Verified credential
A signed C2PA claim validates and can describe origin, tools and edits.
Inspect images and text for verifiable provenance, metadata hints, and hidden characters. Free and private in your browser.
Drop an image here
JPEG, PNG or WebP • up to 25 MB
Your file stays on your device. Analysis runs locally in your browser.
Each tool answers a narrow question honestly rather than producing one vague score. Pick the one that matches what you need to know.
Signed provenance inspection
Validate C2PA Content Credentials in JPEG, PNG, and WebP images privately in your browser. Inspect the signer, claim generator, recorded actions, ingredients, and validation issues.
Content provenance made legible
Read and verify Content Credentials without uploading anything. Inspect signed provenance, edit history, digital source types, signer identity, and ingredient chains in your browser.
Evidence, not a score
Check images for signed provenance and AI-related metadata in your browser, and learn exactly which AI watermarks can be verified independently and which cannot.
Provider strings and editable hints
Inspect EXIF, IPTC, XMP, GPS, and AI-provider metadata in JPEG, PNG, and WebP images locally. See which generator strings are present without uploading your photo.
Every field, nothing uploaded
View EXIF, IPTC, XMP, and GPS data from JPEG, PNG, and WebP photos directly in your browser. See camera, lens, exposure, timestamps, and location without uploading the file.
Unicode inspection
Find zero-width spaces, joiners, directional controls, soft hyphens, and byte order marks in pasted text. Every character is named with its code point and position, locally in your browser.
Clean text, kept intact
Remove zero-width spaces, byte order marks, soft hyphens, and other invisible Unicode characters from text. Runs in your browser, keeps the characters your language needs.
Observable artifacts only
Inspect text for observable Unicode artifacts and understand why no third-party tool can verify a Claude statistical watermark or produce a reliable Claude probability.
A careful answer
A careful answer on Claude text watermarking: what statistical watermarking is, what Anthropic has published, why hidden Unicode is a myth, and why no third-party detector is reliable.
Provider-controlled watermark
How Google SynthID differs from C2PA, what can be inspected locally, and why independent SynthID verification requires Google's own detection flow.
Signed provenance and editable metadata are not the same. Results are grouped by the strength of the evidence instead of producing a vague AI percentage.
A signed C2PA claim validates and can describe origin, tools and edits.
A credential exists, but a signature, hash or trust check reports a problem.
Editable EXIF, IPTC or XMP strings mention software or a provider.
Nothing verifiable was found. That is not a human-made verdict.
Longer explanations of how provenance works, and where the popular shortcuts go wrong.
A practical, honest method for checking whether an image was generated by AI: what provenance can prove, what metadata hints at, why visual tells fail, and when to stop.
What OpenAI attaches to generated images, why the marking disappears so easily, and how to check any specific image yourself instead of relying on a general claim.
Content Credentials explained without jargon: what they record, how the signing and binding work, what they prove, what they do not prove, and why they disappear so easily.
Em dashes, hidden characters, 'delve', and clean formatting are widely cited as AI tells. Here is why each one fails, and what actually distinguishes machine-written text.
The browser is the boundary. Your files, text, metadata, and results stay in the current tab.
Your image is opened through browser APIs; there is no upload endpoint.
Preview URLs live only in the tab and are revoked when replaced or closed.
Filenames, text, metadata, manifests and results are excluded from analytics and advertising.
C2PA Reader and SDK resources are disposed after every analysis.
No. It reports verifiable C2PA provenance when present and labels ordinary metadata as hints. Missing evidence never proves human authorship, and no AI percentage is produced.
No. File bytes, metadata, manifests, pasted text and scan results are processed in your browser. There is no upload endpoint on this site, and none of it reaches analytics or advertising.
It is a cryptographically signed provenance record whose structure, signature and binding to the file all validate. It can describe tools, actions and source types, but it does not guarantee the truth of what an image depicts.
Screenshots, social networks, editing, export settings and format conversion all remove C2PA manifests and ordinary metadata. An empty result is the normal case on the open web.
No. Claude's described watermark concerns statistical word-selection patterns, not hidden Unicode. This tool reports only directly observable characters, and no public detector for statistical text watermarks exists.
Yes. There is no account, no usage limit and no paid tier. The site is funded by advertising, which is kept out of the checker, the results and the policy pages.