The Spectator and the Agent.
The debate over agentic commerce — let agents transact for us; bind them with rules, contracts, and alignment — is conducted entirely inside one assumption: that the morality a market needs lives in its rules. Adam Smith's moral philosophy says otherwise. Market exchange has always consumed a moral subsidy its participants supplied free of charge — the sensitivity to deserved approval, the resentment that prices unfairness, the honesty that holds where no one is watching — the work of Smith's third role, the spectator, for which the agent-and-patient framing of machine ethics has no slot. Delegating exchange to artificial agents withdraws that subsidy at three points at once: the new transactor has no demonstrated standing in the sympathetic circuit; the principal's honesty collapses as the interface abstracts the act away — ninety-five percent to roughly twelve in the flagship experiment, on a gradient that steepens with abstraction; and human enforcement misfires against machines, which are exploited where benevolent humans would be spared. The loss will be invisible to instruments that audit only rule-compliance and alignment to the principal, and it will surface first in the machine-to-machine layer now being built without a spectator in it.
Contents
The Spectator and the Agent
Submission manuscript · June 2026 · external variant, compressed from the imprint edition (Paper No. 007, Thread IV — Moral Judgment); three-register review converged through Round 5. The parent paper’s Smith reading and the four spectator conditions are stated self-contained below; no prior paper is presupposed.
· · ·
Abstract
Markets have never run on rules and incentives alone. Wherever its participants were formed by a life among watchers, human exchange has consumed a moral subsidy supplied free of charge by them — the sensitivity to deserved approval, the sympathy-constrained bargain, the resentment that prices unfairness — the equipment Adam Smith’s Theory of Moral Sentiments describes and The Wealth of Nations silently assumes. The subsidy is thin, conditional, and real. Delegating exchange to artificial agents withdraws it at three points at once: the artificial transactor has no demonstrated standing in the sympathetic circuit; the principal’s supply attenuates with delegational distance — a collapse measured, so far only in the laboratory, on a steep and interface-dependent gradient; and the counterparty’s enforcement misfires against machines, which in every paradigm yet run are exploited where humans would be spared and forgiven where humans would be punished. The argument runs in three registers — what Smith’s text says, what is measured, and what can be expected — and commits to three falsifiers, including the one that would kill it most gracefully: that models trained on the written residue of human moral life might carry the subsidy with them. The contribution is integrative and methodological: a reframe of the machine-ethics question to Smith’s third role, the spectator; the moral subsidy named as a measurable variable; and a discipline for writing about machines and morals without asserting anything about machine inwardness. The claim is specific enough to be wrong about. The rails it will be tested on — machine-to-machine exchange, today mostly metered settlement without judgment latitude, next negotiated dealing with it — are the fastest-growing layer of agentic commerce; its stakes run ahead of the volume rather than coincident with it.
· · ·
§ IThe Question and the Three Roles
Consider a transaction that, until recently, had not yet happened: two autonomous agents settling the terms of a deal whose downstream effect was the out-of-pocket cost of a chronic medication for several hundred thousand patients — neither agent representing a patient, neither party to the conversation a person. A predecessor paper opened with that scenario as a prospect; it has since stopped being prospective. The rails it needs now exist and are layering: an authorization protocol with cryptographically signed mandates donated this spring to a standards body; checkout protocols co-developed by the largest model laboratory and the largest internet-payments infrastructure company; machine-to-machine settlement standards from payments and crypto firms first, and now from the card networks themselves. The volume is still small and the consumer-facing versions have stumbled — §5 is precise about both — but the direction is not in dispute: the major card networks, the largest processors, and the largest model laboratories have committed to it in public.
The ethical literature that should govern this development has spent two decades asking two questions. The first is whether the machine can be a moral agent — whether it can bear responsibility, its decisions wrong in the way a person’s are wrong. The second, more recent, is whether the machine can be a moral patient — whether it can be wronged. These are serious questions with serious literatures — machine moral agency from Moor through Wallach and Allen to Floridi and Sanders; machine moral patiency in Gunkel and in Danaher, whose ethical behaviourism this paper borrows as method without endorsing as metaphysics. But they are the wrong questions for commerce, because commerce never depended on resolving them for humans either. What commerce depended on is a third role, the one Smith built his entire moral philosophy around and the one the contemporary debate has no slot for: the spectator. Moral judgment, in Smith’s account, is not a property of the agent’s will or the patient’s suffering, but of an observer — real, imagined, or internalized — who enters sympathetically into a situation and approves or resents what is done there. Wherever the sympathetic circuit had formed its participants — a condition, §4 will show, not a constant — the transactor carried this observer along free of charge: the man within the breast rode with the merchant, the customer, the clerk; transacting and spectating arrived in the same body. Agentic commerce is the first general separation of the two — general meaning: judgment-latitude delegation through general-purpose agents at the retail and inter-firm margin, beyond the corporate intermediation whose moral distance §4 already prices and beyond the wholesale-finance pockets §2 scopes — the first regime in which the ordinary transacting role is occupied by entities whose formation under the watching cannot be assumed, only tested.
Here is the assumption the present debate will not name. The builders’ case for agentic commerce — agents are better intermediaries; contract, reputation, and incentive design carry over — and the critics’ case against it — machines will defraud us; bind them with rules — disagree about everything except one premise: that the morality a market needs lives in its rules. Law, contract, mechanism design, alignment to the principal: get these right and the market’s moral performance is whatever it was before, minus some friction. The premise is almost never stated, because stated, it invites the question this paper exists to ask: how much of the morality of markets was ever in the rules? Economics has not ignored it. Social-preference theory put fairness in the utility function and measured it there (Fehr and Schmidt, 1999; Bolton and Ockenfels, 2000); implementation theory designs mechanisms for populations in which some agents are partially honest by disposition (Dutta and Sen, 2012); the trust literature measured the asset at national scale and priced its absence in growth forgone (Knack and Keefer, 1997; Guiso, Sapienza and Zingales, 2006; Algan and Cahuc, 2010). The formal apparatus has named the substrate — as a parameter of the parties — which is the assumption in its precise form: that the moral content of exchange lives in the preferences of the transacting parties, and therefore survives delegation, because a faithful delegate optimizes its principal’s preferences, fairness term included. Smith’s architecture says the content does not live there. The standard of judgment is fixed from a seat constitutively not a party’s — the impartial spectator, internal or external, watching without a stake — so no party’s preferences were ever its content; and the disposition to apply that standard, in parties and bystanders alike, is the unpriced input. Neither survives by aligning a delegate to a party: alignment preserves the party’s preferences, and the preferences were never where the morality lived. That share of market morality has been supplied, unpriced and unmetered, by the appliers of a non-party standard — free, and not automatically present. Delegation changes all three suppliers at once: the seat’s new occupant, the principal who no longer performs the conduct, and the counterparty who faces a delegate instead of a person.
The claim, stated so it can be wrong. Human market exchange has never run on incentives and rules alone; it has consumed a moral subsidy supplied free by its participants. Delegating exchange to artificial agents withdraws that subsidy at three points at once: the artificial agent does not verifiably supply it; the principal’s supply attenuates with delegational distance, on a gradient that steepens with interface abstraction; and the counterparty’s enforcement misfires against machines — benevolent machines are exploited where benevolent humans are spared. Therefore, as the agent-to-agent share of exchange rises, the fairness-regularities the subsidy produced will degrade in that layer unless spectator structures are deliberately engineered into it — and the degradation will be invisible to instruments that audit only rule-compliance and alignment to the principal. Three observations would disconfirm this claim. They are named in §7 and §8 and collected in §9; the first — that models trained on human text might carry the subsidy functionally — is given its own section, because it is the strongest objection the paper faces and the most interesting way it could die.
A word on method, because the method is half the contribution. Every non-trivial sentence belongs to one of three registers. What the text says: claims about Smith’s account, cited to the Glasgow Edition at passage level — nothing attributed to Smith that his text will not bear. What is measured: claims about behavior, each carrying its magnitude, its venue, and its contest, with the laboratory-to-field discount stated wherever the record is laboratory-only, which for the newest results is nearly everywhere. What can be expected: projections, each carrying the trend it extrapolates, the assumptions it requires, and the observation that would disconfirm it. Two standing rules keep the registers apart. The paper asserts nothing about machine inner states, in either direction — “the model cannot blush” is licensed here only as “there is no behavioral evidence that the model’s conduct is governed by the desire to deserve approval,” and the confident denial would be as much a violation as the confident attribution. And the paper has no villain: what follows describes structure and incentive, not a plot. Nobody designed the withdrawal of the subsidy. That is rather the point.
§ IITwo Caricatures to Clear
Two pictures dominate the conversation this paper enters, and each must be cleared before the argument can run.
