The Scarcity of Reality

Imagine commissioning a report on what customers bought last week. One supplier can analyse recorded transactions. Another can generate a convincing estimate from an existing model of customer behaviour. Both may provide useful information, but they are not selling the same evidential relationship with the market. Only the first is necessarily reporting observations from the week in question; the second is useful to the extent that its assumptions and inputs justify the estimate.

Generative AI makes that distinction commercially important. An organisation can obtain more descriptions, scenarios and plausible explanations without obtaining more independent evidence about the world they describe. The question is not whether synthetic information has value. It is which decisions still depend on observations that generation alone does not supply, and what businesses will pay to obtain them.

This is not a claim that information previously required witnessing an event. Mathematics, deduction, fiction and simulation long preceded generative models. Nor does a model have to repeat familiar statements: it can combine knowledge, derive implications and propose something genuinely new. The narrower distinction concerns fresh external evidence. A novel prediction about an event is not itself an observation that the event occurred.

Suppose a manufacturer uses simulations to explore possible failures of an established component. Those simulations could identify useful tests and reduce the number of physical experiments required. A measurement from an actual test serves a different function because the result can expose a condition missing from the simulation. The two activities are complements when each is used for the question it can answer.

A similar distinction applies to generated customer responses. A synthetic population might help a business explore how different preferences could affect demand. A small real-world trial might instead reveal a preference or obstacle the model did not represent. Neither the number of generated responses nor the number of real participants establishes quality by itself. Sampling, measurement and relevance determine how informative either exercise is.

Research on model collapse illustrates one way in which the distinction can matter. Shumailov and colleagues, writing in Nature in 2024, demonstrated deterioration under recursively generated training-data regimes, with uncommon parts of the original distribution particularly vulnerable. That supports concern about indiscriminate replacement of evidential material with generated output; it does not establish that every use of synthetic data damages learning.^[1]

Indeed, Gerstgrasser and colleagues found that retaining and accumulating original and synthetic data could avoid collapse in the settings they studied. The findings concern particular models and training arrangements, not a guarantee that any mixture is safe. Taken together, the studies direct attention towards how generated material is selected, retained and validated rather than towards a simple division between good real data and bad synthetic data.^[1][2]

The economic question begins where particular observations remain difficult to obtain and useful substitutes remain limited. An authenticated record of an unrepeatable event, a fresh measurement from inaccessible equipment or a scarce set of observed outcomes can have value that another generated account does not replace. Yet calling such evidence useful does not establish that its market price must rise as generation becomes cheaper.

For a price increase to follow, buyers would have to want more of that particular evidence relative to the supply available on acceptable terms. Alternatives would matter, as would exclusivity, timeliness and the ability to use the material. A widely shared dataset might become essential to many applications without generating substantial revenue for its original collector. Importance to users and income to producers are different outcomes.

The supply side can change too. In a hypothetical laboratory, AI might help choose informative experiments, organise instruments and interpret measurements more efficiently. Observation could then become cheaper at the same time that its applications expanded. A growing demand for grounded evidence would not necessarily outrun that improved supply. The same technology could increase the value of collecting evidence while reducing the price of obtaining it.

Acquisition and distribution must also be separated. Conducting an experiment may require scarce time or material, but sharing its recorded result need not require repeating it. A sensor network could be expensive to establish while supplying further copies of an existing measurement at little additional cost. Freshness, rights of access and the ability to verify origin may therefore matter more commercially than whether a dataset is described simply as real.

Provenance helps answer some of these questions, but not all. C2PA's Content Credentials are designed to convey signed information about origin and history and help identify alteration. The specification explicitly explains that provenance information alone cannot establish whether content is true, accurate or factual. A generated image can have valid credentials, while a photograph can authentically record something staged or be interpreted misleadingly.^[3]

This means an evidential chain is more than a digital signature. A buyer assessing a measurement may need to understand how the instrument was calibrated, what was sampled and which transformations occurred before the conclusion was presented. An intact history can improve that assessment without certifying every inference made from the underlying material. Authentication should make scrutiny easier, not become a substitute for it.

The potential commercial value is therefore specific. A business with continuously refreshed transaction data could detect changes that an older model fails to anticipate. An organisation able to conduct reliable tests might help customers evaluate uncertain predictions. Such capabilities could justify higher prices than weaker alternatives, but only where customers value the difference and comparable evidence is not readily available elsewhere.

There is an equally plausible market in which verification becomes routine and inexpensive. Buyers could expect provenance as a standard feature rather than pay extra for it. Automated collection might make some forms of observation abundant. Competition could pass the gains to users even as the total amount of valuable evidence increases. None of those outcomes would make empirical grounding less important.

The practical question for information businesses is consequently not whether reality as a whole is becoming scarce. It is where fresh, relevant and credible evidence remains a bottleneck, whether technology can ease that bottleneck, and who can retain value from doing so. A forecast of a universal scarcity premium would obscure the very differences that make some assets commercially distinctive.

For users, the corresponding question is what a piece of information actually establishes. A simulation may be sufficient to screen options. A decision involving an unfamiliar condition may require an external test. An existing observation may be relevant but no longer current. Better systems can combine these forms of information and show their limits, rather than assigning confidence according to fluency or apparent detail.

Generative intelligence could therefore increase both the demand for evidence and the efficiency with which evidence is obtained. In some markets, scarce observation may become more valuable; in others, grounded information may become cheaper and more widely accessible. The durable distinction is not between human truth and machine invention. It is between an account supported by the evidence needed for its purpose and an account whose plausibility is doing the work of that evidence.