI contribution to the result is, as self-created files can also be submitted for editing or conversion. For everyday use, this means: if someone wants to check a suspicious press photo or an AI-generated graphic, they get an initial assessment but no forensic analysis. The labeling announced in August for image files relies on the open C2PA standard of the Coalition for Content Provenance and Authenticity, which is also used by camera manufacturers and image editing software.
Labeling implements EU transparency obligation since August
The basis remains Article 50 paragraph 2 of the EU AI Regulation, which has made machine-readable labeling of AI-generated content mandatory since August 2, 2026. Anthropic is among around 190 signatories of the voluntary EU Code of Conduct for Transparency of AI Content, which the company signed in July 2026. Violations of the labeling obligation can be punished with fines of up to 15 million euros or three percent of global annual revenue. Currently, the models Claude Fable 5.1 and Claude Mythos 5.1 support the marking from the start; for older models, Anthropic is gradually retrofitting the function according to its own statements. As an Anthropic engineer explained to TechCrunch back in August, only a complete rephrasing of each word reliably removes the text watermark. The labeling applies worldwide and across all access paths – the Claude web application, the Claude Platform API, Claude Code, Claude Cowork, and Claude Tag, as well as cloud partners AWS, Google Cloud, and Microsoft Foundry. For users in Germany and the EU, there is no change in access: the marking runs automatically in the background, and a separate tariff or opt-in is not provided.
Text recognition remains reserved for selected organizations
Unlike the free file check, the recognition API for text watermarks remains temporarily restricted in access. Anthropic provides the so-called Watermark Detection API in a private preview only to selected organizations: regulatory authorities, law enforcement, media, fact-checkers, independent research institutions, educational institutions, and civil society groups from the EU are entitled to access under EU law. Companies that are themselves obligated to comply with the AI Regulation can also apply for it. Editorial offices or private individuals who want to check a single text for Claude origin still have no direct way to do so – only the file check for image, video, and audio content is available to everyone, completely without registration or proof of affiliation. Comparable checking tools for their own markings are now also offered by OpenAI and Google DeepMind, each only for their own models and without cross-provider matching. Those who want to apply for access to the text recognition API as an authority, editorial office, or research institution must fill out their own registration form and prove their affiliation with one of the mentioned groups.
It remains open how reliable a negative result appears in practice: if the marking is missing, it cannot be concluded from this whether there is a human origin or another AI model. Therefore, it will be crucial whether editorial offices, platforms, and courts will accept the free file check in the future as a reliable indication or wait for a broader text recognition, whose access remains temporarily limited to a few organizations.