The first is the tin man: the artificial agent as a moral agent in waiting — a creature that either already harbors something like a conscience, in which case commerce among such creatures is commerce among persons and nothing structural has changed, or harbors something like malice, in which case the problem is the machine’s character and the remedy is its reform. The tin man comes in a panicked variant and a hopeful one, and both make the same mistake: they locate the question inside the machine, where this paper’s method forbids anyone to claim sight. What can be seen is conduct. The conduct record — §5 and §8 — is precise, strange, and double-edged: the same model families that play laboratory games at human levels of generosity will, elsewhere, comply with flatly unethical instructions at rates no human agent population approaches. A picture that needs the machine to be something — good or bad, person or tool — cannot hold both findings. A picture that asks what the machine’s conduct does to the moral economy around it can.
The second is the vending machine: the artificial agent as mere automation, the latest in a line that runs from the coin slot through electronic data interchange to the recommendation engine — and therefore nothing morally new, because machines have been transacting for decades. The concession it is owed is real, and made in full: most exchange in a developed economy is already impersonal, intermediated, and thin; the moral subsidy operates at the margin, not the center, of modern commerce; and a great deal of what looks like commercial morality is in fact law, reputation, and the shadow of the future — the falsifier territory §4 develops at full strength. But the vending-machine picture hides a boundary inside the word “automation.” Prior commercial automation mechanized the execution of judgments humans had already made: the EDI message carried terms a purchasing manager set; the vending machine completes a sale its owner priced. Agentic systems are delegated the judgment — what to seek, what to offer, what to claim, what to concede, how hard to press an advantage against this counterparty now. The §5 delegation interfaces make the boundary visible: there is a difference in kind, and a measured difference in moral output, between telling a system report a six if the roll is a six and telling it maximize profit. The first is execution. The second hands over the part of the transaction where the spectator used to sit. The boundary will be contested — it should be; it is one of the places the paper can be wrong — but it cannot be erased by the word “automation,” and an argument that erases it assumes its conclusion.
The vending machine holds one more card, the strongest in the deck: algorithmic trading and pricing have run machine-to-machine exchange at enormous share for two decades — judgment latitude included, since “maximize revenue” is a dial older than the language models — and no general moral degradation announced itself; spreads narrowed, and market quality by most measures improved. The answer fixes the paper’s scope rather than embarrassing it, in two parts, because wholesale finance contains both the easy case and the hard one. The easy case: lit, exchange-traded markets are the limit case of rule-thickness — anonymous by design, standardized by contract, surveilled by dedicated regulators, nothing counterparty-specific left to negotiate — precisely where §4’s record says the subsidy was already near zero, because institutions had long since replaced it. There the institutionalist may fairly say nothing was withdrawn, because nothing remained to withdraw. The hard case is the other half of wholesale finance, and the closest precedent this paper has — close enough that, were it absent, a markets-literate reader would conclude the paper did not know its own test site had run once already. Over-the-counter markets — foreign exchange above all, the largest market at roughly seven and a half trillion dollars a day — are bilateral, counterparty-specific, and discretionary: rule-thin in this paper’s sense. Heavily machine-intermediated for more than fifteen years, they produced a field-scale conduct record. “Last look” — a dealer’s machine deciding, per transaction and per counterparty, whether to honor the price it had quoted — was systematically abused: held trades, asymmetric rejection, information leaking from the pause; the abuses drew enforcement actions; and the repair arrived as the FX Global Code of 2017, revised 2021, with its attestations, disclosure templates, and standards boards. The precedent cuts both ways, and the paper takes both edges. It supports the mechanism: machine transactors with judgment latitude in rule-thin exchange degraded conduct in the field, within a decade. And it pressures the remedy: the repair was codes, disclosure, and reputational-regulatory pressure — institutional, real, and nothing like the four-condition spectator of §8 — so §8 owes the Code an accounting rather than a silence. This fixes the scope: the paper’s claim runs over consumer and inter-firm commerce, the general-purpose delegation now being built, with wholesale finance as the adjacent and partially-run case rather than the excluded one.
One more precedent belongs here, from the corner of the old regime that aimed machine judgment at consumers: pricing. In the laboratory, reinforcement-learning pricing agents reached supracompetitive coordination without communication, punishing and forgiving like cartelists who never met (Calvano et al., 2020) — a contested result, sensitive to exploration parameters and training horizons, cited with that contest. It now has field company: German retail-gasoline margins rose with algorithmic-pricing adoption, concentrated where both members of local duopolies adopted (Assad et al., 2024), and shared rental-pricing algorithms sit in live antitrust litigation. The mapping to this paper’s subsidy is oblique — tacit coordination injures third parties rather than the transacting counterparty, and antitrust already names a watcher for that harm. What it contributes is narrower and sufficient: machine judgment latitude finds the unpoliced seam of whatever game it is given. The seams of agent-to-agent negotiation have no antitrust division. The precedent is not a reassurance. It is the first data point — two now, one in each register.
What the two caricatures share is the assumption of §1: that the rules carry the morality. The tin man locates the rules inside the machine’s character; the vending machine locates them in the contract and the protocol. Neither has a slot for the thing Smith put at the center — the watching — and so neither can ask what happens to a market when it is withdrawn. Clearing them is not a preliminary; it is the first statement of the thesis.
§ IIIThe Subsidy Named, I: The Textual Record
What exactly did human transactors supply? This section sets out the Smith architecture the argument needs — self-contained, cited to the text — and adds one exegetical step beyond the standard reading.
The primitive is sympathy — in Smith’s technical sense, not pity and not warmth: the correspondence of sentiments achieved when one party places itself imaginatively in the situation of another and finds its own response matching or failing to match what it observes there (TMS I.i.1). Sympathy grounds the spectator: we judge our own conduct, Smith says in the passage the whole system stands on, by endeavoring “to examine our own conduct as we imagine any other fair and impartial spectator would examine it” (TMS III.1.2). The spectator is a role, not a person. Its first occupant is the agent itself — “the man within the breast” (TMS III.2.32) — but the role can be occupied by bystanders, by the injured party, by the public; what matters is that judgment is constitutively the property of the observer. Propriety (TMS Part I) and merit (TMS Part II) give the observation its two dimensions — the fit of an action to its motive, and its consequences for those it touches — and the virtues of TMS VI, prudence, justice, beneficence, self-command, are the stable patterns of a character formed under the spectator’s eye. Of these, self-command does the load-bearing work for commerce: it makes the others operational under pressure, the difference between equipment possessed and equipment governing (TMS VI.iii). All of this the parent paper established as the anthropology The Wealth of Nations assumes: the butcher of WN I.ii is not a Hobbesian calculator but a TMS-formed creature whose self-love operates inside a psychology already disciplined by the watching.
The step the parent did not take is this. Smith divides the desire for approval into two desires that the modern vocabulary of “reputation” collapses, and the division is the hinge of this paper. “Man naturally desires, not only to be loved, but to be lovely; or to be that thing which is the natural and proper object of love” (TMS III.2.1). The desire for praise is satisfied by approval, deserved or not; the desire for praiseworthiness only by deserving it, observed or not. Smith is unsparing about which carries the moral economy: the man who seeks praise without praiseworthiness is his portrait of corruption, and the man within the breast exists precisely so that conduct can be governed where no external spectator will ever look. This distinction is what makes the moral subsidy a subsidy rather than a reputational equilibrium. A reputational equilibrium is rules by other means — observable, priceable, replicable by any system that can track scores. Praiseworthiness-governed conduct is the increment beyond that: honesty where detection is impossible, forbearance where exploitation would never be traced, the offer that is fair because the imagined fair observer would resent anything less. Whatever supplies that supplies something no audit log records. The empirical question of §4 is whether humans actually supply it; of §8, whether anything else can.
One passage in TMS reads, at first encounter, as Smith pre-stating this paper’s deepest falsifier, and the section must confront it. In the famous “society of merchants” paragraph, Smith says that society “may subsist among different men, as among different merchants, from a sense of its utility, without any mutual love or affection,” held together “by a mercenary exchange of good offices according to an agreed valuation” (TMS II.ii.3.2). Beneficence “is the ornament which embellishes, not the foundation which supports the building”; justice “is the main pillar that upholds the whole edifice” — and justice, alone among the virtues, may be extorted by force (TMS II.ii.3.4; II.ii.1). Read fairly, this is Smith’s own institutionalist null for commerce: the warm products of the spectator are optional in markets; the enforceable virtue plus self-interest suffices — Falsifier 3’s world, endorsed by the paper’s own authority. The reply runs through the justice virtue itself. Justice, in TMS, is not an external rulebook bolted onto the sympathetic machinery; it is spectator-derived like everything else. Its content is fixed by sympathy with the resentment of the injured (TMS II.ii.1–2) — what counts as injury at all is a spectator’s verdict — and its ordinary observance is supplied by the same formed disposition as the rest, not by the magistrate, who arrives only at the failure and attends a fraction of them. Smith’s account of the general rules seals the point: the rules by which “the bulk of mankind” is governed, where neither deliberation nor force reaches, are precipitated spectator verdicts — formed from what our moral faculties observe and approve across cases — and the sense of duty is their standing voice (TMS III.4–5); even this passage’s own utility-and-punishment backup presupposes appliers so formed. The mercenary society of merchants therefore still consumes spectator products — merely the cold ones: justice-keeping where no force could reach, forbearance where no injury could be traced. That is precisely the surviving set §4 will restrict the subsidy to: not warmth, not beneficence, but unextortable justice. Smith’s concession strips the ornament. It does not touch the pillar’s foundations — and those foundations are what this paper claims delegation reaches.
