Anthropic launched a tool to check whether a file came from Claude

Anthropic published a public web page at claude.com/check-content where any visitor can upload an image, audio, or video file up to 100 MB and see whether it carries a Claude-signed content credential. The tool is browser-only: the file is read locally, its embedded credentials are checked, and nothing is uploaded to Anthropic's servers. For text, a separate detection API is available in private preview for eligible organisations.

What it actually checks

Anthropic marks Claude-produced content in two ways. Images, audio, and video files get a cryptographically signed C2PA content credential attached when downloaded from Claude, "the same used by camera manufacturers and in photo-editing software", per Anthropic's page. Text output can carry an embedded watermark. The public detector on claude.com/check-content only handles the media formats: JPG, PNG, GIF, WEBP, TIFF, HEIC, AVIF, SVG, DNG, JXL for images, and MP4, MOV, AVI, WAV, MP3, M4A, FLAC for audio and video.

What Hacker News commenters found on day one

The Hacker News thread on the launch found two things quickly. Downloaded files kept their credentials in normal cases: one commenter reported asking Opus 5 for "a jpg of a cat" through the web chat, and the downloaded file did carry injected C2PA metadata. But the same commenters showed the credential is easily removed. Standard tools like exiftool strip it, and packaging the file inside a zip archive removes it during extraction on many platforms. As one commenter put it, "stripping the C2PA data is easy, but faking it is hard."

The other limit is scope. The public checker does not handle text or PDF files, which is where many practical questions about AI-written content come up. Text detection is only available through the private-preview API, open only to eligible organisations, so a teacher or an editor who suspects an essay was written by AI cannot use this page to answer their question.

For a team building a workflow that involves Claude-produced media, the checker is a first-party way to verify provenance when the credential is intact. It only works in one direction: a valid C2PA credential is strong evidence the file came from Claude, but the absence of one is not evidence the file did not, because a save-and-reupload removes it. Any team relying on this for editorial or compliance decisions should treat a passing check as reassurance and a failing check as inconclusive.

For a team producing content with Claude, the credential is now something a later reader can verify without contacting Anthropic. That is worth knowing when publishing an image or an audio clip that has to be attributable later. Whether it matters in practice depends on whether the publishing process keeps the credential from download to publication, which most compression, resize, and re-encode steps do not.

Source

Primary source: Anthropic, check-content. Community discussion: Hacker News.