The AI-Provenance Problem: Why Trusted Data Will Become Strategic Infrastructure

The cost of producing plausible content has gone to roughly zero. The cost of establishing that a particular piece of content is what it claims to be has not moved. That divergence is the whole of the provenance problem, and it is getting wider every quarter.

Why this becomes a procurement question

For most of the last two years, provenance has been discussed as a societal concern — misinformation, deepfakes, trust in media. That framing is correct and it is not the one that will drive spending. The commercial version arrives when an enterprise buyer asks a supplier to warrant that its training data was licensed, that its outputs are attributable, and that neither will produce a claim eighteen months later.

That question is already appearing in enterprise procurement. It is difficult to answer, and most vendors currently answer it with an assurance rather than evidence.

Three things have to be true at once

Authenticity: that a piece of content originated where it claims. Permission: that its use is within the rights actually granted. Attribution: that downstream use can be traced to a source. Systems that solve one and gesture at the other two do not survive diligence, because a buyer's exposure is the union of all three, not the maximum.

Why the hard part is not cryptographic

Signing content is straightforward. Establishing that the signer had the right to sign, in a form that means something to a court in a jurisdiction that has never considered a content credential, is not. The registries that hold authoritative rights data — collecting societies, copyright offices, guilds — largely sit outside the systems being built, and a provenance layer that ignores them is asserting a parallel truth.

The teams we find most credible started from how rights disputes are actually resolved and worked backwards to the technology. The ones we do not started from a data structure.

What we are watching for

Attestation with something at stake — a signer whose reputation or balance sheet is exposed if they attest wrongly. Supersession rather than deletion, because rights positions change and immutability without correction is a liability. And integration with the registries that already hold the authoritative record. Those three are what separate infrastructure from a demonstration.