Provenance (C2PA)

A photo on the internet doesn’t reveal on its own where it came from. It could have been taken with a camera, altered on a computer, or entirely invented by a program. Provenance is the attempt to deliver this backstory along with the file: the file gets an attached record noting who created it, with which device or program, and what was later changed about it. C2PA is the name of the most important set of rules for such records. Behind it stands an alliance of major companies like Adobe, Microsoft, Sony, and the BBC, who have agreed on a common format. You can imagine it like a package insert that’s firmly stuck to the file, describing its journey.

Why proof of authenticity is becoming more important than forgery detectors

Programs that generate deceptively realistic images and voices are accessible to everyone today. This causes an old principle to lose its validity: for a long time, an image was a fairly reliable piece of evidence that something had actually happened. It is exactly this evidence that is currently crumbling, and newsrooms, courts, and insurers face the problem of no longer being able to simply believe recordings.

The obvious countermeasure would be a detector that recognizes artificially generated images. Such tools exist, but they are unreliable. They only provide probabilities, make mistakes with unusual subjects, and become outdated as soon as a new generation of image generators appears. It’s a race that the checkers are structurally bound to lose.

Provenance flips the question around. Instead of proving that something is fake, you prove that something is genuine. A camera signs its recording directly at the moment of shooting. An editing program then adds which steps it performed. If this proof is missing, the image is not automatically false. But it is unverified, and in certain contexts that alone can already be an important clue.

Signatures, manifests, and the question of trust

Technically, C2PA creates what is called a manifest. This is a small block of data within the file that contains information about the creator, the time, the device used, and the edits performed. This also includes an indication of whether part of the image was generated by an AI model, meaning a program that computes images from a text description.

So that this manifest cannot simply be rewritten, it is cryptographically signed. Put simply: a kind of checksum is calculated from the image content and the accompanying data, then sealed with a secret key. Anyone can use the matching public counterpart to check whether the seal is valid. If someone changes even a single pixel, the checksum no longer matches and the software raises an alert.

Edits build on one another in the process. Each step creates a new entry that references the previous one, forming a chain. The best-known weak point: anyone who photographs an image off a screen or takes a screenshot removes the signature entirely. So C2PA does not prevent forgery. It merely makes complete proof possible wherever it exists.

Content Credentials in cameras, chatbots, and social networks

Under the name Content Credentials, the standard is already appearing in products. Adobe Photoshop can record the provenance data along with the file, and DSLR and mirrorless cameras from Leica, Nikon, and Sony sign recordings directly in the device. Image generators from major providers are now also marking their outputs this way.

On platforms like LinkedIn or TikTok, corresponding content gets a small icon. Clicking on it shows the stored history. News agencies are testing the standard so they can prove to readers that a press photo has remained unchanged since it was taken. For investors, the field is also economically interesting, since a distinct market is forming around provenance verification.

A common misconception is that C2PA is a watermark. A watermark hides an invisible pattern within the image itself and often survives even screenshots, but reveals little about the backstory. Provenance, conversely, delivers a lot of detail, but is easily lost when copied. In practice, the two methods are therefore often combined.