So here is the uncomfortable question: if the picture itself can't be trusted, what can?
The most credible answer the industry has produced is a way to make media carry its own paperwork, not a better way to look at media. The standard is called C2PA, and the consumer-facing version is called Content Credentials. Think of it as a receipt that travels with the file.
The Content Authenticity Initiative likes to call it a "nutrition label for digital content," and that's a fair start. A Content Credential is a small bundle of data (a manifest) attached to an image, video, or audio file. It can record who or what created the asset, when, with which tool or camera, and what edits happened along the way. If an image came out of a generative model, a well-behaved credential says so, including the model.
The important part is the word verifiable. This is cryptographically sealed, not the kind of metadata you can open in a text editor and rewrite.
Strip away the jargon and there are two moving parts.
First, the file gets hashed. A hash is a mathematical fingerprint of the exact bytes of the image. Change a single pixel and the fingerprint changes completely. C2PA uses SHA-256, the same hashing standard that secures much of the modern internet.
Second, that fingerprint, plus the history of the file, gets digitally signed using standard X.509 certificates, the same kind of certificate technology behind the padlock in your browser's address bar. The signing format, COSE, is a recognized standard also used in secure, low-resource devices.
The effect is a sealed link between the file and its story. If someone alters the image after signing, the fingerprint no longer matches and the seal breaks. A verifier (a website, an app, a browser) can check the math and tell you the credential was tampered with.
One word matters enormously here, and the standard's own authors insist on it: tamper-evident, not tamper-proof. C2PA cannot stop anyone from messing with a file. It can only make the messing show. That distinction is the whole reason this article is honest and not a sales pitch.
This is where a lot of writing about C2PA goes wrong, promising a future as if it were the present. So let's stick to what is genuinely live.
Adobe has the deepest implementation. Content Credentials are written across Creative Cloud, including its Firefly generative tools, so edits and AI generation get recorded in the manifest.
OpenAI attaches C2PA metadata to images from DALL·E and ChatGPT, and joined the C2PA steering committee in May 2024. In 2026 it layered an embedded SynthID watermark on top of the credential for supported generations.
Google made Pixel 10 the first smartphone to sign photos with Content Credentials in its native camera, and is extending capture to video on Pixel 8, 9 and 10. It has begun bringing C2PA verification and SynthID detection into Search and Chrome.
Camera makers are in. Leica shipped the first production camera with Content Credentials, the M11-P, and Sony, Canon, and others now sign in-camera with device certificates.
Social platforms read credentials on the way in. TikTok joined the effort in May 2024 and began labeling AI content via Content Credentials. Meta applies "AI Info" labels on Instagram and Facebook when it detects C2PA data.
That is real, broad, and growing. It is also, crucially, not universal, and the gaps are the whole story.
Adoption is not the same as flawless. In 2025 Nikon added C2PA to a camera body, then suspended its authenticity service after a security vulnerability was found in it. Provenance systems are software, and software has bugs. A signed file is only as trustworthy as the chain that signed it.
Here is the part you must internalize before you ever lean on a credential.
Metadata gets stripped. Constantly. Standard credentials are embedded inside the file, and that data is fragile. WhatsApp, iMessage, and many platforms re-encode images on upload and silently discard the manifest. A perfectly authentic, signed photo can arrive in your feed with nothing attached, simply because it passed through a pipe that didn't care.
Screenshots break the chain entirely. Photograph or screenshot a credentialed image and you've created a brand-new file with no manifest at all. The seal doesn't travel through a screen grab. This is the single easiest way to launder provenance out of existence, and it requires no skill.
Absence proves nothing. This is the rule that catches careful people. A missing credential does not mean "AI" and does not mean "fake." It usually means the file went through something that dropped the data, or was never signed in the first place. You cannot reason backward from no-receipt to a verdict.
The industry's answer to stripping is Durable Content Credentials, which pair the embedded manifest with an invisible watermark and a content fingerprint, so a stripped credential can sometimes be recovered from a cloud repository. It helps. But the same engineers will tell you watermarks can be removed or spoofed and fingerprint matching is fuzzy: useful as a hint, not proof. Even a U.S. government advisory on the technology frames provenance as one layer among several, not a silver bullet.
So C2PA isn't a lie detector. What is it good for?
It is the strongest positive signal available when it's present. A valid Content Credential that says an image came straight from a Leica with no edits is far better evidence than your gut squinting at the shadows. A valid credential naming a generative model is a near-certain tell that the image is synthetic.
Use it as a one-way test. Present and valid tells you a great deal. Present and broken is a loud warning. Absent tells you almost nothing. Fall back on the slow, boring fundamentals: where did this come from, who published it first, does the source survive a search.
That's the honest hierarchy. The receipt, when you can get one, beats the picture. When you can't get one, you're back to doing the work yourself, which is the whole muscle a daily round of real-or-AI is quietly building.
If an image has no Content Credentials, is it fake?
No. Absence proves nothing. Plenty of genuine photos carry no credentials, and most platforms strip the data on upload anyway. A missing receipt is not evidence of guilt, just a missing receipt.
Can Content Credentials be faked or removed?
The signed data can't be silently altered: tampering breaks the cryptographic seal and shows up. But the whole manifest can be stripped out entirely, and a screenshot leaves it behind completely. It's tamper-evident, not tamper-proof, and not unstrippable.