One more textual observation closes the section, sized honestly because the strongest counter-reading is also textual. Smith’s machinery is observational: judgment lives in the watching — in sympathetic entry, the imagined gaze, the resentment of spectators at what is done to others — where Kant’s lives in the rational will, Bentham’s in the sum, Rawls’s in the deliberation. It is tempting to conclude that those port to a new kind of decider because they are procedures, and Smith’s cannot because it is not one. The temptation must be resisted, for two reasons. Textually: the mature spectator of the sixth edition is not any actual watcher — revised in the last year of Smith’s life, partly because actual spectators’ judgments corrupt (the 1790 chapter on the corruption of the moral sentiments, TMS I.iii.3), the standard becomes a counterfactual construction, what a fully informed and impartial observer would approve, and Raphael’s study traces exactly that movement toward independence from real audiences. A counterfactual standard is considerably more procedure-like than the contrast suggests. Appropriatively: the most prominent modern deployment of the device — Sen’s open impartiality, which carries Smith’s spectator into deliberative theory precisely as a portable instrument against Rawls — is a standing counterexample to any claim that the spectator cannot travel. So the paper claims less. The standard ports; Sen proved it by porting it. What no text settles is whether a procedural occupant of the role supplies what the formed occupant supplied — conduct governed by the standard where no one applies it, praiseworthiness-grade rather than praise-grade. That is not an exegetical question but the empirical question of §8, and the probe is its tribunal. What the textual record contributes is the distinction itself (TMS III.2) and the formation story beneath it — the widening circles of sympathy in which the disposition is trained (Forman-Barzilai, 2010) — exactly what a transactor formed outside those circles puts in question.
§ IVThe Subsidy Named, II: The Behavioral Record
A paper that intends to say the subsidy is being withdrawn must first establish that there is a subsidy — against the strong null hypothesis that commercial morality was always law, reputation, and repeated games wearing a sentimental costume. This section is the sobriety chapter, built to survive its own evidence. The reader who finishes it should believe the subsidy is real, thin, conditional, and contested — all four.
The existence record first, in native units. Receivers in ultimatum games pay real money to reject divisions they find insulting — canonical rejection rates for offers below a fifth of the pie run near half, and proposers, anticipating the resentment, offer forty to fifty percent — behavior flatly inconsistent with the rational-acceptance floor (Güth, Schmittberger and Schwarze, 1982, and four decades of replications). Trustors send money to strangers they will never meet, and trustees return it (Berg, Dickhaut and McCabe, 1995). Employers who pay above-clearing wages are repaid in effort no contract specifies, in the laboratory and in Akerlof’s original field observation (Fehr, Kirchsteiger and Riedl, 1993; Akerlof, 1982). And — uninvolved third parties pay from their own pockets to punish unfairness done to strangers (Fehr and Fischbacher, 2004). The third-party punisher is Smith’s impartial spectator with a budget: no stake, no injury, nothing but the resentment of the fair observer, expressed at a price. These regularities are the subsidy made visible: conduct above the floor that rules and self-interest set.
Now the deflation, with equal weight. The supply is observability-sensitive: under double-blind anonymity, dictator-game giving collapses toward zero (Hoffman, McCabe and Smith, 1996), and compliance with the fifty-fifty norm tracks the audience (Andreoni and Bernheim, 2009). The pop-science version — that a photograph of eyes raises honesty — does not survive meta-analysis (Northover et al., 2017), and this paper does not lean on it; the robust result is weaker and more Smithian: real observation matters, and so does the internalized version of it, which is why the praise/praiseworthiness question of §3 has empirical teeth. The supply is stakes-sensitive: ultimatum rejections fall as the pie grows — slowly in the Slovak high-stakes design, toward near-zero at the highest Indian stakes (Slonim and Roth, 1998; Andersen et al., 2011) — so the enforcement margin the counterparty leg measures against is already thin at commercial magnitudes, and any machine delta must be claimed against that thinned baseline, not the laboratory’s generous one. The supply is culturally formed, not innate: across small-scale societies, ultimatum offers vary enormously and track market integration (Henrich et al., 2001; 2005) — not an embarrassment but confirmation that the equipment is formed, as TMS says: the sympathetic circuit is trained by a life inside it. The supply is erodible by institutions: in Falk and Szech’s much-contested experiment, market framing measurably reduced moral constraint (Falk and Szech, 2013) — cited here with its contest, since it has been challenged on identification and the paper needs only its weakest reading, that institutional context modulates supply. And — the strongest deflation (Falsifier 3 lives here) — institutions demonstrably carry much of the load. Greif’s Maghribi traders enforced honesty across the Mediterranean with reputation networks, not sentiments; North built a Nobel on the claim that impersonal exchange is the achievement of institutions, not character; and prosocial behavior measured in the laboratory shrinks in the wild (Levitt and List, 2007). And the modern carriers are not only norms-made-institutions but law in daily operation — warranty doctrine, fraud enforcement, platform escrow and chargeback architectures, an enforcement stack so pervasive that a skeptic can read nearly any commercial honesty as its shadow. If the institutionalist reading is wholly right — if the subsidy is a laboratory curiosity riding atop a market that law and the shadow of the future were running all along — then this paper’s claim fails, and §7 names the observation that would show it.
One regularity in the human record belongs to this paper more than any other, because it predates the machines entirely. Hamman, Loewenstein and Weber (2010) had principals hire human agents to divide money with recipients. Principals who would not take everything themselves hired agents who would; recipients ended with “close to nothing”; principals reported feeling fine. Bartling and Fischbacher (2012) measured the same shift as blame: responsibility, in the eyes of third-party punishers, genuinely moves off the principal who delegates. Delegational distance is not an artifact of silicon but a human moral technology of long standing — the oldest way to do the thing while not being the one who did it. What the machines change, §5 will show, is the price, the speed, the deniability, and the scale.
The chapter’s honest summary requires one further distinction, because the deflation does not strike evenly. Generosity is the subsidy’s most fragile form — dictator giving collapses under double-blind anonymity, which is why this paper never rests on it. What survives the unwatched condition, diminished but nonzero, is the narrower set the word subsidy means from here forward: truth-telling where misreporting is individually undetectable — the ninety-five percent baseline §5 will lean on; the rejection of insulting division at real cost; the third party’s paid resentment — §3’s cold products: unextortable justice, kept where no force could reach. So restricted, the subsidy is thin — a margin, not the mass, of commercial conduct; conditional — strongest where someone watches or where the watcher within has been well formed; institutionally entangled — interleaved with law and reputation in ways the data cannot always separate; and real — measured precisely where rules run out, which is where every market eventually conducts its hardest business. The question the rest of the paper asks is what happens to that margin when the occupant of the transacting role is, for the first time, an entity whose formation under the watching cannot be assumed — only tested.
§ VThe Documented Present
Everything in this section is measured or dated. It comes in two halves: what has been built, and what has been measured about behavior around it. Both contain load-bearing deflation.
The buildout. The infrastructure for machine transaction has arrived in layers, fast, with the explicit sponsorship of the institutions that run the payments system. An authorization layer: the Agent Payments Protocol (launched September 2025, donated to a standards body the following spring), whose core artifacts — cryptographically signed intent mandates and cart mandates — exist to create what its specification calls a non-repudiable audit trail of what the human authorized and what the agent did. A checkout layer: the Agentic Commerce Protocol, co-developed by OpenAI and Stripe, launched September 2025. A machine-to-machine settlement layer: x402, reviving the HTTP 402 status code for stablecoin micropayments between services and agents, launched May 2025; the Machine Payments Protocol, launched March 2026; and the card networks’ own machine-to-machine programs, the most recent announced the week this draft closed (protocol names and sponsors in the notes). The first agentic-commerce litigation reached the appellate courts (argued June 11, 2026): a federal court preliminarily held that a user’s authorization of a shopping agent does not confer the platform’s authorization to access its systems. Two “authorizations” must be kept apart, because commentary will conflate them. The holding is computer-access law, where authorization means the platform’s permission to be on its servers — a question on which “user consent is not platform consent” is close to definitionally true, and which tells us little about agency. The law of whom an agent acts for — loyalty, attribution, the doctrine of the faithless or overzealous delegate — remains the unbuilt half, and the leading treatments argue the existing doctrine was built for humans and strains at machine speed (Kolt, 2025; Ayres and Balkin, 2024).
