Content Provenance
Why it matters
Generative models can now produce photorealistic images, convincing video and cloned voices cheaply. Trying to detect fakes after the fact is a losing race, because detectors are unreliable and fall further behind with every model generation.
Provenance turns the problem around. Instead of trying to prove that something is fake, it lets honest sources prove what's real, and lets AI tools label what they generate. That can't stop deception by itself, but it gives journalists, platforms and ordinary viewers a signal they can check.
Deep Dive
Content provenance means attaching trustworthy information about origin and history to a photo, video, audio clip or document: which camera or AI model made it, who published it, what edits were applied. It's often compared to a nutrition label for media.
The leading standard is C2PA, from the Coalition for Content Provenance and Authenticity, founded in 2021 by Adobe, Arm, the BBC, Intel, Microsoft and Truepic. Google, OpenAI, Meta, Amazon and Sony joined later. Its consumer-facing name is Content Credentials, shown as a small "CR" icon that you can click to inspect a file's history.
How C2PA Works
When a camera, an editing app or an AI generator creates or changes a file, it writes a manifest: a set of statements such as "captured by this camera model", "generated by this AI model" or "cropped and color-adjusted". The manifest is bound to the content with a cryptographic hash and digitally signed with the tool's certificate, the same kind of technology that secures websites.
Each tool in the chain adds its own signed manifest, so the file accumulates a history. A verifier can confirm that the manifests are intact, that the signers are on a trust list and that the content hasn't changed since signing. Any tampering breaks the signature.
Adoption
- AI generators: OpenAI attaches Content Credentials to images from DALL-E and its later image models and to Sora videos. Adobe Firefly, Microsoft's image tools and Google's models do the same in many products.
- Cameras and phones: Leica shipped the first camera with built-in credentials in 2023, followed by Sony, Nikon and Canon models aimed at photojournalists. Google's Pixel 10 phones sign every photo from the camera app.
- Platforms: LinkedIn, TikTok, YouTube and Meta's apps read C2PA data to label AI-generated or camera-captured media, with uneven coverage.
The Limits
- Metadata is easy to remove. A screenshot, a re-encode or an upload to a platform that strips metadata deletes the manifest. This is the standard's biggest weakness.
- Absence proves nothing. Most media has no credentials, so a missing label doesn't mean something is fake. Bad actors simply use tools that don't label.
- Provenance isn't truth. A credential says where a file came from, not whether what it shows is accurate. A real, signed photo can still be staged or captioned misleadingly.
- Trust in signers. The system is only as good as the security of signing keys and the list of trusted certificate holders.
- Privacy. Identity information is optional by design, because mandatory identification could endanger photographers and whistleblowers.
Provenance, Watermarks and Fingerprints
Three techniques complement each other:
- Signed metadata (C2PA) is detailed and verifiable but easy to strip.
- Invisible watermarks, such as Google's SynthID, are embedded in the pixels or audio themselves. They survive many edits but carry little information and can be degraded by a determined attacker.
- Fingerprinting stores a signature of the content in a database so that a stripped file can be matched back to its record.
Recent versions of C2PA combine them as "durable credentials": if the metadata is removed, a watermark or fingerprint can be used to look up the original manifest.
Regulation
Law is beginning to require this. Article 50 of the EU AI Act, which applies from August 2026, obliges providers to mark AI-generated content in a machine-readable way and requires deepfakes to be disclosed. Systems already on the market have until 2 December 2026 to mark their content. California's AI Transparency Act requires large generative AI providers to embed provenance data and offer detection tools, and China's labeling rules for AI-generated content took effect in September 2025.