Content Credentials Checker

Add an image and this content credentials checker reads any C2PA data stored inside the file, right in your browser. The image is not uploaded and nothing is sent to a detection model. It shows what the credentials say: who signed them, which app or AI tool they name, how the image was made according to the record, and whether the signature checks out.

Credentials describe the file's history; they don't judge the pixels. For an opinion on the picture itself, use the AI image detector on the home page.

Content Credentials is the public name for the C2PA standard, an open specification from the Coalition for Content Provenance and Authenticity. A credential is a small signed record, called a manifest, embedded in the file. It can name the tool that produced the image, list actions such as created or edited, and record a digital source type from the IPTC vocabulary, such as "Created using Generative AI" or "Edited using Generative AI".

The record is signed with a certificate and tied to the image data by a cryptographic hash, so changes made after signing can be detected.

Some AI image services sign the images they deliver. In our test, original files from OpenAI, Google, Black Forest Labs (FLUX) and BytePlus (Seedream) carried AI credentials, while the Midjourney and Stable Diffusion XL files carried none.

Cameras are joining in. The Content Authenticity Initiative describes the Leica M11-P as the first camera with Content Credentials built in, and Google says Pixel 10 attaches them to every JPEG taken with Pixel Camera. In our set, real photos from a Google Pixel 9 and a Google Pixel 10 carried camera credentials. Editing apps can add entries of their own, so one credential can tell a short history: captured on a phone, then edited in an app.

The credential lives inside the file. A screenshot is a new file your phone creates from whatever is on screen, so it starts with no credential, and none of the 10 screenshots in our test had one. Many platforms also strip metadata from uploads.

Some systems try to reconnect lost credentials using invisible watermarks or fingerprints, which the C2PA specification calls soft bindings; Adobe, for one, says its Content Credentials stay connected even after a screenshot. This content credentials checker reads the credentials stored in the file you give it, so a stripped copy will show none.

An empty result from a content credentials checker is the normal case: most images reach you with no credentials, AI or not. More than half of the AI images in our test had none: every Midjourney and Stable Diffusion XL file, every screenshot and a few others. Google points to research on what it calls the implied truth effect: when only AI content carries a label, people wrongly assume unlabeled content is real.

Treat a missing credential as no information and move on to the four-step check.

When the content credentials checker finds a record, it also checks the signature on it. The result is one of three states, and each tells you how far you can rely on what the record says.

• Valid: the record is well formed, its signature checks out, and the parts of the image it covers haven't changed since signing.
• Trusted: valid, and the signing certificate chains to a trust list the checker recognises, so you know which organization signed.
• Invalid: something failed. The image may have been changed after signing, the record may be damaged, or the signature doesn't match. Invalid doesn't automatically mean fake, because ordinary software can break a credential just by rewriting the file, but you can't rely on what an invalid record says.

Even a trusted credential only tells you who signed it. A record signed by Google says a Google tool made or edited the image; it doesn't make the scene real or the caption accurate.

Yes. It runs entirely in your browser, the image isn't uploaded, and no detection model is involved, so reading credentials costs nothing.

It may be a screenshot or a re-saved copy, the platform may have stripped its metadata, or the tool that made it may not sign its files: the Midjourney files in our test carried none.

It is an IPTC digital source type for an image that was changed with a generative AI model, for example with inpainting, rather than created from scratch. Our detector reports it as Edited with AI.

Anyone can write a record, but they can't sign it with another organization's certificate unless that organization's signing key is stolen. That's what the signature check is for: an unknown signer or an Invalid result means you shouldn't trust what the record claims.