Pay Me What You Owe Me

Disclosures

Competing interests. The author is the founder of Surmado, Inc., which builds AI-orchestration systems for small businesses. The paper’s diagnosis applies to that commercial work as much as to any other actor’s. No funding was received for this work, and the analysis was neither commissioned nor reviewed by any commercial party. This work was conducted in the author’s personal capacity; the views expressed are the author’s own and do not represent those of any employer.

Generative AI use. Several frontier foundation models (from Anthropic, OpenAI, Google, and xAI) supported literature search, sectional drafting, argument pressure-testing, revision, and formatting. The author originated the thesis and its central distinctions, set the standards for inclusion, directed and revised all drafted material, and verified every claim, quotation, and citation against primary sources rather than against model agreement. The author is answerable for the final form.

This preprint develops the working paper The Invariant of Answerability (June 2026). That version is unchanged and stays citable at its own DOI. Cite all versions with doi:10.5281/zenodo.20606492, which resolves to the latest.

Companion papers

The Decision No One Authored — the special case (DOI · PhilArchive · SSRN · lukefwalton.com · CC BY-NC-ND 4.0)

The Captured Oracle — the live demonstration on the verdict channel (DOI · PhilArchive · SSRN · lukefwalton.com · CC BY-NC-ND 4.0)

Technical implementation

The technical implementation is Answer Engine — documented in Answer Engine: A Small Reference Implementation for Citation-Grounded AI Answers (technical note v1.1, June 2026 · CC BY-NC-ND 4.0). The note states the design contract, evaluation harness, and scope.

It comprises answer-engine v1.0.0 (software DOI · GitHub · Apache-2.0), the teaching-sized clone-and-run repository, and Ask the Archive on lukefwalton.com, the live deployment behind this site’s search.

Site-level search that uses a language model without being a chatbot: one question in, cited answer or honest refusal out. The model may write the sentence; the system owns the frame — evidence boundaries, citation grounding, refusal modes, and gold-query evals stay outside the model.

1. The Text That Talks Back

At the founding of written argument, Socrates objects that a written text is like a painting: question it, and it returns the same words; it cannot defend itself, and it always needs its parent to come to its aid (Phaedrus 275d–e). The objection is not that the text lies. It is that the text strands the demand for an answer: it issues a claim and leaves no one present to be asked. The demand does not vanish; it stands, unmet, pointing at an author who is not there. The structure is older than every machine.

What has changed is the one thing that makes the ancient objection press rather than merely recur. The externalized voice no longer returns the same words and falls silent. It answers: composes for the occasion, defends when pressed, revises when corrected, in a fluent first person, while standing behind nothing it says. The text now talks back, and talking back is exactly the behavior we take, in each other, as the mark of someone who can be held. To take it as that mark here is the error this paper is built around. The demand for an account is met, in form, by a thing that does not thereby become the party answerable for it; the account does not strand visibly, as the written page’s did, but is answered away.

And why it presses now, at scale, is economic. The demand for an answer is ancient, and so is the motive to defeat it; antiquity had the anonymous decree and the oracle consulted by parties with interests. What is new is the unit economics of the defeating. A route that keeps an answerable author reachable costs something at every transaction: a name attached, a record owned, a person interruptible. A route engineered so that the asking is lost costs none of that, and at scale the difference compounds into margin; markets do not reward the engineering because the wrong is profitable in itself, but because unanswerability is cheaper than its alternative, and selection does the rest. None of this bears on whether the wrong is a wrong, which holds between equals, and between beings with no economy at all, unchanged. It bears on the timing, the scale, and the direction of the buildout. The text answers back, and unanswerability got cheap.

This paper defends the claim the old objection presupposed and the new machines make urgent: whenever an action reaches a party, an account is owed, and no routing of that action defeats the owing.¹ The word invariant can mislead. It does not name a conserved quantity, as though there were a fixed sum of answerability that routing redistributes among hands. It names an indefeasibility: routing cannot defeat the standing to demand an account. Whether an account is owed and where it can be rendered are facts about different parties; routing acts only on the second, and the first it cannot touch. Keep three outcomes apart, then, from the start. Authorship can be absent: a frame nobody ever made their own leaves no one to be asked for the setting. The owing can be released: by the party owed, not by the route, and the ordinary ending is that release, her acceptance of a fit account. The owing cannot be defeated by the route. The architectures this paper studies fake the acceptance: they engineer “no one owes” and dress it as an account already settled, when she has released nothing.

