When Generation Becomes Free, Attention Becomes the Scarce Good

The hall at five in the morning

A wholesale fish auction is one of the oldest working solutions to a problem that is usually discussed as though it were new. The catch arrives in the dark. It is graded by size and bedded in ice before the floor opens. Buyers walk the rows for perhaps twenty minutes, lifting a gill cover, pressing a flank, checking the clarity of an eye. Then the auctioneer starts, and lots that took a boat three days to land are priced in a few seconds each.

Almost nothing in that hall is there to produce fish. The ice, the grading, the shallow trays that hold one layer rather than three, the raised rostrum, the fixed opening hour, the hanging balance — all of it exists to make quality cheap to judge. The institution is a verification-cost-reduction machine, and its architecture is legible because it is physical.

This is the frame worth holding when a technology makes a class of output effectively free to produce. The interesting variable is not the supply of the thing. It is what happens to the cost of checking it, and to the institutions whose entire function was to keep that cost low.

Simon’s sentence, and why it is older than the network

The canonical formulation is Herbert Simon’s, delivered in 1971 in a volume on computers and the public interest, and it is more precise than its usual paraphrase. Simon began from an analogy about rabbits and lettuce — a rabbit-rich world is a lettuce-poor world — and then applied it directly:

in an information-rich world, the wealth of information means a dearth of something else: a scarcity of whatever it is that information consumes. What information consumes is rather obvious: it consumes the attention of its recipients. Hence a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it ^[1].

Two features of this deserve emphasis. First, it is a resource-allocation claim, not a lament about media quality. Simon’s next move is not to complain about noise; it is to ask how an organisation should be designed when attention is the binding input. Second, it contains no reference to any particular channel. The claim is about a ratio — information available to processing capacity — and that ratio can be pushed in either direction by printing, by telephony, by databases, or by anything else. This is why the framing predates the internet and survives it. It was already true of a 1960s corporate information system, and it will be true of whatever succeeds the current one.

Falkinger later formalised the same intuition, treating limited attention as the scarce resource in an information-rich economy and modelling what happens when the number of signals competing for a receiver exceeds that receiver’s capacity to process them ^[8]. The formal result matters less here than the framing: attention is a capacity constraint on a process, and the process it constrains is not consumption. It is verification.

The asymmetry that actually binds

Separate two costs that ordinary language runs together. Let one be the cost of producing a claim, and the other the cost of establishing whether that claim is true, or good, or what it purports to be. Call them generation cost and verification cost.

For most of recorded economic history these two moved together, at least loosely, within any given class of claim. A forged coin required a die-sinker. A fabricated ledger required someone who understood bookkeeping. A plausible technical report required someone who could write one. The correlation was never a law; it was an artefact of the fact that both production and imitation drew on the same scarce human skills.

An accounting identity makes the consequence visible. Let $G_t$ be the number of claims placed before some audience in period $t$, let $A_t$ be that audience’s total attention budget, and let $c_v$ be the average attention cost of verifying one claim. The number of claims that can be verified, and the fraction $\phi_t$ of what arrives that is actually checked, are then

This is bookkeeping, not a finding. Its only content is what it forbids. Simon’s point was that $A_t$ is bounded by hours and by human processing capacity, and that this bound is not responsive to technology in the way that $G_t$ is. If $G_t$ rises by an order of magnitude, $\phi_t$ falls by an order of magnitude unless $c_v$ falls proportionally or $A_t$ expands proportionally. Institutions are the mechanism by which societies have historically pushed on $c_v$. Nothing pushes usefully on $A_t$.

It is worth saying plainly what this framework does not license. It does not permit a claim about how much machine-generated text exists, because that quantity is not reliably measured and estimates of it circulate without provenance — which is itself a small instance of the problem under discussion. The structural argument does not need the number. It needs only the observation that $c_v$ and $G_t$ are set by different mechanisms.

Not every claim costs the same to check

Verification cost is not a constant across goods, and the standard taxonomy is due to Nelson and to Darby and Karni. Nelson distinguished search qualities, which a buyer can determine by inspection before purchase, from experience qualities, which are learned only by consuming the good ^[4]. Darby and Karni added a third category: credence qualities, which the buyer may never verify even after consumption, because doing so would cost more than the information is worth. Their paper drew the uncomfortable but correct conclusion that where verification is genuinely expensive, the socially optimal amount of fraud is not zero, since driving it to zero would cost more than the fraud does ^[5].

