Claude's public checker is for files, not essays
The first file in the queue is still a document. A cover letter. A take-home. Someone pastes it into a "Claude watermark" box and waits for a score.
That box is reading the wrong layer.
Anthropic already shipped a public file check
Claude marks two things. Generated text gets an embedded watermark that still needs Anthropic's Detection API, which is private preview. Generated files get a Content Credential: signed C2PA metadata on the PNG, JPEG, SVG, or clip you download. Anthropic's own tool at claude.com/check-files only reads that credential. It does not decode the essay.
A writing-style classifier is not that tool. It never was.
The credential is a signature, not a vibe
When Claude exports a supported file, it can attach a small, cryptographically signed note that the file was made or processed with Claude. Verification is whether that note is present and Trusted under Anthropic's Content Credentials Root CA. No Gemini in the loop. No guess about tone.
I built the Anthropic Claude Watermark Checker API to return that result as JSON. You upload the still, the SVG, the clip. Each file becomes one dataset row: whether a trusted Claude credential showed up, whether it matches Claude, the C2PA validation state, and a limitations field that says a credential is not authorship and a missing mark is not a human.
Three steps: upload the file. Start the run. Read the credential flag next to the limitations note, not as a red card on a person.
Who this is for
If you already hold a Claude export and need a row you can store, this is for you. If you screen charts, product stills, or motion that may have come out of Claude, this is for you.
If your queue is essays, this API will refuse them. That is the correct refusal. Text watermarks are not public yet.
Do not treat a Content Credential as legal proof of who sat at the keyboard. Claude may have only converted or edited someone else's material. Screenshots and re-saves can strip the mark. Older models never attached it.
Input, pricing, and the rest of the caveats live on the Apify Store README for this checker.
For Gemini writing, I check SynthID-Text. For ChatGPT, skip the essay; I check images and audio. If the next job is disclosing the image rather than detecting the credential, I stamp official EU AI Act labels into the pixels. The other posts on rainminer cover those labs when the file is not a Claude export.
The old way was a drop zone and a screenshot
The old way is a browser tab and a PNG of the result. It does not give you a row you can keep next to the file.
If you already have the export, upload it. Store the row.