Who Pays to Check the Machine

Generative artificial intelligence has collapsed the cost of producing plausible assertion while leaving the cost of establishing whether an assertion is true substantially unchanged. This paper argues that the resulting asymmetry, and not the automatable share of tasks, governs the economic effect of the technology in the sectors where it is most heavily adopted; and it argues that the asymmetry does not resolve itself, because verification cost is not the kind of cost that amortises. The argument has three parts. The first is a re-specification. In markets for goods whose quality the buyer cannot assess even after consumption—law, audit, medicine, compliance, scholarship—the binding constraint on what a production technology delivers is the cost of verification and not the cost of production, so a supply shock to the second does not become productivity growth. It becomes verification rent, a redistribution of surplus towards whoever can certify, and a re-allocation of liability towards whoever cannot. The growth literature has argued at length about how much labour the technology displaces. It has not asked what happens when the marginal cost of a claim falls to zero and the marginal cost of checking it does not. The second part is the paper's central theoretical claim and it is a claim about cost structure. The dominant explanation of the present productivity shortfall treats the verification and governance practices that must be built around a new technology as intangible capital—a stock, accumulated once, thereafter yielding returns, and therefore transitional. This paper argues that verification cost divides, and that the division matters more than the total. Establishing that a cited authority exists is mechanisable and does amortise: it is a stock, and building the register is the investment. Establishing that a real authority supports the proposition it is advanced for is a fresh judgement about a fresh pairing every time the pairing is made. That is a flow, it is the larger part, and it does not amortise. A discriminating test is stated: does the measured verification burden in a sector fall with cumulative experience of the technology, or scale with the volume of output produced? The third part is an original measurement, and its result is a negative one about measurement itself. Using the Retraction Watch corpus against Crossref annual publication counts, the paper computes, for every publication cohort from 2005 to 2023, the proportion of journal articles retracted within a fixed window after publication. Over two decades in which similarity detection, image forensics, post-publication review and paper-mill screening were all built out, the three-year detected retraction rate rose 6.3-fold, and the rate for integrity-type retractions specifically rose 11.4-fold, from 0.55 to 6.25 per ten thousand articles published. The integrity share of retractions rose from 38 per cent across the first five cohorts to 70 per cent across the last five. That series admits of two readings. Either the production of defective work rose roughly elevenfold, in which case the record is deteriorating quickly; or production was roughly stable and detection improved elevenfold, in which case the 2005–2015 record contains an undetected residue of about ten times what was found. Both are serious and they are not the same. Nothing observable from inside the system distinguishes them, because separating them requires knowing the rate at which defective work is produced, which is precisely the quantity the verification constraint makes unobservable. A verification apparatus operated for twenty years over a corpus it substantially controls cannot determine, from its own output, whether the underlying rate of defective production has risen or fallen. That is the paper's strongest single result, and it is a result about what the measurement cannot reach. The paper closes with a proposal that follows from the analysis rather than from preference: a verification duty which divides the cost of checking rather than moving it. The component a register can discharge is allocated to the party that captures the saving from cheap production; the component no register can discharge is left, expressly and cumulatively, with the professional. It is framed as a statutory duty rather than an incremental extension of the common law, for reasons given at §20.1A. The proposal is positioned against the transparency obligations of the Artificial Intelligence Act, which came into application on 2 August 2026, and against the new Product Liability Directive, and the residual claim is stated at the width those instruments leave. A companion paper applies the same cost structure to a different institution. Where this paper asks what happens to a market when producing a claim becomes cheaper than checking it, the companion asks what happens to a body of law when taking an action becomes cheaper than proving who took it. The two share a structure and not a cause, and neither depends on the other. Preprint; not yet peer reviewed.