News · 2026-09-02

Anthropic shipped a content checker that cannot tell you if Claude wrote it

Anthropic released a free, public content checker at claude.com/check-content, and the most important sentence on the page is a disclaimer. The tool "identifies only the content credential," Anthropic writes, and "it can't tell whether Claude was involved in creating the content." It is a provenance reader, not an AI detector, and the distinction is the entire point of the product.

Key facts

• The checker is live at claude.com/check-content and reads C2PA content credentials attached to files.
• It runs locally in the browser: "Your file never leaves your device," per the page, with a 100 MB limit.
• It supports 17 image, video, and audio formats including JPG, PNG, SVG, MP4, MOV, WAV, MP3, and FLAC; plain text is not covered.
• The EU AI Act's Article 50 transparency obligations became applicable August 2, 2026.

The confusion this tool is built to avoid has been running for three years. People want a button that answers "was this written by AI," and a long parade of products has claimed to provide one, mostly by looking at statistical properties of text and guessing. Those tools produce false positives on non-native English writers, on formal prose, and on anything edited enough to smooth out its rhythm. They have cost students grades and writers contracts.

A content credential works from the opposite direction and makes no guesses at all. When a file is generated, the generating system attaches signed metadata recording what made it, using the C2PA standard that a coalition of media and technology companies developed. Checking a credential is a cryptographic verification, not an inference. If the tag is present and valid, you know precisely what it says. If there is no tag, you know nothing whatsoever. That is a much smaller claim than "this was AI-generated," and it is the only kind of claim that is actually reliable.

This is why the disclaimer matters more than the feature. A file that came out of Claude and then went through a screenshot, a format conversion, or any of a dozen ordinary editing steps loses its credential, and the checker will report nothing. A file that never touched an AI system also reports nothing. The tool cannot distinguish those cases, and Anthropic says so on the page rather than in a footnote. That is a rare piece of product honesty in a category built almost entirely on overclaiming.

Anthropic draws a second line in its support documentation, between two different signals it produces. Generated text gets an imperceptible watermark embedded in the text itself, which the company says travels when text is copied and pasted and may survive some editing. Generated files get signed provenance metadata attached externally. The public checker only inspects the second one. For the text watermark there is a separate Detection API, still in private preview, which means the signal most people would want to check is the one they cannot check. Anthropic also lists the failure modes for the text watermark candidly: heavy editing, paraphrasing, translation, mixing with other writing, very short passages, and any process that strips file metadata.

The regulatory context explains the timing. The European Union's AI Act sets transparency obligations under Article 50 that became applicable on August 2, 2026, and Anthropic states that Claude models launched on or after that date support machine-readable marking at launch, naming Fable 5.1 and Mythos 5.1 as currently supported. This is compliance infrastructure shipped as a consumer tool, which is the usual way these things arrive. Anthropic previously added the text watermark and pointed at the same EU deadline.

Two honest caveats. First, Anthropic publishes no false-positive rate for either the file checker or the private-preview text detector. For a cryptographic credential check that is arguably fine, since verification either succeeds or it does not, but the absence of a published figure for the text detector is a gap worth noting before anyone builds a policy on it. Second, and more fundamental: provenance marks are removable by anyone who wants them removed. A tool that strips SynthID and C2PA marks passed 4,900 stars on GitHub earlier this year. Content credentials are designed to survive ordinary handling, not deliberate attack.

Which leaves the real use case, and it is narrower than the headlines about AI detection suggest. This is a tool for confirming that a file is what it claims to be when someone is cooperating, in a newsroom checking a supplied image, or a platform verifying an upload from a publisher who wants provenance preserved. It is worthless against anyone determined to hide, and it says so. In a field where provenance and watermarking claims routinely outrun what the technology can do, shipping a tool alongside an accurate description of its limits is the notable part.

Primary source, verified: read the paper →

Ground Truth. (2026, September 2). Anthropic shipped a content checker that cannot tell you if Claude wrote it. Ground Truth. https://groundtruth.day/news/anthropics-content-checker-cannot-tell-you-if-claude-wrote-it.html

@misc{groundtruth:anthropics-content-checker-cannot-tell-you-if-claude-wrote-it,
  title  = {Anthropic shipped a content checker that cannot tell you if Claude wrote it},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/anthropics-content-checker-cannot-tell-you-if-claude-wrote-it.html}
}

Topics: provenance · watermarking · c2pa · anthropic · regulation · eu-ai-act · tools