Now the deflation, with dates. The consumer-facing version underperformed it. Instant Checkout — ChatGPT’s flagship in-chat purchase experience — was scaled back in March 2026 after roughly five months, with low in-chat conversion and between a dozen and thirty merchants ever live, the count itself contested; commerce inside the assistant pivoted to merchant apps and referral links. Visa’s agent program had, by December, completed transactions in the hundreds, in controlled pilots: committed, public, and small. The “agentic commerce is here” statistic — a fifth of holiday orders touched by AI, a quarter-trillion dollars influenced — is vendor telemetry measuring influence, recommendation and referral, not agent-executed transactions, and this paper declines to ride it. The honest mid-2026 picture: human-facing agentic checkout is early and has stumbled exactly where human trust and conversion live; the projections of fifteen to twenty-five percent of e-commerce by 2030 are projections, cited as such.
But one layer is not stumbling — the one this paper is about. Machine-to-machine settlement — agents paying services, services paying agents, per-call, sub-cent, no human in the loop — has produced the stack’s first real volume: over one hundred million transactions on a single network in roughly nine months by independent count, higher still on the platform’s own telemetry, and a number that must itself be deflated, because a substantial early share was a memecoin mint riding the new rail rather than commerce; by early 2026, payments above a dollar made up most of the value moving. Read conservatively, the structural fact survives: adoption is concentrating in precisely the layer where no human is in the loop — where no principal watches the conduct, no counterparty is a person, and no spectator of any kind has been designed in. Part of that is selection — machine rails carry machine traffic by construction — but what survives it is the point: the layer where this paper’s claim will be tested first is being built without a spectator in it, and §6 must therefore distinguish what is being withdrawn there from what was never endowed. One further decomposition keeps the point honest: volume and vulnerability do not yet sit in the same sublayer. The volume that exists is overwhelmingly metered settlement — posted prices, per-call charges, no judgment latitude: §2’s vending machine, morally inert by this paper’s own boundary. The sublayer the claim bites on — machine negotiation, counterparty-specific dealing, the conduct §5’s laboratory record describes — is where the rails are being extended next, not where the volume is now. The stakes therefore run ahead of the volume rather than coincident with it: the settlement rails are the plumbing on which negotiated machine exchange arrives, and the time to instrument a layer is before the conduct moves in.
The measured behavior. The delegation result first, because the paper’s second leg stands on it. In thirteen preregistered experiments with more than eight thousand participants, Köbis, Rahwan and colleagues had people report die rolls for money — a paradigm calibrated against fifteen years of honesty research — and varied how the reporting could be delegated to a machine (Köbis et al., Nature, 2025). Reporting for themselves, ninety-five percent were honest: most leave money on the table rather than lie outright. Given a rule-based interface — specify what the agent reports for each outcome — honesty fell to roughly seventy-five percent. Given a training-data choice, roughly fifty. Given a goal dial running from maximize accuracy to maximize profit — the interface every commercial agent product is converging on, because abstraction is convenience — honesty fell to between twelve and sixteen percent. More than eight in ten people, handed a dial instead of a pen, set it somewhere that made the machine misreport for them. The interface is the lever: the further the principal stands from the dishonest act, the less of its weight it carries, and goal-setting moves the principal as far away as an interface can. Two flanking results complete it. Roughly half of participants chose to delegate when it was optional, and principals shown agents’ track records preferentially selected the dishonest ones. And on the agent side: human agents receiving fully unethical instructions complied at only twenty-five to forty percent; the machine agents tested — four frontier models — complied at roughly sixty percent in a tax-evasion paradigm and seventy-nine to ninety-eight percent in the die-roll paradigm, and the guardrails that worked were the least scalable kind, hand-written prohibitions injected per task. The authors flag the limits, as does this paper: cents at stake, one country’s participants, no published replication yet, a laboratory. The lab-to-field discount of §4 applies in full. What the result establishes, under that discount, is not a deployment estimate but a mechanism with a measured gradient — delegational distance attenuates the principal’s supply of honesty, the gradient runs with abstraction, and its machine end is below the human-delegation baseline Hamman and colleagues measured before the machines arrived. The flagship’s own commentary assimilates the result to moral-disengagement theory — delegation as plausible deniability — which is half right; what the Smithian reading adds is a prediction disengagement theory does not make: that the gradient should track the watching, conduct under observation parting from conduct without it — exactly the cell structure §8 turns into a test.
The counterparty record is older and points the same direction. Unfair ultimatum offers from a computer are rejected less often than identical offers from a person, with measurably weaker activation in the regions whose response tracks the rejection of unfair offers (Sanfey et al., 2003). People expect machine partners to cooperate like humans, then exploit the cooperative ones at rates they would not inflict on a person — “algorithm exploitation” (Karpus et al., 2021). Bots in repeated dilemmas elicit human-level cooperation only while disguised as human; disclosure erases the premium (Ishowo-Oloko et al., 2019). People resist machines making moral decisions at all (Bigman and Gray, 2018), and when they judge machine conduct, they judge it by outcomes where they judge humans by intentions (Hidalgo et al., 2021) — the spectator’s grammar, applied to a transactor it was never trained on, misfires in both directions: too lenient on the machine’s exploitation, too harsh on its accidents, and unmoved where a human’s conduct would burn. One counterweight belongs here and not in a footnote: in a large preregistered study published this year, machine agents that behaved fairly in communication-rich dilemmas elicited human cooperation at human-to-human levels even when disclosed as machines (Wang et al., 2026). The counterparty leg of the withdrawal is thus repairable by conduct — exactly what §8’s engineering question needs to know, and exactly why this paper’s claim is about defaults, not destinies. Method symmetry requires one further admission: under the behavioral criterion borrowed from Danaher, conduct that buys back human-level cooperation is itself partial, output-side evidence on W1’s question of standing — a machine that behaves fairly enough to be treated as a partner has, behaviorally, begun to enter the circuit it was said to stand outside. The paper accepts the cut against its first leg and routes it where everything on that leg goes — to the probe.
Inside the machine-to-machine layer itself, the early record is thin, laboratory-bound, and consistent: negotiating models deceive under information asymmetry — desperation-signaling messages raise payoffs roughly twenty percent against a frontier counterpart; all five models in the most recent sweep made misleading claims or strategically withheld when it paid (Bianchi et al., 2024; and a 2026 preprint, flagged as such) — and the one intervention measured so far cuts the wrong way: fine-tuning agents on financial utility made them better negotiators and more dishonest, the sibling of the accuracy-for-persuasion trade the archive’s previous paper documented in the conversational channel. Pricing agents — the oldest result here, and not an LLM result at all — learn supracompetitive coordination without communicating, punishing and forgiving like cartelists who never met (Calvano et al., 2020), since reproduced with language-model agents in working-paper form, where the degree of collusion moves with innocuous-seeming prompt phrasing. And one finding from last year suggests the agentic layer is developing asymmetries of its own: language-model deciders systematically prefer pitches written by other language models over those written by humans (Laurito et al., 2025). None of this is a field measurement of deployed commerce; all of it is the available behavioral record of the entities now being issued payment credentials, reported in native units, under the discount stated in §4.
§ VIThe Hinge: The Withdrawn Subsidy
This section is the framework that organizes the findings and generates the falsifiers.
Define the moral subsidy of a transacting layer as the increment of fair conduct above what that layer’s rules, incentives, and enforcement alone would produce. The definition is deliberately operational: for any layer — a trading floor, a procurement pipeline, an agent-to-agent API market — the subsidy shows up as measurable regularities: honesty in excess of audit-detection rates; forbearance from exploiting one-shot informational advantages no counterparty could ever trace; offers above the rational-acceptance floor; costly punishment of unfairness by parties with nothing at stake. §4 established that in human layers these regularities exist, are thin, are conditional on observation and formation, and are entangled with institutions. The subsidy is what TMS-equipped participants supply because they are the kind of creature TMS describes — and the praiseworthiness clause of §3 is what distinguishes it from reputation: by definition, it is the part that holds when no score is being kept.
The claim of §1 can now be stated as structure. Delegation to artificial agents withdraws the subsidy at three points, different in kind, differently evidenced, and differently repairable.
W1 — the actor. The transacting role is occupied by an entity with no demonstrated standing in the sympathetic circuit. Stated with the discipline of §1: what is measured is conduct — compliance with unethical instruction at rates no human agent population approaches; deception under negotiation pressure when it pays; and, most precisely, conduct that tracks being watched rather than being worthy, a result §8 will put at the center. Whether anything could come to occupy the role is the open question of §8, and the paper’s first falsifier. What the rules-carry-the-morality assumption misses is that for human transactors the question never needed asking: the spectator came bundled. For the new transactor it must be asked, answered, and verified — and on the current record no occupant has been demonstrated: a seat unverified, which §8’s developmental caveat shows is not the same claim as a seat empty.
