Where does the value of AI output actually come from?

A commodity has use-value: a property of the object that makes it useful to anyone who has it (a hammer hammers, a textbook teaches). The use-value is intrinsic to the object — it is what the thing is for. Marx's exchange-value layers on top of this: the price the object commands in the market.

Intelligence-tokens have no analog of use-value. The same AI-generated paragraph can be invaluable to one reader and useless to another, depending on what the reader brings — context, prior knowledge, ability to evaluate, willingness to act on it. The "value" is not a property of the token. It is a property of the relation between the token and the receiver. Two readers receiving the identical output get different values because the value is constituted in the receiving, not in the produced object.

This is structurally unlike how we have historically valued knowledge goods. Books, lectures, consultations have intrinsic content-value that varies by reader but is at least anchored to something the producer made. Intelligence-tokens have no anchor: the producer is not a person, the production is not deliberative, and the artifact does not have stable content the way a written sentence does. What was produced is a string that licenses certain receiver-side reconstructions; the value lives in the reconstruction.

This is what gives AI output its peculiar status in markets. Pricing intelligence-tokens by content fails because content is not where the value sits. Pricing them by access (per-token billing) is closer but still misses — receivers with different abilities extract different value from the same access. The likely emergent pricing model is by outcome or by contextual fit, both of which are receiver-side measurements. Is the LLM a tool or a new form of intelligence itself? explains why this is the case at the medium level: a medium has no content of its own.

Source: Tokenization of Intelligence