What C2PA provenance actually proves

Check your starting point in 30 seconds

Before we dig into C2PA, place yourself.

If you saw a "Content Credentials" badge under a dramatic news photo, what would you assume it means?

If you picked the first one, this article will correct some overtrust. If you picked the second or third, it will give you a clearer, practical way to use the badge without treating it like magic or noise.

C2PA in one line, then the signed manifest model

Here is the one line version:

C2PA provenance proves that someone with a specific cryptographic key claims to have created or processed this exact file with a specific chain of tools. It does not and cannot prove that the scene happened, that the caption is honest, or that no steps were left out.

C2PA (from the Coalition for Content Provenance and Authenticity) defines a way to attach a signed manifest to a media file. The current 2.x spec treats the file as an object with a history. Each step in that history can record things like the device or app name, a timestamp, and a description of the operation.

You can think of it as a tamper-evident shipping log. A camera app, a photo editor, or an AI generator can "check in" a new step, then sign the updated manifest. When you later view the file on a supporting platform, your browser or app verifies the signatures and shows you a readable version of that log.

The key move is this: the cryptography protects the link between the pixels and the manifest. It does not judge the content. A staged press photo can have perfect provenance. An honest citizen video can be unsigned.

Example 1: a press photo with a clean C2PA trail

Start with a relatively friendly case.

Imagine a wire service photo: a politician at a podium, distributed by a major agency that participates in the Content Authenticity Initiative. You click the Content Credentials badge in your browser.

You likely see a short chain. It might show a camera body model, the agency's ingest system, and an editor's workstation. The signer is the agency, which holds a private key registered in the C2PA ecosystem.

What this provenance does prove in this case: that the agency (or its staff) claims to have taken this exact photo and processed it with those tools, and that the image file was not altered after the final signature without breaking the manifest.

What it does not prove: that the event was not staged, that the caption is honest, that no earlier versions existed with different framing, or that there was no off-platform editing before the camera first wrote a manifest.

As of mid-2026, several major newsrooms and agencies are piloting or rolling out this workflow. They see C2PA as one input into their trust relationship with readers, not as a replacement for trust.

Example 2: a Midjourney v7 AI image with Content Credentials

Now take a clear AI image.

You see a stylish image of "a neon-lit dragon flying over Tokyo" credited to a creator on a popular art platform. A Content Credentials badge appears, and you open it.

Here the manifest might show that the file originated from Midjourney v7, or another AI system that has integrated C2PA output. It could include the model name, some prompt or settings metadata, and then a publishing step from the hosting platform.

In this case, C2PA provenance tells you something powerful: this was not a photo that someone "lightly edited". It is the output of a text-to-image model. The log separates it from documentary photography without you needing to inspect pixels.

It still does not tell you whether the creator's caption about "leaked game art" is honest. The provenance also does not say whether someone composited AI elements into a camera photo somewhere else, then fed that into the generator. You know the file is model output. You do not know the full creative process outside the logged steps.

This is where people often overread the badge. A clean C2PA log from an AI tool is honest about what it is, but that honesty does not grant it documentary status.

Example 3: a human edit of an AI image

Now combine the two.

You find a political meme that looks partly realistic and partly cartoonish. A badge is present. You open the provenance.

You might see a chain that starts with a generative model, such as Midjourney v7, then passes through an image editor, such as Photoshop, then perhaps a mobile app that added text overlays. Each step can record that it cropped, recolored, or composited parts.

What this proves: the file you are looking at is derived from AI output and then edited by a human with specific tools, in a specific sequence, at roughly specific times. If the chain is intact, any pixel-level change after the last step would break the signature.

What it does not prove: that you are seeing the full chain. A bad actor could strip provenance entirely by exporting or screenshotting. A more subtle actor could start with unsigned assets, then only sign the last few steps. The C2PA spec allows optional fields, and not every tool logs the same level of detail.

The manifest is a map of where the file claims it has been, not a CCTV recording of its entire life.

What "verified" means on major platforms in 2026

In mid-2026, most platforms that support Content Credentials or similar labels mean something narrow by "verified".

Roughly, it means: this file contains a C2PA manifest, the signatures are cryptographically valid, and the manifest entries line up with known keys and tool identifiers. That is all. It is a technical verification of the log, not an editorial endorsement of the content.