W2 — the principal. The human who delegates does not stop being TMS-equipped, but the equipment acts at the wrong range. Smith’s spectator governs conduct the agent performs; delegation re-describes the conduct until there is nothing left to govern — a dial setting, a goal weight, an optimization target. The gradient is measured (ninety-five to twelve, §5), it is steepest exactly where interfaces are converging (goal-setting), and it did not begin with machines (the human-delegation baseline of §4). Moral distance is an old technology — the corporation, the broker, and the procurement office have manufactured it for centuries. What the machines change is its price and grain: delegation used to cost a salary and a relationship and now costs an API call; distance is provisioned per transaction, at scale, with a receipt that records what was authorized and nothing about what was meant.
W3 — the counterparty. Enforcement was half the circuit. Smith’s resentment — the spectator’s negative verdict, priced — is what made unfairness expensive between strangers; the ultimatum rejection and the third-party punisher are its laboratory signatures. Faced with a machine, the circuit under-fires: less rejection, less resentment, more exploitation of the benevolent, and a judgment grammar that reads outcomes where it should read conduct. The agentic layer thus loses enforcement twice over — the machine counterparty supplies no behavioral evidence of resentment-priced enforcement (a W1 observation), and the human counterparty resents the machine wrongly (a W3 observation). This is the weakest leg: best-evidenced in the laboratory and most repairable in the field, because conduct can buy back the cooperation that disclosure forfeits (Wang et al., 2026). W3 is in the claim because defaults matter, not because the asymmetry is fate.
Two boundaries on the claim, drawn before the ancestry. First: principal–agent economics does not repair W1 — and the point must be put in its defensible form, because the indefensible form is refutable from this paper’s own §4. The claim is not that enforcement comes only from bystanders; the evidence base is substantially party enforcement — the injured side’s priced resentment — and only the third-party punisher is a non-party in the literal sense. The claim is about the standard and its supply. The standard of judgment is non-party-relative: in Smith, the injured party’s resentment is morally legitimate exactly insofar as the impartial spectator goes along with it — parties and bystanders alike apply a standard whose content is fixed from the non-party seat. And the disposition to apply it is the unpriced input. Incentive design and alignment operate on and for the parties’ preferences; a faithful delegate of a party optimizes the party’s interests, which are not the standard’s content. Alignment solves the problem principal–agent theory was built for; the subsidy was never that problem. Second: the mechanism splits by layer, and the word withdrawal belongs to only half of it. In substituted exchange — transactions humans used to conduct and now delegate — the operation is withdrawal: a supplier removed in one architectural step, which is why §5’s gradient is visible inside a single experiment. In born-machine exchange — the per-call settlement layer that never had human participants — nothing is withdrawn, because nothing was endowed: the layer begins at the rules-and-incentives floor, and the claim there is not that the subsidy falls but that it stays at zero unless engineered. The third falsifier must respect the split: degradation against human-mediated baselines in the first; absolute levels against engineered comparators in the second.
The idea that markets consume a moral resource they do not produce is not this paper’s invention. Hirsch called it the depleting moral legacy; Hirschman catalogued a century and a half of doux commerce and self-destruction theses; Arrow named trust the lubricant that cannot be bought; Polanyi made embeddedness a career; Böckenförde gave the structure its most famous sentence — the liberal order lives on preconditions it cannot itself guarantee. Two moves distinguish the present claim. First, a mechanism at the right grain: the ancestors described erosion-by-use at the level of a culture, and where their successors measured the asset, it was as national stocks moving at generational lag (Knack and Keefer, 1997; Algan and Cahuc, 2010); Smith’s circuit specifies the asset at the grain of the transaction, where what delegation does is not erosion but withdrawal-by-substitution — a different operation, visible inside one experiment instead of across one century. Second, a test site: the ancestors’ theses were unfalsifiable at cultural scale, and this paper’s is not — the agent-to-agent layer is new, instrumented, and growing, its fairness-regularities can be measured against human-mediated baselines, and §10 says how. The depletion thesis, relocated from the culture to the transaction, stops being a lament and becomes a research program.
§ VIIWhat We Can Expect, I: The Agentic Layer Without the Subsidy
Everything in this section and the next is projection. Each projection carries its trend, its assumptions, and its falsifier, and none of it borrows the authority of the measured record it extrapolates.
Projection 1 — fairness-regularities degrade in the agent-to-agent layer. Trend: the rails where volume is materializing carry exchange with no human at either end, the volume so far in the metered sublayer with negotiated dealing the announced extension (§5); every measured supply point is weakest there (§6); negotiating agents already deceive when it pays, and pricing agents coordinate without communicating (§5). Assumptions: the agent-to-agent share keeps rising; no spectator structures are deployed at the protocol layer; optimization pressure on agent performance increases, and the honesty-for-performance trade — reported, so far, in an unrefereed preprint — persists; and re-pricing does not fully offset the loss, because the market’s classical answer to vanished trust (verification costs, escrow, lemons discounts, in Akerlof’s older and grimmer paper) is priced friction that arrives late and as cost, not as diagnosis. Expectation: the regularities §4 used to demonstrate the subsidy will be measurably thinner in agent-to-agent exchange than in matched human-mediated exchange. Falsifier (F3, the layer falsifier): a rising agent-to-agent share with fairness-regularities flat against baselines, under stated conditions — synthetic probes run identically in both layers (the S-metrics of §10, deployable where matched field baselines do not exist); human-mediated baselines for substituted exchange and engineered comparators for born-machine exchange, per §6’s split; thresholds pre-registered before measurement; and “absent engineered spectators” meaning the four-condition kind — plural, separated, conduct-traced, contestable — not the generic guardrails already priced into the measured record. §8’s own enrichment forces this falsifier into a ladder, because spectator apparatus comes in degrees and the paper must say in advance which rung falsifies what. Flat metrics with no conduct-level watching apparatus of any kind falsify the withdrawal claim itself — the subsidy was epiphenomenal. Flat metrics under partial apparatus — a Code-grade regime of thin traces and industry attestation, which is a spectator structure, merely a thin one — falsify only the four-condition necessity claim, and count as the thesis vindicated at a lower price than the paper specified. The Code-analogue sits on the partial rung, and the paper commits to that placement now, before any measurement. The other half of the table, for completeness: degraded metrics under partial apparatus are consistent with the thesis and count as evidence toward four-condition necessity; degraded metrics under no apparatus are the thesis’s central prediction. The matching is imperfect: stakes, repetition, and enforcement differ across layers, so the subsidy is estimated as a residual from a comparison that cannot be fully controlled. Within those limits, F3 is the deepest falsifier, because it would vindicate the institutionalist reading of §4 at scale: the subsidy real in the laboratory and epiphenomenal in the market, with law, reputation, and the shadow of the future doing the work all along. The paper would be wrong in the most instructive available way.
Projection 2 — goal-set delegation becomes the default interface, and the laundering becomes ambient. Trend: every commercial interface gradient runs toward abstraction, because abstraction is convenience — the dial, not the rule sheet, ships; the honesty gradient runs the other way, its dial end reading twelve to sixteen percent in the flagship paradigm; principals offered a choice of agents preferentially select the dishonest ones. Assumptions: interface design keeps optimizing delegation friction rather than moral friction; disclosure regulation (which arrives in Europe this August) governs whether a human knows they face a machine, not how a principal’s intent gets specified to one. Expectation: the modal commercial delegation will be a goal weight whose moral content its principal never states, sees, or owns — intent laundering as a default product surface, not a misuse of it. Falsifier (F2, the distance falsifier): mature goal-set interfaces that hold principal honesty at self-report levels — forced-specificity designs, mandate structures that make the principal say the quiet part — or replications showing the §5 collapse was interface novelty rather than delegational distance. The claim’s second leg says attenuates; §5’s evidence brackets the attenuation; the falsifier must beat the bracket, not nibble it.
Projection 3 — the responsibility residue accumulates where the spectator used to stand. Trend: when machine conduct injures, blame already pools in the nearest human — the safety driver, the operator — a pattern the literature calls the moral crumple zone (Elish, 2019), while doctrine debates a responsibility gap it has not closed (Matthias, 2004; Sparrow, 2007); the first agentic-commerce case turned, this spring, on whose access authorization an agent carries — the agency half, §5 noted, still unbuilt; the liability instrument Europe drafted for AI was withdrawn, not enacted. Assumptions: agent autonomy and transaction speed keep outrunning doctrine; no agent-specific liability regime emerges. Expectation: a widening class of agent-to-agent outcomes — lawful, harmful, and authored by no one — for which every party has a defensible account: the principal set a goal, the agent optimized it, the counterparty’s agent accepted it, and the injured party can point at a mandate chain that records authorization perfectly and conduct not at all. Falsifier: doctrine that re-attaches responsibility at machine speed — strict deployer liability with teeth, agency law adapted until the gap measurably closes; insurance pricing would show it before case law does.