The fish on the ice is close to a pure search good. Gill colour, eye clarity, firmness and smell are cheap and reliable indicators, available to any trained buyer in seconds, and the auction hall is engineered to make them available: bright even light, one layer per tray, ice to hold condition, a window before bidding. The institution moved the good leftward along that taxonomy, and the auction format is viable because of it.

The relevant structural change when generation becomes free is that a large class of textual claims moves in the opposite direction. A written argument used to carry search qualities: fluency, structure, correct citation form and internal consistency were inspectable in seconds and correlated with the effort behind them. As those surface properties become free to produce, they stop being search qualities and the underlying quality becomes a credence quality — knowable, if at all, only by redoing the work. The reader’s inspection window still exists. What it inspects has stopped being informative.

What happens when verification is too expensive

Akerlof supplied the canonical model of a market in that condition. Where sellers know quality and buyers do not, buyers can rationally offer only the expected value of what is on offer; sellers whose goods are worth more than that price withdraw; the expected quality of what remains falls; the price falls again. The equilibrium is a fixed point of the form

where $v(q)$ is the buyer’s value of quality $q$ and $c(q)$ the seller’s reservation value. Only sellers whose reservation value lies below the prevailing price participate, and the price is the expectation over exactly that self-selected pool ^[2].

Two things about this result are routinely misstated. It is not a claim that bad always drives out good; it is a claim that the market for the high-quality segment can fail to exist, which is a different and worse outcome than low average quality. And Akerlof’s own paper does not end in fatalism. It identifies the counteracting institutions directly — guarantees, brand names, chains, licensing and credentials — as devices that let a seller commit to quality when quality is not observable. The economics of adverse selection has always been half a theory of institutions.

The condition for the failure is worth stating carefully, because it is more specific than “information is asymmetric”. Unravelling requires quality dispersion, private information on the seller’s side, and a verification cost high enough that no buyer finds it worth paying. Where any of the three is absent, the mechanism does not bite. The fish auction defeats it by attacking the third: a floor inspection costs a buyer a few seconds, and a grader has already done the expensive part once on everyone’s behalf.

Signals work because they are expensive

Spence’s contribution was to explain how information can cross the asymmetry without direct verification. An agent takes a costly action — in his original case, education — and the action separates types only under a specific condition: the marginal cost of the signal must be lower for the type the signal is supposed to identify. In the standard notation, with signal level $s$ and type $\theta$ increasing in ability, the single-crossing condition is

and it is this inequality, not the content of the signal, that does the work ^[3]. Spence’s model famously does not require the signalling activity to be productive at all. It requires only that it be differentially costly.

That is the precise statement of what happens when effort signals become cheap to fake, and it is sharper than the usual complaint. The problem is not that fakes appear; forgery is as old as the hallmark. The problem is that when a signal becomes free for every type, the cross-partial goes to zero, the separating equilibrium collapses into pooling, and the signal conveys nothing — no matter how good the underlying work of the high type actually is. The high type loses the ability to prove it. Nothing about the signal’s content protects it, because the content was never the mechanism.

Applied to written output, the list of signals whose cost differential has narrowed is long and familiar: length, polish, structural completeness, correct reference formatting, fluent register, apparent breadth of reading, responsiveness within hours. Each was once a plausible proxy for effort or competence. Each was a cost differential, not a causal indicator, and cost differentials are exactly what a general-purpose generation technology erodes.

Reputation is a bond, not a badge

The obvious repair is reputation, and Shapiro showed why it works and what it requires. In his model, buyers cannot observe quality before purchase, so sellers have an incentive to cut quality and collect short-run gains before buyers adjust. The equilibrium that prevents this has high-quality goods selling at a premium above cost, and that premium performs two jobs simultaneously: it deters quality-cutting, and it compensates the seller for the earlier investment in building a reputation ^[6].

The mechanism is therefore closer to a posted bond than to a badge. The seller has sunk something recoverable only through continued honest trade, and cheating forfeits it. This makes the failure condition specific and testable: reputation stops disciplining behaviour when identities are cheap to create and cheap to abandon, because then nothing has been sunk. Pseudonymity itself is not the problem — a durable pseudonym can carry a large bond. Non-durability is the problem. An identity that can be discarded and reissued at negligible cost cannot post a bond, and a reputation system built on such identities is decorative.