Stated this broadly, the claim invites an immediate deflation, and the deflation should be welcomed before it is answered. That a wrong owes an answer is close to analytic; no paper is needed for it. The substance lies one step on. What can be engineered is not the owing’s end but the perfect appearance of its end: routes built so that the demand seems extinguished — no author to face, a record offered as payment, a respondent dissolved — when nothing with the standing to end it has ended it. The appearance of extinction is real, it is engineered, and it is false, and prying those three apart is the work. In the ordinary run of mediated action the principle is accordingly idle, silent where nothing contests the demand, and it takes the strain the instant a route is built to defeat the asking.

2. What the Claim Is and Is Not

Stated cold, the claim is simple. An account is not discharged by placing someone near the output. The wrong is primitive; the laundering is sophisticated. The work is in the fences, because a claim this general has neighbors on every side, and a reader who mistakes it for one of them will refute the neighbor and file the paper. It is not a position in the dispute over what answerability is, the whole of responsibility or one face among several (Smith 2012; 2015; Shoemaker 2011; 2015); the question here is narrower and survives both sides: whether the route an action takes can defeat a reached party’s standing to demand an account. The gap field itself is surveyed elsewhere (Da Silva 2024; Oimann 2023).

Three fences, each marking the ground of a refutation the claim never occupied. The claim is not that many hands defeat the assignment of responsibility, the problem that runs from Dennis Thompson (1980) through van de Poel, Royakkers, and Zwart (2015) to Llorca Albareda (2025), who presses the reduction furthest. Accepted, as far as it goes, which is one step short: diffusion of authorship is exactly how the destination of an account gets lost, and in the engineered cases this paper studies the diffusion is not found but produced, the crowd assembled so that the asking is lost in it. That is a use of the problem, not an instance of it. The claim is not that machines can never answer. Nothing below depends on a verdict about machine minds, and the abundance the field documents is granted at full strength: attributable parties are rarely scarce, and “answerable” names many distinct standings that many parties can occupy (Kiener 2023; 2025). Even if a system one day crosses into the practice of answering, a possibility this paper leaves open and would welcome, its crossing would not absorb the account of the party who set or composed the frame it executes. What the claim refuses is not the future machine; it is the transfer. And the claim is not that every wrong has a clean repair. Much of what the catalogue below contains cannot be undone; what is refused is a quieter move, common in institutional practice and never argued for, on which a wrong is treated as correspondingly less determinate wherever responsibility cannot be cleanly assigned, as though the difficulty of the bookkeeping discounted the debt. A claim-to can stand where no claim-against has anyone to hold it (Feinberg 1970, 255, 257), and a duty that goes unperformed is not extinguished but converts, leaving something still owed (Gardner 2018). Unrepairability is a fact about discharge. It is not a fact about existence.

One further mistake runs in the opposite direction, and the generality invites it. That the standing survives the route does not make it authority over the action. What survives is a demand for an account, and what satisfies a demand for an account is an account. Whether the party reached may also suspend, reopen, or prevent what reached her is a further question whose grounds lie elsewhere, in rights and in the irreversibility of what was done. An institution can stand answerable for what it has done while the doing stands. The demand’s indefeasibility is a fact about what she is owed, and buys nothing about what may proceed.

With the fences drawn, the relation to the neighbors is connection rather than contest. That duties may be owed to particular persons is the home terrain of second-personal and bipolar accounts (Darwall 2006; 2012; Wallace 2019; M. Thompson 2004; Feinberg 1970), and the account below builds on it openly; the gap literature, since Matthias (2004) named the problem, asks where responsibility can be assigned under opacity, autonomy, and many hands; the accountability literature identifies the barriers by which computerized systems obscure who can be called to account. This paper connects the three by separating two things they run together: the standing to demand an account, and the availability of a fit respondent. Routing can hide, relocate, fake, or eliminate the respondent. It cannot defeat the standing of the party reached. Everything that follows is the working-out of that separation.