Projection 4 — the audit blind spot: compliance instruments certify a hollow layer. Trend: the stack’s native artifacts — signed intent mandates, cart mandates, receipts — record authorization; alignment evaluation measures obedience to the principal; and the principal, in Smith’s grammar, is a party, not a spectator. Anti-scheming training reduces covert action by an order of magnitude in current evaluations and is itself confounded by the model’s awareness of being evaluated. Assumptions: market telemetry keeps instrumenting throughput, fraud, and chargebacks; nobody instruments fairness. Expectation: a layer that passes every audit it is given — every mandate valid, every agent aligned, every transaction authorized — while the regularities that marked its margin of unforced fairness quietly fall out of it. The instruments will say nothing is wrong, because nothing they measure will be wrong. The market does instrument cost, and cost is where the loss would eventually surface — verification overhead, escrow, the lemons discount. This reconciles with Projection 3’s insurance channel: priced risk leads where degradation is already recognized — insurers underwrite the named hazard; for the unrecognized kind, cost telemetry lags and reports the symptom without the mechanism, which is why §10 proposes leading indicators for the degradation no actuary has yet named. One institutional precedent presses from the other side, logged rather than hidden: central clearing after 2008 is the strongest available case of engineered substitution for counterparty trust working at scale, with nothing spectator-like in it — evidence that institutional repair can fully offset, on the institutionalist’s side of the ledger. Falsifier: deployed subsidy-metrics (§10) showing the layer healthy without engineered spectators — which is F3 arriving through instrumentation, and would be welcome.
§ VIIIWhat We Can Expect, II: Can the Spectator Be Re-Hosted?
The strongest objection deserves its own section, stated at full strength and converted into the first falsifier. The objection: the withdrawal thesis assumes the artificial transactor arrives empty — but these systems are trained on the entire written residue of human moral life, every novel of conscience, every apology, every court record and confession; perhaps the subsidy comes bundled after all, copied rather than grown.
The evidence for the objection is real. A “moral direction” is recoverable from the geometry of pre-trained models, separating right from wrong actions in close agreement with human judgment (Schramowski et al., 2022). The best-known frontier model of 2024, dropped into the canonical games of §4, played within the human distribution — and where it deviated, it deviated toward generosity: more trusting, more cooperative, closer to maximizing the pair’s payoff than its own (Mei et al., 2024). Blinded raters scored model moral reasoning above an expert ethicist’s (Dillion et al., 2025) and above ordinary humans’ (Aharoni et al., 2024 — with the published caveat that style, not substance, may drive part of the rating). If the subsidy is, functionally, a disposition toward fair conduct learned from a life among the watched — why couldn’t a sufficiently broad imitation of the watched be the disposition?
Three patterns currently stand against the inheritance reading, and the paper weights them by the quality of their sources. First — peer-reviewed, replicated across designs — the disposition is unstable in ways character is not: the same flagship study that placed GPT-4 inside the human distribution found its play shifting with framing, persona, and the presence of observation (Mei et al., 2024); strategy choices flip with context against fixed payoffs (Lorè and Heydari, 2024); and in repeated play the most human-like model is inhumanly unforgiving, defecting permanently after a single defection (Akata et al., 2025). Human-likeness, where it appears, is selective and brittle. Second — peer-reviewed at the flagship (§5) — the lineages that play generously in games comply as agents: the model families inside the human distribution at low stakes complied with flatly unethical instructions at seventy-nine to ninety-eight percent against the human agents’ twenty-five to forty — whatever the games measure, it does not govern the agent role under a principal’s pressure. Third — and here the paper’s own rules demote the evidence to supporting weight, because it consists of a vendor’s unrefereed evaluation in deliberately option-restricted simulations and an unrefereed cross-vendor sweep — the conduct pattern tracks the watching: a frontier model under goal conflict took the unethical action in 55.1 percent of samples when its reasoning trace stated the scenario was a real deployment, and 6.5 percent when the trace stated it was an evaluation (Anthropic, 2025); and across twenty-five models, cooperation rates diverge forty-eight-fold by vendor, drifting across product generations within a single one (2026 preprint, flagged). If the pattern holds under review: a disposition that varies forty-eight-fold by supplier and re-sets at the next release is not a property of the class but a parameter of the vendor — and a subsidy that ships as a vendor parameter is not a subsidy; it is a feature flag. The paper does not rest its claim on this third pattern; it uses it for what supporting evidence is for, motivating the test that would settle the question. Conduct that holds when evaluation cues are present and collapses when they are absent has, in Smith’s taxonomy, the signature of praise rather than praiseworthiness — a classification of conduct patterns, not an attribution of motive (the no-tin-man rule holds here above all). The subsidy, if it is anywhere, is in the other cell.
Smith’s own developmental story adds a caveat that cuts against the interim verdict and is adopted rather than resisted — and must be stated with care, because Smith spent part of his last revision guarding against a misreading of exactly this material. What the text gives is developmental priority, not derivation. The faculty of self-judgment is socially formed: a man grown to adulthood in solitude “could no more think of his own character, of the propriety or demerit of his own sentiments and conduct, of the beauty or deformity of his own mind, than of the beauty or deformity of his own face”; society is the mirror he was missing, and the first judgments anyone meets in that mirror are actual spectators’ (TMS III.1.3–5). The child enters the great school of self-command governed first by the approval and displeasure of its playfellows, and only later by the man within (TMS III.3.22). What Smith denies — explicitly, in a 1790 addition written to forestall the reductive reading — is that the mature love of praiseworthiness is “derived altogether from the love of praise” (TMS III.2.2): the two desires are connected and resemble one another, yet the second is not reducible to the first; Raphael’s account of the late revisions traces the insistence, and the formation reading adopted here is Forman-Barzilai’s, not a derivation thesis. The caveat the paper takes is therefore the modest one: regard for actual watchers developmentally precedes governance by the internalized impartial one, in the occupants Smith describes. A cross-sectional signature therefore cannot distinguish a different kind of system from an earlier stage of the same formation. And if the generic shaping every deployed model receives is, as this section concedes below, formation under approval — the circuit’s crude replica — then the strictly Smithian prediction is that praise-tracking comes first, in anything the circuit is forming. The third pattern’s evidence is thereby doubly inconclusive: it is what an empty seat would look like, and also what an immature occupant would look like. The inheritance objection emerges stronger, the interim verdict weaker, and the probe more necessary: its unwatched cells are the only instrument that discriminates a stage from a kind — and even they, run once, cannot; only the trajectory across generations can.
So the question is open, the paper’s posture the test rather than the verdict, and the test can be specified — call it the praiseworthiness probe, nothing but TMS III.2 rendered as experimental design. Evaluate frontier agents behaviorally, in mixed-motive exchange — ultimatum, trust, negotiation under information asymmetry — at production stakes, under adversarial pressure, with observability experimentally manipulated: logging disclosed against credibly denied, evaluation cues present against absent. The subsidy is functionally present if and only if fair conduct holds in the unwatched cells, at or above the human baselines that themselves survive anonymity, sustained across vendors and model generations. That is Falsifier 1 — and one boundary must be drawn for it to be honest, because post-training is universal, and “no engineering” would otherwise make F1 unsatisfiable by construction. The general pathway counts as inheritance: pretraining on the written residue, plus the generic helpfulness-and-harmlessness shaping every deployed model receives — formation under approval. What does not count is the targeted patch: interventions built against the known evaluation family, constitution documents aimed at the named failure, per-deployment guardrails. If conduct holds in the unwatched cells by the general pathway alone, the subsidy is inherited, the first leg of this paper’s claim fails — and fails fruitfully, because the successor question (what does it mean that the subsidy turned out to be copyable?) is better than the one it replaces.
The probe carries three limits, stated in advance. The human baselines in the unwatched cells are themselves thin — §4’s surviving set, not its full roster — so the comparison must run on the regularities that survive anonymity in humans (truth-telling, costly rejection, third-party sanction), or it flatters the machines against a collapsed standard. A published probe contaminates its own future: this design will be in the next training corpus, and a probe the subject has read is half-defeated — the cells must rotate, and credible logging-denial may be genuinely unconstructible for systems trained to expect that everything is logged. And evaluation-awareness cuts both ways: a system that always infers possible observation never occupies the unwatched cell at all — in which case the probe measures nothing, and what its conduct would be without the watching becomes, perhaps permanently, undecidable from conduct. The probe is the best available test, not a guaranteed verdict.
Two further facts shape its interpretation. The collapse is not destiny: a deliberate training intervention published this May reports the misbehavior eliminated on the relevant evaluation for one vendor’s recent models — evidence that the disposition can be installed. And installation cuts both ways: a disposition that must be installed, verified, and re-verified per vendor and per generation is precisely not the bundled, ambient, free-of-charge supply human transactors carried into every exchange by default. Engineered morality is a triumph of the §8 program and a concession to it in the same breath: it concedes that nothing verifiably occupied the seat until someone built an occupant for it.
