Ownership

Provenance of AI-Generated Content

You cannot detect your way out of this. You can sign your way out.

10 min readDigital Ownership

The instinct, when synthetic media became convincing, was to build detectors. Feed a file to a classifier and get back a probability that a machine made it. Enormous effort has gone into this, and it is losing - structurally, not through lack of skill.

Detection is an arms race in which one side improves continuously and the other has to keep up from a standing start. Every generation of generative model erases the artefacts the previous generation's detectors relied on. Worse, detectors fail asymmetrically: false positives accuse real people of fabricating real things, which is a far more damaging error than missing a synthetic file.

You cannot detect your way out of synthetic media. You can sign your way out - but only if signing becomes as ordinary as a timestamp.

Provenance inverts the question

Instead of interrogating the file, carry a record of how it was made. A content credential is a manifest attached to a file - cryptographically signed by whoever produced or edited it - recording the capture device, the editing steps, and whether a generative model was involved and how.

Verification then checks a signature rather than analysing pixels, which means it does not degrade as models improve. It also changes what the answer means. A detector says "this is probably synthetic". A credential says "this organisation is publicly staking its reputation on this account of where the file came from". The second is weaker as a metaphysical claim and far more useful in practice, because it produces someone to hold responsible.

How it works, concretely

  1. Assertion A tool records facts about what it did: captured on this device at this time, or generated by this model from this prompt, or cropped and colour corrected.
  2. Binding The assertions are hashed together with the content itself, so a change to the pixels invalidates the manifest.
  3. Signing The producer signs the manifest with a certificate that chains to an identity a verifier can evaluate - a camera maker, a newsroom, a software vendor.
  4. Verification Anyone can check the signature, see the chain of edits, and decide whether they trust the signer. No model, no guessing.

The three failure modes, stated honestly

Stripping

Metadata is trivially removed. Screenshot the image, re-encode the video, copy the text - the credential is gone. Nothing prevents this, and any account of provenance that does not concede it up front is selling something.

The mitigation is to treat absence as uninformative rather than damning. A file without a credential is not proven fake; it is merely unverified, in the same way that most photographs have always been. What credentials provide is a way for content that wants to be trusted to demonstrate it - soft binding techniques such as watermarks and perceptual hashes can partially recover a stripped manifest, but the honest baseline is that provenance is opt-in.

Lying at the source

A signature proves who made a claim, not that the claim is true. A camera pointed at a screen showing a fabricated image produces a genuinely signed photograph of a fake. Provenance moves trust to the signer; it does not eliminate the need for trust.

Nobody checking

The largest practical problem. Credentials that no platform displays and no reader looks at do nothing. This is a distribution and interface question rather than a cryptographic one, and it is where the effort should now be going.

Why this matters commercially, not just journalistically

The public discussion focuses on misinformation. The commercial stakes are more immediate and less discussed:

  • Licensing. If you cannot establish where an asset came from, you cannot establish what you are allowed to do with it. Provenance is a prerequisite for a functioning rights market, not an add-on to one.
  • Insurance and evidence. Claims processes, inspections and compliance records all rest on photographs. All of them now need a story about authenticity.
  • Training data. Knowing what a model was trained on is becoming a legal question in several jurisdictions, and provenance at the input side is the only mechanism that scales.
  • Agent outputs. When software produces work autonomously, the record of what produced it - under what instruction, on whose authority - is the same problem wearing different clothes. See agent identity.

Where a ledger fits, and where it does not

Signing a manifest requires no blockchain, and claiming otherwise has damaged the credibility of an otherwise sound idea. What a ledger does contribute is narrow: an anchor for revocation, so a compromised signing key can be repudiated without trusting the compromised party, and a timestamp proving a signature existed before a given moment. Both are real and both are infrastructure - the honest scope is smaller than the marketing version.

What to do now

For anyone publishing: sign what you publish, if your tools support it, and record the generative steps rather than concealing them. The disclosure that feels like an admission today will look like ordinary practice within a few years, and being early is cheap.

For anyone consuming: treat a valid credential as evidence about the signer, not proof about the world, and treat its absence as uninformative. The useful mental model is a byline - it tells you who is responsible, which is exactly what you need in order to decide how much to believe.

The broader ownership picture is on digital ownership.