The problem: trusting digital media
A photograph used to be evidence. Today, anyone can make a convincing image of anything. How can anyone know where a piece of content really came from?
By the end of this chapter you will be able to
- Explain why ordinary file metadata can’t be trusted.
- Contrast detecting fakes with proving provenance.
- Describe, at a high level, what C2PA and CAWG are and who makes them.
- State clearly what Content Credentials can and cannot tell you.
A crisis of evidence
For most of the history of photography, faking a convincing image took skill, time and equipment. That friction gave photographs a kind of default credibility. Then three things happened in quick succession:
- Editing became effortless. Desktop and phone editors made retouching, compositing and cloning trivial.
- Distribution became instant and context-free. An image is screenshotted, cropped, re-uploaded and shared millions of times, losing every clue about where it started.
- Generative AI arrived. Systems can now synthesise photorealistic images, voices and video from a sentence of text.
The result is sometimes called the liar’s dividend. When fakes are easy, it isn’t only fakes that cause harm. Real evidence can also be dismissed as “probably AI”. A journalist’s genuine photo of a war crime and a fabricated one look identical on a phone screen.
Why not just look at the metadata?
Most image files carry metadata: extra information stored alongside the pixels. Cameras write EXIF fields such as camera model, exposure, date and GPS location. Photo agencies add IPTC fields such as caption, creator and copyright. It is tempting to treat these as facts. Try the demo below.
Three properties make ordinary metadata useless as evidence:
- No integrity. Nothing detects a change. Edit a field and the file is exactly as “valid” as before.
- No attribution. Nothing records who wrote a field. “Artist: Jane Photographer” might have been typed by anyone.
- No durability. Many platforms strip metadata on upload, for privacy and file size, so even honest metadata rarely survives.
By the end of this course you will see how C2PA addresses the first two directly with cryptography, and partially addresses the third.
Detection vs. provenance
There are two broad strategies for dealing with untrustworthy media.
| Detection | Provenance | |
|---|---|---|
| Question asked | “Does this look fake?” | “What does this content’s verifiable history say about it?” |
| How | Classifiers look for artefacts of manipulation or generation. | Creators and tools attach signed, tamper-evident records. |
| Weakness | An arms race: generators improve and detectors fall behind. Results are probabilistic. | Only works for content that carries credentials, and credentials can be removed. |
| Output | A probability (“87% likely synthetic”). | Cryptographic facts (“signed by X; unchanged since”). |
The two approaches are complementary. Provenance flips the burden of proof: instead of trying to catch every fake, it lets authentic content prove itself. A newsroom can show that a photo came from a specific camera, was edited in specific ways, and was published by them. An AI image generator can label its output as AI-generated in a way that is hard to remove silently.
Enter C2PA and Content Credentials
The Coalition for Content Provenance and Authenticity (C2PA) is a standards body formed in 2021 by companies including Adobe, Microsoft, Intel, the BBC, Truepic and Arm, joined since by camera makers, AI companies, news organisations and platforms. It publishes the C2PA Technical Specification, an open standard for attaching provenance information to digital content.
That provenance information is called a manifest. When shown to people, for example via the small “cr” icon, it is branded Content Credentials. A manifest can record:
- what kind of device or software made or changed the content (a specific camera model, an editing app, an AI model),
- how the content was made (captured by a camera? generated by AI? a composite?),
- what was done to it (cropped, colour-adjusted, combined with other images),
- which earlier assets it was made from (its ingredients), each with its own history.
Everything in a manifest is cryptographically signed and bound to the exact content, so any change to the content or the manifest is detectable. The rest of Part II explains precisely how.
…and CAWG
You’ll notice the list above says what device or software made the content, not what person made it. That is deliberate. Since version 2.0, the C2PA signature identifies the product that created the manifest, not the human using it. Identity, meaning who the creator is and what role they played, is handled by the Creator Assertions Working Group (CAWG), now part of the Decentralized Identity Foundation (DIF). CAWG defines additional assertions that plug into C2PA manifests, most importantly the identity assertion. Part IV covers it in depth.
Content Credentials prove who signed what, and that it hasn’t changed since. They do not prove that the content is true. A trustworthy camera can photograph a staged scene. An honest editor can sign a misleading crop. Provenance gives you verifiable facts about origin and history. Judging the content remains a human job.
The roadmap
To understand how a signature on a manifest can be trusted, you need four cryptographic building blocks. Each gets its own chapter:
- Hash functions: compact fingerprints that change if anything changes (chapter 2).
- Key pairs: a private key that only you hold and a public key anyone can see (chapter 3).
- Digital signatures: proof that the holder of a private key approved exactly these bytes (chapter 4).
- Certificates and time-stamps: how you know whose key it is, and that it was valid when used (chapters 5–6).
Part III then assembles these into C2PA, and Part IV adds CAWG identity.