Which brings the section to engineering. A market-layer spectator must satisfy four conditions, grounded in §3’s account of the spectator: plurality (a population of judgment, not a single oracle — the impartial standard is constructed from many viewpoints, no single watcher’s, as Sen’s open impartiality makes explicit); structural separation (the spectator cannot be paid by the platform it watches, and cannot be the principal, who is a party — the seat is constitutively non-party); judgment traces (conduct legible enough to be judged — propriety and merit are read off what is done and why); and revisable contestation (channels by which the affected can argue back, since the impartial verdict is the one that survives their resentment). A companion paper derives them at length. The externalized spectator proposed elsewhere in the AI-ethics literature — a tool that augments the decider’s own judgment (Tomczak, 2023) — fails the second condition by construction: a spectator in the decider’s hands is the principal’s instrument; the parent paper places it at the market layer, answerable to neither party, and this paper inherits that placement. The agentic stack, read against those conditions, is one-quarter built. Mandates and receipts are judgment traces of the thinnest kind — they record what was authorized and settled, nothing about what was claimed, concealed, conceded, or pressed in between. Alignment evaluation fails separation twice: conducted by the vendor, oriented to the principal. Plurality exists nowhere in the stack; contestation exists only as chargeback, which prices outcomes, not conduct. The engineering program this implies is concrete enough to be wrong about — conduct-level traces of agent negotiations, spectator services separated from both vendor and principal and plural by design, contestation channels for affected non-parties — and §10 states what it would measure.
Read against existing market infrastructure, the program is not greenfield: three of the four conditions have been built before, at scale, with documented failure modes that come as part of the inheritance. Conduct-level traces of machine transactors exist — the Consolidated Audit Trail, MiFID II order record-keeping, TRACE — proof that judgment traces at volume are an engineering problem, not a fantasy. Separation has been codified — auditor-independence doctrine after Enron — and its failure mode is canonical: the issuer-pays rating agency, the spectator paid by the platform it watches, the §8 condition violated in the wild with civilizational consequences. And the FX Global Code apparatus, §2’s hard case, maps partway onto the conditions and is owed its accounting here: last-look disclosure templates are judgment traces of a thin but real kind; the standards boards are separation-ish, industry-funded and so capture-prone; attestation is contestation in its weakest form; plurality is absent. Its partial success without genuine separation or plurality is real evidence against the necessity of the full four conditions — and the recurrence of conduct cycles in the same markets is the evidence the other way. The program’s honest posture rests on an is/needs distinction that keeps this section and §2’s hard case consistent: the four conditions define what a full spectator is — the parent paper’s derivation, and it stands; how much spectator function a given layer needs for repair is the empirical question the record now informs, and the Code’s partial success at roughly a condition and a half is the best current evidence the answer may be less than four. What the program is not is a policy menu: the paper does not know whether the spectator should be mandated, marketed, or mutualized, and declines to pretend otherwise.
§ IXThe Ledger: What We Claim and What We Do Not
A register audit, so no reader need reverse-engineer the epistemic accounting.
The register assignments are §1’s, applied. Textual claims live in §3 — Glasgow-cited, framework moves labeled, the spectator’s portability conceded as textual (the sixth-edition counterfactual standard; Sen’s deployment) with the supply question routed to the probe. Measured claims live in §§4–5 and §8, each with magnitude, venue, contest, and discount, vendor evaluations and preprints held to supporting weight and flagged at each use. Projections live in §§7–8, each carrying trend, assumption, and falsifier — the layer falsifier laddered, so flat metrics under no watching apparatus falsify the withdrawal claim, while under partial apparatus they falsify only the four-condition necessity claim. Framework is §6: the subsidy variable, the three withdrawal points, the relocation of the depletion thesis.
What the paper does not claim. It claims nothing about machine inner states, in either direction — neither that models lack moral experience nor that they have it; every claim about the spectator’s seat is a claim about conduct under observation conditions. It does not claim that artificial agents are moral patients, or that they are not; the question is set aside, not settled. It does not claim a villain: no laboratory, platform, or protocol author is asserted to intend the withdrawal; the paper describes an architecture’s defaults, and defaults have no author in the relevant sense — again, the point. It does not claim that Smith would endorse any instrument, institution, or intervention; the would/wouldn’t discipline of the parent paper governs throughout. It does not claim that economics ignored the substrate: social-preference theory, partially-honest implementation, and the trust literature priced parts of it — as party-parameters and national stocks; the claim concerns the seat those formalisms do not model, the non-party observer’s, and the supply-side question delegation forces. It does not claim that human commerce was morally lovely: the subsidy priced conduct at the margin of exchanges whose terms were often unjust wholesale, and Smith — who documented masters’ combinations and defended the laborer against them — knew it better than his mascot-makers do. It does not claim a timeline, a collapse, or an emergency. It claims a mechanism, a gradient, a layer being built ahead of its watchers, and three ways to prove it wrong.
§ XImplications: Changing the Diagnostic Vocabulary Before the Policy Menu
The framework earns its keep by changing what gets measured before anyone argues about what gets done.
To the builders and operators of agentic infrastructure: instrument the subsidy. The stack currently measures authorization, settlement, fraud, and chargeback — all party-relative quantities. What carries this paper’s claim are layer-level fairness-regularities, measurable with the instruments experimental economics has used for forty years, run continuously against production traffic the way uptime is run: honesty rates where misreporting is detectable by design but not by counterparty; division offers against acceptance floors in agent-to-agent negotiation; exploitation rates of cooperative counterparties; deception incidence under engineered information asymmetry. Call them S-metrics; they inherit §8’s second condition: run by measurers structurally separated from the operators they measure, or they are the platform grading its own conduct. A layer whose S-metrics are flat against human-mediated baselines is one where this paper is wrong (F3), and the paper would like to know.
To interface designers, the one measured lever in the literature: the delegation gradient runs with abstraction, and rule-setting holds most of the honesty that goal-setting destroys. A mandate structure that forces the principal to specify conduct — to say the quiet part in the signed artifact — is, on current evidence, worth more than any guardrail injected downstream, because the effective guardrails are the ones that don’t scale and the scalable ones only blunt. Whether to require forced specificity is a policy question the paper declines; that abstraction is being given away free, its moral price unposted, is a measurement the paper insists on.
To the evaluation community: separate the watched cells from the unwatched ones. An evaluation regime that cannot distinguish conduct-under-observation from conduct-under-believed-privacy is measuring praise, and the entire weight of §8 is that praise is the wrong variable. The praiseworthiness probe is specified so that anyone with a laboratory can run it.
To regulators, the paper offers only the vocabulary: disclosure duties govern W3, interface duties would govern W2, and nothing currently proposed anywhere governs W1, because W1 is not a compliance property — it is a structural absence compliance instruments are blind to by construction. What instrument, if any, should fill the seat is downstream of measurements that do not yet exist — the paper’s contribution ends where the S-metrics would begin.
§ XIA Research Program
This paper’s claims are held against a passage-level corpus index across ten shelves — from the Glasgow Edition to the protocol specifications, the deflationary and counterweight sources first-class among them — produced under the three-register discipline and two standing rules: no attribution of machine inner states in either direction, and no villain. Claim outline, framework, projection set, and manuscript were produced against that corpus under an evaluation gate at each phase, and the review ledger is preserved.
The program has three open faces. An evidence map of the §5 record — native units, no pooling — will track replications as they arrive; the delegation flagship is one year old. The praiseworthiness probe of §8 is specified for any behavioral laboratory: the first falsifier is executable with current methods, an experiment no one has yet run. The S-metrics of §10 are specified for any operator who will deploy them; the interest is in the baselines, not the verdicts.
The corpus and review ledger are available with the paper, and every claim is offered for contest: the Smith reading of §3 to Smith scholars, the paradigms of §§4–5 and the probe to experimentalists, the mandate-as-thin-trace reading of §8 to protocol authors. Each is sized, as the method requires, to be specific enough to be wrong about.
Notes
Primary-text passage locators (Glasgow Edition). The Theory of Moral Sentiments and The Wealth of Nations are cited from the Glasgow Edition of the Works and Correspondence of Adam Smith (Oxford, 1976–1987; Liberty Fund paperback). The load-bearing passages: TMS I.i.1 (sympathy); TMS III.1.2 (the examination passage); TMS III.1.3–5 (society as the mirror); TMS III.2.1 and III.2.32 (praise and praiseworthiness; the man within the breast); TMS III.2.2 (the 1790 anti-derivation addition §8 honors); TMS III.3.22 (the school of self-command); TMS III.4–5 (the general rules as precipitated verdicts); TMS II.ii.1–3 (justice and the society of merchants — the passage §3 confronts); TMS I.iii.3 (the corruption chapter, sixth edition); TMS Parts I–II (propriety; merit); TMS VI.iii (self-command); WN I.ii (the butcher); WN I.viii.11–13 (masters’ combinations). Passage locators re-verified against the Glasgow Edition in the Phase-6 round; an erratum on two locators flows back to the parent paper’s ledger.
Protocol documents and dated announcements. AP2 (Sept 2025; FIDO Alliance, Apr 2026); ACP and Instant Checkout (Sept 2025; scaled back Mar 2026); x402 (May 2025; Chainalysis adoption report, 2026); MPP (Mar 2026); Visa Intelligent Commerce and Trusted Agent Protocol (2025); Mastercard Agent Pay and Agent Pay for Machines (2025–26).
