Identify the artifact first when checking for a Claude watermark. Anthropic’s public file checker can inspect supported images, video, and audio for a Claude-issued Content Credential. It does not check text. Eligible organizations currently reach Claude text watermark detection only through a private-preview API. A public website claiming to find that exact text mark cannot independently verify its result against Anthropic’s private key.
That distinction matters in a document dispute. A file credential, a statistical text watermark, and an ordinary AI-writing score answer different questions. None alone establishes who wrote the underlying ideas.
Choose the right check for the material
Claude applies marks through two channels. Anthropic describes, for supported text, a pattern in the words the model chooses. For supported files, Claude attaches signed C2PA Content Credentials to the file. Copying words into a document does not turn the text mark into C2PA metadata. Inspecting a PNG’s metadata does not test text inside it.
| What you have | Mark to check | Available route |
|---|---|---|
| A passage of prose | Statistical text watermark | Anthropic Detection API, currently private preview |
| A supported image, video, or audio file | Claude-issued C2PA Content Credential | Public Claude file checker |
| A pasted screenshot or re-saved file | Any credential that remains in that copy | Check the copy, then compare with the original if available |
Anthropic’s current model-support table distinguishes text marking from file credentials. Do not treat a model’s file-credential support as a sign that its text output is marked. Anthropic is extending marking to older models, so support is changing. Check that table at the time of an investigation.
How to check a file for Claude Content Credentials
A few steps complete the public check. Open Claude’s file checker, drag in the original file, and read whether it finds a Claude-issued Content Credential. Save the result together with the exact file you checked. A later export may have different metadata. The checker says it runs in the browser and does not send the file to Anthropic’s servers.
JPG, PNG, GIF, WEBP, TIFF, HEIC, AVIF, SVG, DNG, JXL, MP4, MOV, AVI, WAV, MP3, M4A, and FLAC are accepted on the page, with a 100 MB limit. That list is the checker’s accepted input, not a promise that Claude marks every format or every product surface. A file can pass the format check and still have no Claude credential.
Record a found credential as evidence that Claude may have made or processed this file. The credential does not identify the user, prove that Claude originated every pixel or sound, or reveal the prompt. If nothing is found, ask for the original export before drawing a conclusion. Re-saving, screenshots, and format conversion can remove metadata.
Can you detect Claude’s watermark in text?
As of September 16, 2026, Anthropic’s text watermark Detection API is in private preview for eligible groups, including certain regulators, media, fact-checkers, researchers, educational organizations, civil society groups, and enterprises with relevant EU obligations. Anthropic provides an access request form. The public file checker says explicitly that it does not check text.
A keyed choice among plausible next words produces the text mark. Anthropic says it adds no hidden characters to the prose. A zero-width-character scan can identify unusual characters. It cannot test this watermark. A generic AI-writing classifier also lacks Anthropic’s key. It predicts writing style rather than checking the mark. Anthropic makes that distinction in its technical explanation.
Recent ClaudeAI discussion asked whether a public detector exists. The practical answer is narrower than many checker pages suggest. Eligible organizations can seek official API access. Everyone else can inspect supported file credentials and preserve source records. If an authorship decision is needed today, retain drafts, version history, and the original file. Do not treat an unofficial detector score as proof.
What do positive and negative results mean?
Claude was likely involved in processing the content when a supported mark is positive. The mark does not allocate authorship. A human may have supplied a draft that Claude revised, translated, or formatted. Anthropic also says light proofreading may leave too little statistical signal to detect, because Claude chose few words. A longer translation gives the watermark more opportunities to appear.
Several explanations can produce a negative result. The model or product surface may not yet support the relevant mark. A short passage may be too small for a reliable text test. Heavy editing can weaken the pattern. A re-saved file can lose its credential. Record even an official negative result as “no supported mark detected in this sample.” Note the model, file version, and test method.
Cleanup and its verification limits, for material you own, are covered in our separate Claude watermark removal guide. Detection should come first. Inspect the original, preserve it, and decide what claim the evidence can actually support.
Frequently asked questions
Can I check a passage of text for Claude’s watermark?
The text watermark Detection API from Anthropic is in private preview for eligible organizations. The public Claude file checker does not inspect text. A generic AI detector cannot verify Anthropic’s keyed watermark.
How do I check an image or file made with Claude?
Add a supported file in Anthropic’s free Claude file checker. The checker looks for a Claude-issued C2PA Content Credential in the file. A match indicates Claude may have made or processed it.
Does no detected Claude mark prove a person made the content?
No. Older or unsupported models may not mark output. Short or heavily edited text may not carry a detectable signal. File metadata can be lost when a file is re-saved or converted.