C2PA Content Credentials Missing? What It Actually Tells You

Witness Team · Editorial

Key takeaways

• C2PA (Content Credentials) is a cryptographic provenance standard. When an image carries it, you can verify what created it and with what tool.
• Absence is not a verdict. In most contexts it's close to neutral — but in specific contexts it's genuinely informative, and knowing which is which is the whole skill.
• Social platforms strip provenance metadata on upload. So does taking a screenshot. This is the standard's biggest practical limitation.
• Camera-side signing is expanding. As it does, absence becomes progressively more informative — which is the argument for building this layer now.
• Witness reads provenance alongside statistical detection and reports them as separate signals, not one fused score.

What C2PA is

The Coalition for Content Provenance and Authenticity is an open standard developed by Adobe, Microsoft, Google, the BBC, Intel, Sony, Nikon and others. It lets a camera, editing tool, or publishing platform cryptographically sign an image, attaching a manifest recording what created it, what tool was used, and what edits followed.

The signature travels with the file. Edit the image in a C2PA-aware tool and a new claim appends to the chain. Export it through anything that isn't C2PA-aware and the chain ends there.

The key property is that a manifest can't be forged without the signing key. The signature chains back to a certificate issued to the camera maker or software vendor, so you can verify that a particular camera model produced a file, or that a particular generator did. What you can't do is fabricate a manufacturer's signature without their private key. That's a strong guarantee, but not an absolute one — keys can leak and certificates can be mis-issued, and a valid signature on a maliciously constructed manifest is a recognized threat in the standard's own security model.

How provenance is actually distributed

The pattern matters more than any single number, and it breaks down along four lines.

Recent cameras from brands that have adopted the standard sign at capture. Leica, Sony, Canon, Fujifilm, and Panasonic have shipped C2PA-capable cameras; Google's Pixel 10 is the first mainstream smartphone to sign every photo at capture by default. But this is still a small share of the installed base. Most cameras in active use predate the standard entirely, and their images are perfectly real with no signature at all.

Some AI generators sign at generation, and most don't. DALL·E 3, Adobe Firefly, and Bing Image Creator embed a manifest that explicitly asserts the content is machine-generated. Stable Diffusion, Midjourney, and Flux have optional C2PA export that isn't on by default, and anything generated locally or through a non-signing API carries nothing.

Publishers are adopting it faster than consumers. Adoption is expanding through CAI publisher integrations — newsrooms signing and countersigning their imagery — which makes provenance meaningfully more informative in a professional news context than a general one.

Social platforms strip it. Every major platform removes provenance metadata on upload. This is deliberate rather than an oversight — platforms don't want to be in the business of re-signing third-party content — and it's the single biggest practical limit on the standard's usefulness today.

What absence tells you

A valid signature from a credible camera issuer is positive evidence of authentic capture. That's the strongest signal available.

A valid signature from an AI generator is positive evidence of synthetic origin — and notably, this holds even when our statistical detectors don't fire. Two independent signals can diverge; when they do, the provenance signal is the one that doesn't depend on a model's training distribution.

Absence is where judgment comes in, and the honest answer is that it usually tells you less than people assume.

The pattern: absence is weak evidence in isolation and stronger when you know how the image reached you. Distribution path is the variable that does most of the work.

And this changes over time. Every camera generation that ships with signing enabled makes absence a little more meaningful. The question is drifting from "can we detect the fake" toward "why does this supposedly authentic photo carry no credential" — and a verification tool that only reads one direction will be answering the wrong question in a few years.

Why metadata goes missing

Provenance is fragile by design. Re-encode a file, strip its metadata, or capture it as a screenshot, and the manifest is gone — the image is unchanged to the eye, and the chain is severed. Screenshotting is the most common path by which signed synthetic content reaches circulation unsigned, and no amount of cryptographic strength prevents it, because nothing about the original file is being attacked.

This is precisely why provenance can't stand alone. The adversarial case is the one where the signature is missing, so the statistical layer has to carry the weight exactly when the provenance layer has nothing to say.

There's a partial compensation. Screen capture layers its own artifacts onto whatever was underneath — display gamma, rendering pipeline effects, moiré from the pixel grid, and re-compression at capture quality. Our detectors can often still identify synthetic content through that layer, though it's harder and we return Uncertain more often on such inputs.

Why this is uncommon

Among the detectors we've evaluated, we haven't found one that surfaces the provenance chain alongside its verdict. They return Real or Synthetic and nothing else.

That's a category error, because the two answer different questions. Detection asks whether an image carries the statistical fingerprints of generation. Provenance asks whether there's a verifiable record of what made it. Those come apart constantly:

• A real news photograph with heavy artistic processing can trip detection while carrying clean camera provenance and a newsroom countersignature.
• Output from a generator too new for our models to have seen can pass detection while carrying an AI generator's own signature declaring exactly what it is.

Reading both isn't conceptually hard, but it isn't free either — it means parsing manifests, validating certificate chains against issuer roots, and handling partial or broken chains gracefully. Most detection tools are built as classifiers with no provenance path anywhere in the architecture, so adding one isn't a feature toggle.

We report the two as separate states rather than collapsing them into a single score:

The last row is the one that matters most and the one a single fused score destroys. When two independent signals disagree, that disagreement is the finding.

Verifying an image in practice

Check provenance first. A credible camera signature is strong evidence; an AI generator signature is close to definitive. No signature is usually neutral — don't stop there.

Then run detection. Statistical models catch what provenance can't: unsigned generators, stripped metadata, everything that reached you through a platform.

Weigh the distribution path. Social media arrival means absence is expected and uninformative. A raw file attachment makes absence worth more.

Take Uncertain seriously. Mixed detector signals often mean post-processing has disrupted the statistical layer — which is itself a reason to lean harder on provenance and sourcing.

Detectors can be confidently wrong in both directions, and the false-positive direction is the one people forget. A system that labels authentic footage as AI-generated does real damage — it discredits genuine evidence, and it hands anyone caught on camera a ready-made excuse. That failure mode is the strongest argument there is for reading provenance and detection as independent signals rather than trusting either alone: when a classifier misfires on real footage, an intact provenance chain is what contradicts it.

Try it

Upload any image at witness.vision/scan. Witness reports the detection verdict alongside provenance — signed or unsigned, and by which issuer.

Doing this at volume? Talk to us about the API.