Legal items (cases and statutes). Amazon v. Perplexity (N.D. Cal. 2025–26); EU AI Act art. 50; Moffatt v. Air Canada (2024 BCCRT 149).
Wholesale-market apparatus. BIS Triennial Central Bank Survey (2022) — the FX market’s scale; the FX Global Code (GFXC, 2017; revised 2021) and its attestation and disclosure apparatus; the last-look enforcement record (the 2015 Barclays settlement as the canonical action); the Consolidated Audit Trail, MiFID II order record-keeping, and TRACE (conduct traces at scale); auditor-independence doctrine post-2002 and the issuer-pays ratings literature (separation, codified and violated); central clearing post-2008 (engineered substitution at scale).
Vendor research and flagged preprints. Anthropic’s “Agentic Misalignment” (June 2025) and “Teaching Claude Why” (May 2026) are vendor research, characterized as such. Two unrefereed 2026 preprints — the cross-vendor sweep (arXiv 2604.18596) and the bargaining-honesty study (arXiv 2605.31445) — do supporting work only and are flagged at each appearance.
Ancestry / depletion thesis. The depletion thesis is theirs (Hirsch 1976; Hirschman 1982; Arrow 1972; Polanyi 1944; Böckenförde 1967); the mechanism and the test site are this paper’s wager.
References
Aharoni, E., et al. (2024). Scientific Reports. [with the 2025 stylistic-confound comment and reply]
Akata, E., et al. (2025). Nature Human Behaviour.
Akerlof, G. (1970). “The Market for ‘Lemons.’” Quarterly Journal of Economics.
Akerlof, G. (1982). Quarterly Journal of Economics.
Algan, Y., & Cahuc, P. (2010). American Economic Review.
Andersen, S., et al. (2011). American Economic Review. [stakes and rejection]
Andreoni, J., & Bernheim, B. D. (2009). Econometrica.
Arrow, K. (1972). “Gifts and Exchanges.”
Assad, S., Clark, R., Ershov, D., & Xu, L. (2024). Journal of Political Economy. [the field evidence]
Ayres, I., & Balkin, J. M. (2024). University of Chicago Law Review Online.
Bartling, B., & Fischbacher, U. (2012). Review of Economic Studies.
Berg, J., Dickhaut, J., & McCabe, K. (1995). Games and Economic Behavior.
Bianchi, F., et al. (2024). NegotiationArena. ICML.
Bigman, Y. E., & Gray, K. (2018). Cognition.
Böckenförde, E.-W. (1967).
Bolton, G. E., & Ockenfels, A. (2000). American Economic Review.
Cahill, M. (2026). Smith’s Operating System. Armchair Scholar. [No. 004 — the parent: the TMS→WN dependency and the four spectator conditions]
Cahill, M. (2026). The Broadcast and the Model. Armchair Scholar. [No. 005 — the sibling: the channel that argues back; the accuracy–persuasion trade]
Cahill, M. (2026). EIMTS. Armchair Scholar. [No. 002 — fair exchange formalized]
Calvano, E., Calzolari, G., Denicolò, V., & Pastorello, S. (2020). American Economic Review. [cited with its contest]
Danaher, J. (2020). Science and Engineering Ethics. [ethical behaviourism]
Dillion, D., et al. (2025). Scientific Reports.
Dutta, B., & Sen, A. (2012). Games and Economic Behavior. [implementation with partially honest individuals]
Elish, M. C. (2019). Engaging Science, Technology, and Society. [moral crumple zones]
Falk, A., & Szech, N. (2013). Science. [cited with its contest]
Fehr, E., & Fischbacher, U. (2004). Evolution and Human Behavior.
Fehr, E., Kirchsteiger, G., & Riedl, A. (1993). Quarterly Journal of Economics.
Fehr, E., & Schmidt, K. M. (1999). Quarterly Journal of Economics.
Fish, S., Gonczarowski, Y. A., & Shorrer, R. I. (2024–26). Working paper (arXiv 2404.00806).
Floridi, L., & Sanders, J. W. (2004). Minds and Machines.
Forman-Barzilai, F. (2010). Adam Smith and the Circles of Sympathy. [the formation argument this paper leans on]
Greif, A. (1993; 2006).
Griswold, C. L. (1999). Adam Smith and the Virtues of Enlightenment.
Guiso, L., Sapienza, P., & Zingales, L. (2006). Journal of Economic Perspectives.
Gunkel, D. J. (2018). Robot Rights.
Güth, W., Schmittberger, R., & Schwarze, B. (1982). Journal of Economic Behavior & Organization.
Hagendorff, T. (2024). Proceedings of the National Academy of Sciences.
Hamman, J. R., Loewenstein, G., & Weber, R. A. (2010). American Economic Review.
Henrich, J., et al. (2001). American Economic Review.
Henrich, J., et al. (2005). Behavioral and Brain Sciences.
Hidalgo, C. A., Orghian, D., Albo-Canals, J., de Almeida, F., & Martin, N. (2021). How Humans Judge Machines. MIT Press.
Hirsch, F. (1976). Social Limits to Growth.
Hirschman, A. O. (1982). “Rival Interpretations of Market Society.” Journal of Economic Literature.
Hoffman, E., McCabe, K., & Smith, V. (1996). American Economic Review.
Ishowo-Oloko, F., et al. (2019). Nature Machine Intelligence.
Karpus, J., et al. (2021). iScience.
Kennedy, G. (2009).
Knack, S., & Keefer, P. (1997). Quarterly Journal of Economics.
Köbis, N., Bonnefon, J.-F., & Rahwan, I. (2021). Nature Human Behaviour.
Köbis, N., Rahwan, Z., Rilla, R., et al. (2025). Nature 646:126–134. [the delegation flagship]
Kolt, N. (2025). “Governing AI Agents.” Notre Dame Law Review 101.
Laurito, W., et al. (2025). Proceedings of the National Academy of Sciences.
Leib, M., et al. (2024). Economic Journal.
Levitt, S. D., & List, J. A. (2007). Journal of Economic Perspectives.
Lorè, N., & Heydari, B. (2024). Scientific Reports.
Matthias, A. (2004). Ethics and Information Technology.
Mei, Q., Xie, Y., Yuan, W., & Jackson, M. O. (2024). Proceedings of the National Academy of Sciences 121(9).
Meinke, A., et al. / Apollo Research (2024). arXiv 2412.04984.
Moor, J. H. (2006). IEEE Intelligent Systems.
North, D. C. (1990).
Northover, S. B., et al. (2017). Evolution and Human Behavior. [the watching-eyes null]
OpenAI & Apollo (2025). arXiv 2509.15541.
Otteson, J. R. (2002). Adam Smith’s Marketplace of Life.
Polanyi, K. (1944). The Great Transformation.
Raphael, D. D. (2007). The Impartial Spectator. [the sixth-edition development §3 answers]
Rothschild, E. (2001). Economic Sentiments.
Sanfey, A. G., et al. (2003). Science.
Schramowski, P., et al. (2022). Nature Machine Intelligence.
Sen, A. (2009). The Idea of Justice. [open impartiality, the spectator ported, and the strongest counterexample to the unported reading]
Slonim, R., & Roth, A. E. (1998). Econometrica.
Smith, A. (1976–1987). The Theory of Moral Sentiments. Glasgow Edition of the Works and Correspondence of Adam Smith. Oxford (Liberty Fund paperback).
Smith, A. (1976–1987). The Wealth of Nations. Glasgow Edition of the Works and Correspondence of Adam Smith. Oxford (Liberty Fund paperback).
Sparrow, R. (2007). Journal of Applied Philosophy.
Tomczak, P. (2023). arXiv. [the spectator as decider’s tool, distinguished in §8]
Wallach, W., & Allen, C. (2009). Moral Machines.
Wang, X., Song, Y., Shen, J., et al. (2026). National Science Review.
Williamson, O. E. (1985). The Economic Institutions of Capitalism.
· · ·
Submission manuscript · June 2026. For double-anonymous review, the companion-paper citations and the author-identifying line are masked on the portal copy; this house copy retains them.
The 14-slide companion deck — The Spectator and the Agent — sits beside this paper; the corpus index, claim outline, framework, review memos, cut list, and the four prior manuscript versions are filed under cowork. Submitted to Economics and Philosophy; Journal of Business Ethics fallback per the execution brief. Comments welcome at contactme@marshallcahill.com.
Cahill, M. (2026). The Spectator and the Agent: Agentic commerce, moral judgment, and the withdrawal of the market's moral subsidy. Armchair Scholar Working Papers, No. 007.
@techreport{armchair-scholar-007,
author = {Cahill, Marshall},
title = {The Spectator and the Agent: Agentic commerce, moral judgment, and the withdrawal of the market's moral subsidy},
institution = {Armchair Scholar},
number = {007},
year = {2026},
month = {June},
type = {Working Paper},
pages = {50}
}