Certification, and the certifier’s own incentives

Dranove and Jin survey the theory and practice of quality disclosure and certification, and the questions they organise the literature around are exactly the ones that matter here: why sellers do not voluntarily reveal quality even though unravelling arguments suggest they should, whether mandatory disclosure improves welfare, and whether independent certifiers can be trusted ^[7]. The last question is the one that generalises. A certifier reduces verification cost for everyone downstream by paying it once and issuing a portable token. But the certifier is a firm with its own incentives, and it is very often paid by the party being certified.

Attention as bottleneck, attention as commodity

These are two different economic objects and conflating them causes real confusion.

Attention as a bottleneck is a capacity constraint on a process. Its price is a shadow price: what it costs the constrained party, in foregone alternatives, to spend a unit of capacity here rather than elsewhere. Nobody receives this price. It is an opportunity cost internal to the reader.

Attention as a commodity is something else: a quantity assembled by an intermediary and sold to a third party. It has an actual market price and an actual revenue line. The IAB and PwC report puts United States internet advertising revenue at a record figure of roughly 259 billion dollars for 2024 ^[16]. That is the price of the commodity, and it is paid by advertisers.

The structural consequence follows immediately. The party purchasing attention is not the party whose verification budget is being consumed, so the allocation that maximises revenue in the attention market need not be the allocation that maximises verification in the epistemic process. There is no contradiction here and no conspiracy required; it is an ordinary externality. When accuracy is what attracts attention — in professional trade data, in credit assessment, in tools people pay for directly — the two allocations align. Where attention is monetised by a third party, alignment is coincidental. Any argument that markets will automatically solve the verification problem has to explain which of these two prices is doing the work.

Four institutional responses, and how each was gamed

The historical record on cheapened signals is not blank. It shows a repeating pattern, and the pattern includes the failure.

Hallmarking. The oldest recognisable response to an unverifiable quality claim is the assay mark. The United Kingdom’s Hallmarking Act 1973 makes “fresh provision for the composition, assaying, marking and description of articles of, or containing, precious metals, and as to agencies for the implementation and enforcement thereof” ^[13]. The design is instructive: verification is destructive and expensive, so it is performed once by a body with no stake in the sale, and the result is expressed as a mark that is itself costly to apply and hard to counterfeit. The attack surface is equally clear. Historically the response to hallmarking was not better forgery of the metal but forgery of the punch, transfer of genuine marks onto substandard articles, and importation of goods marked under weaker regimes. The mark, not the metal, became the target.

Appellations. Geographical indications extend the same idea to origin and process. WIPO defines a geographical indication as “a sign used on products that have a specific geographical origin and possess qualities or a reputation that are due to that origin”, and notes that protection is obtained by acquiring a right over the sign, which then lets the holder prevent use by anyone whose product does not conform to the applicable standards ^[14]. Two structural weaknesses follow directly from that definition. The right attaches to the sign, so enforcement is a trademark-like exercise against misuse rather than an inspection of quality. And the specification certifies origin and process, not outcome, so a producer can comply fully while making something mediocre. Boundary expansion, relabelling and blending are the recurring gaming strategies, and they exploit exactly the gap between the certified thing and the thing buyers care about.

Auditing. The financial audit is the industrial-scale version: a third party paid to verify, subject in turn to a public regulator. The regulator’s own data show how much residual verification failure the system carries. PCAOB staff reported that in 2023 inspections, “46% of the engagements reviewed in 2023 had at least one Part I.A deficiency”, and that at the United States Big Four firms the aggregate Part I.A deficiency rate “held steady at 26% in 2023 after previously jumping from 12% in 2020 to 16% in 2021 and 26% in 2022” ^[10]. The adjacent case of credit ratings makes the incentive explicit. Bolton, Freixas and Shapiro model rating agencies that understate credit risk to attract business while issuers shop for favourable ratings, and find that inflation rises when more investors take ratings at face value or when the agency’s reputation costs are low; they predict agencies “are more likely to understate credit risk in booms than in recessions” ^[9].

