Digital ownership when most content is machine-made

Four questions hiding inside one word

When someone says they own a digital file, they may mean any of four things, and the differences decide every dispute:

• Control - they hold the keys or the account, and can move or delete it.
• Title - a recognised claim that survives the platform holding the file.
• Licence - permission to use it in specified ways, which is what most commercial "ownership" actually is.
• Provenance - a verifiable account of where it came from, which is not ownership at all but is what most ownership claims rest on.

The 2021 wave of digital collectibles is best understood as a large-scale experiment that answered the first question well, the second partially, and the third and fourth barely at all. Control worked: keys held assets, transfers settled, nobody needed a platform's permission. Title was mostly a pointer to a file somewhere else. Licence terms were usually absent or unenforceable. Provenance covered the token, not the artwork.

What generative models changed

Three shifts, each of which breaks an assumption that used to be safe:

1. Scarcity stopped being evidence. A rare file was once expensive to produce. Now anything can be produced in quantity, so rarity says nothing about origin and everything about enforcement.
2. Authorship became a spectrum. A photograph retouched by a model, a text drafted by one and edited by a person, a design generated from a brief - the binary of human-made or machine-made no longer describes what happened.
3. Training turned works into inputs. A body of work can influence a model that then competes with it, without any individual file being copied in a way the old framework recognises.

Where the law currently stands

Broad strokes, because this moves and jurisdictions differ: purely machine-generated output generally does not attract copyright - US and UK practice both require human authorship in some form, and EU law reasons from the author's own intellectual creation. Human contribution can restore protection, but the protected part is the human contribution, not the whole output. In practice this means a company's AI-generated assets may be freely copyable by competitors unless something else - contract, trade secrecy, trademark - is doing the work.

The training-data question is separate and genuinely unsettled, moving through courts and legislatures on different timetables in different places. Anyone claiming a clear answer today is describing their preferred outcome.

Provenance is the layer that scales

Detection - asking a classifier whether a file is synthetic - is losing and will keep losing, because it is an arms race against systems that improve faster than the detectors. Provenance inverts the problem: instead of interrogating the file, you carry a signed record of how it was made.

A content credential is a manifest attached to a file, signed by whoever produced or edited it, recording the capture device, the edits, and whether a generative step was involved. Verification checks a signature rather than guessing at pixels. It does not tell you a file is true; it tells you who is willing to stand behind the claim about where it came from, which is a far more useful thing to know. How it works in practice.

Licensing that machines can read

If agents are going to use, remix and pay for content, the terms have to be legible to them. A licence in a PDF is invisible to a crawler; a licence expressed as structured metadata alongside a price is something an agent can evaluate and act on. This is the quiet precondition for content having a business model in an agentic web: not blocking machines, but pricing them.

The building blocks exist - machine-readable rights expressions, per-request payment, verifiable identity for the requesting party. What is missing is convention. Until publishers converge on a way to say "this may be read by an agent, for this fee, under these terms", the practical choice remains binary: block the crawlers or give the work away.

What agents own

The final turn is strange but follows directly. If an agent transacts, it holds things: credits, licences, tickets, access rights. Those holdings need to survive the agent being restarted, migrated or replaced, and they need to be provably held rather than merely recorded by whoever operates it. That is the narrow, real reason bearer claims and verifiable credentials keep appearing in an argument that started about copyright.

Whether an agent can own something in the legal sense is a different and much less interesting question - it cannot, and holds everything on behalf of a person or company. But the technical requirement that its holdings be portable and verifiable is immediate, and does not wait for the law.

Frequently asked