Some platforms add light policy language. They might say that verified media meets certain integrity checks, or that they highlight signed content from trusted partners. Under the hood, they are still checking signatures and known keys, not running lie detectors on the pixels.

As a reader, treat "verified" here as you treat "https" in a browser bar. It proves you are talking to someone who controls a certain key and has a route to this file. It does not prove that person is honest.

This is also why vendor marketing around C2PA can feel slippery. Adobe, Microsoft, OpenAI and others push it as part of their trust story. The cryptography genuinely helps with provenance, but the social layer of who is allowed to sign and what they sign is where your judgment still matters.

Signed, unsigned, and broken: the three states

When you encounter C2PA content provenance in the wild, you are usually in one of three states. Think of them like traffic lights, but remember that the colors are hints, not commands.

A signed state is the most informative. You can read the log, see tools, and tie them to keys. Unsigned is common and not suspicious by default. Many citizen photos, older archives, and casual posts will stay unsigned for years.

A broken state is tricky. It can indicate deliberate tampering, a simple resize in a tool that does not preserve manifests, or a platform bug. By itself, "broken" is a reason to pause and seek more context, not grounds to declare something a deepfake.

This is where the screenshot problem shows up most often, which we will turn to next.

Screenshots and missing signatures

Screenshots confuse many people who first meet C2PA.

If you open a signed image in your browser, then take a screenshot, your new screenshot file normally has no C2PA manifest. It is a fresh file written by your operating system or screenshot app. The original signature stayed with the original file.

So if you upload the screenshot to a social network, that platform sees an unsigned image. The content might match the pixels of a signed press photo, but the provenance is gone. That does not prove tampering. It proves only that at least one step in the chain did not preserve or recreate the manifest.

Attackers can abuse this. They can screenshot or re-encode signed content to strip provenance, then circulate it without the badge. Honest users do the same out of convenience all the time. A missing signature in 2026 is weak evidence of anything.

When you see an image without C2PA data, your first question should not be "Is this a deepfake". It should be "What type of content is this, and would I expect the creator to use provenance tools at all".

Who controls the signing keys, and why that matters

The strength of C2PA provenance depends on who holds the keys.

Each signer has a private key that corresponds to a public identifier known to the ecosystem. A camera vendor might embed a key in hardware. A news agency might manage keys for its staff. An AI generator might sign all output under a platform key.

If a key is stolen, poorly managed, or granted to untrustworthy users, the signatures become less meaningful. You might see a technically valid manifest for a harmful fake created by someone who should never have had signing rights.

This is why you should always read

who signed a file, not just whether it is signed.

A press photo signed by a major agency with a strong track record carries more weight than a political meme signed by an unknown group with anonymous operators. Both might be valid C2PA objects. The social trust is different.

Key governance is still evolving. The C2PA spec and the Content Authenticity Initiative describe technical and organizational best practices, but they do not run the trust network. That work is happening in newsrooms, platforms, and standards discussions, and it will likely keep shifting as attackers respond.

As a reader, you do not need to become a PKI expert. You only need to remember that cryptography is like a lock, and keys can be handed to the wrong people.

Practice: read three badges like a pro

Now a short exercise you can do in ten minutes. You only need a laptop or phone and a browser.

1. 1
   Find three images. Pick a signed press photo from a major newsroom, a clearly AI-generated art image from a platform that supports Content Credentials, and one meme or social post that you suspect is AI but that might not be signed.
2. 2
   Open the provenance (when present). Click the badge or "Content Credentials" label and open the full panel. If there is no badge, note that too.
3. 3
   Answer four questions for each image. Who is listed as the signer. Which tools or models appear. Is there any AI system named. Is the image signed, unsigned, or broken.
4. 4
   Write two sentences per image. One sentence starting with "C2PA proves that..." and one starting with "C2PA does not tell me...".

Now check your own work.

Good signals: your "proves" sentence talks about the file, tools, and signer, not about the truth of the story. Your "does not tell me" sentence mentions things like staging, caption honesty, or off-log edits.

Weak signals: your "proves" sentence claims the image is real or unedited, or your "does not tell me" sentence is vague, such as "I still need to be careful" without naming what you would check.

If your answers were weak, retry on a different set of images and slow down. Spend an extra 10 seconds on the signer and tool names. Say them out loud. You are training your eye to link provenance to real-world actors, not to abstract badges.