Peer review. The scholarly version of certification has been under sustained attack from organised fabrication. The joint COPE and STM research report on paper mills concluded that the submission of suspected fake research papers, frequently coupled with fake authorship, is growing and “threatens to overwhelm the editorial processes of a significant number of journals” ^[11]. The downstream count moved with it: Nature reported that “the number of retractions issued for research articles in 2023 has passed 10,000 — smashing annual records” ^[12]. Note that the mechanism of attack was not primarily better fake science. It was capture of the review process itself — fabricated reviewer identities, compromised editorial handling, sold authorship.

The generalisation is uncomfortable and consistent. Verification institutions do not eliminate verification cost; they concentrate it in a specialist, amortise it across many downstream users, and issue a portable token in place of the underlying work. Amortisation is what makes them valuable and it is also the vulnerability, because a single compromised certifier invalidates every token it ever issued. In each of the four cases the attacker went for the token, not the quality, and in each case the institution’s response was to make the token more expensive again: physically harder punches, accredited inspection bodies, an independent audit regulator, identity verification for reviewers.

What is genuinely new

Cryptographic provenance is the current candidate. The C2PA technical specification describes “a model for storing and accessing cryptographically verifiable information whose trustworthiness can be assessed based on a defined trust model” ^[15]. It is worth being exact about what this does. It binds a signed, tamper-evident record of an asset’s history to the asset, so that alterations to the record are detectable. It does not make the assertions in that record true. It relocates the trust question to the signer and the trust list — which is, structurally, the assay office problem again, in software. Same strength, same failure mode, and the same subsequent question of who accredits the accreditor.

If something is genuinely new, it is not fabrication and not forgery. It is that a very wide class of effort signals lost its cost differential at roughly the same time, rather than one signal at a time over decades. Historically, when a signal became cheap, there was slack: another costly signal was available to migrate to. The open empirical question — and it is empirical, not theoretical — is how deep the remaining pool of costly signals is. The plausible candidates are those whose cost is not informational: physical presence, staked capital, legal liability, a longitudinal track record that cannot be backdated, destructive testing, and independent replication. Each of these is expensive in a way that a generation technology does not touch, and each is correspondingly slow and exclusionary. That trade-off, rather than the volume of generated text, is the interesting design problem.

Three conditional predictions

These are speculation, separated deliberately from the theory above, and each is stated with a horizon and a condition that would falsify it.

First: through 2030, institutional demand should shift measurably from artifact-quality signals toward identity-and-liability signals — verified identity, named accountability, staked bonds, insurable warranties. This is disconfirmed if hiring, procurement and publication processes continue to weight artifact inspection as heavily as they did in the early 2020s, or if an automated verifier with independently audited accuracy makes artifact inspection cheap again, which would push $c_v$ down and remove the pressure.

Second: through 2030, verification should move from a bundled externality to a priced input — organisations paying explicitly for checking rather than absorbing it in reviewers’ unpaid time. This is disconfirmed if checking remains unpriced and unstaffed while output volumes rise, which would instead indicate that institutions are absorbing the shift by lowering $\phi_t$ and accepting the resulting error rate.

Third: certifier concentration should increase, and with it the systemic cost of any single certifier’s failure. This is disconfirmed if the number of independent accredited certifiers per domain grows, or if provenance trust lists remain plural and mutually checking rather than converging on a small number of roots.

The constraint that binds

The auction hall does not exist because fish are scarce. It exists because quality is expensive to establish and value has to be set in seconds, and every physical feature of it — the grading, the ice, the single layer per tray, the fixed inspection window, the balance on the rail — is capital invested in making verification cheap.

The equivalent capital in an information economy is institutional rather than physical, and it has the same economics: concentrate the expensive check, amortise it, issue a token, and then defend the token against the people who would rather obtain it than earn it. That defence has never been won permanently. It has been re-won, repeatedly, by raising the cost of the token after each collapse.

The constraint that binds now is not the ability to produce text. Production was never the scarce factor in Simon’s argument and is not the scarce factor in the market-failure literature that followed him. The scarce factor is the capacity and cost of checking, and the historical record is unambiguous about what has absorbed shifts in that cost. Not the technology that caused them. The institutions built afterwards, at considerable expense, by people who understood that a signal is only worth what it costs to fake.