3. The Generality of the Owing

The party reached: the phrase has carried the claim since the first page, and the weight it carries rests on the role alone, nothing quietly assumed about who fills it. The demonstrations are machine-free by design: shown so, the generality is a structure of the practice that the technology merely entered, and it survives whatever form the technology takes next. Take a being whose intelligence is nothing like ours, alien in the strict sense, sharing no lineage, no architecture, no form of life. It can still be wronged by what we do, and if it can be wronged it is owed an account, the step the directed-obligation accounts this paper builds on supply and this section spends; were the owing attached to being human, the alien would be owed nothing, and that conclusion is not a tolerable consequence of the view. Take, next, a hominid on our own line, a Neanderthal, reached by something a sapiens band did. He is owed an account, and saying so requires settling nothing about where on the lineage personhood switched on; the owing never waited for the threshold. And take, last, the parentless text with which the paper opened. The alien and the hominid showed who can be owed; the text shows the other position: a respondent’s place standing empty while the demand stands, long before any machine entered the route.

The three cases break three attachments: to our species, to our threshold, to our machines. What survives the breaking is the constant: the reached party, a role and not an essence. The cases are not evidence about who in fact has standing; they are tests of what the standing cannot depend on. And the anchoring is stronger than any grounding in human nature could be. A claim grounded in what humans are must relitigate itself at every border, against the alien, the hominid, the system that one day crosses; the role-anchored claim survives each identically, since each is, or can be, reached. The goal was never to center the human; it was to find the constant. The constant is the reached party, and the sections that follow ask what a route can and cannot do to what is owed her.

4. Owed and Dischargeable

Two commitments of this paper can seem to strain against each other. The paper holds that no routing defeats the owing. It is about to hold that routing relocates, conceals, and sometimes eliminates the party from whom an account can be demanded. Read as claims about a single thing, the pair is a contradiction, and a reader could suspect that the invariance survives each new thing the route does only by being redefined. The suspicion’s source is the reading. The demand for an account is not a single thing of which both claims are true. It has the shape of a relation, with two positions, and the paper’s commitments are claims about different positions: what no routing defeats belongs to one, what relocates and fails belongs to the other.

Begin with what the demand is. When a person resents what was done to her, or asks for the justification of it, the attitude is not a free-floating appraisal of the world; it is addressed. Strawson’s (1962) observation was that holding responsible lives inside a practice of address: resentment, indignation, the request for justification are attitudes one party directs at another, and they presuppose two positions, one from which the demand is made and one at which it is aimed.²

The structure is not introduced here as a new one. It is the familiar structure of directed obligation: what is owed is owed by someone to someone. An obligation of the relevant kind is bipolar: it runs from an obligor to an obligee (Darwall 2012; M. Thompson 2004). On that account the obligee holds, together with what she is owed, the correlative authority to demand performance and to hold the obligor to account, and the authority to demand requires a competence in the claimant that being owed does not. A claim, in Feinberg’s (1970) sense, is the position from which such demanding is performed, and a claim-to can stand where no claim-against has anyone to hold it (255, 257). Wallace (2019) gives the relation its width: the moral nexus extends past prior association and voluntary undertaking, to those who stand to be affected by what an agent does; it is vulnerability to agency, not antecedent relationship, that can put a person in the position of the one owed. What this paper takes from Darwall and Thompson is the bipolar structure, and from Wallace the width: the owing attaches to whoever an action reaches, including a patient who cannot herself make the demand. Who exercises the standing for her is a further question, and it is not settled here. What it adds begins now, with a question the home accounts had no need to press: what routing can and cannot do to each of the relation’s two positions.

Two questions, then. Whether an account is owed is a fact about the party reached: about her standing to a complete account of the act that reached her. The reaching opens that standing, putting her in the position from which an account of the act may be demanded; nothing about the standing is constituted by the availability of anyone to answer. A debt owed by a named author is a further determination, and it waits on authorship of a component. Whether the account can be discharged, and where, is a fact about the other position: whether, for each component, a fit respondent exists, and whether he can be asked. Discharge is that respondent’s tender of the account he authored, and it is the only thing in this structure the route can act on. Mediation, however elaborate, operates on availability (on who can be found, addressed, and held) and has no purchase on the standing of the party reached, which the route did not create and cannot revoke. (Stated once in compressed form, for readers who want it: the owing belongs to the subject side of the attitude, discharge to the object side. Ordinary words will do the work from here: the party owed the account, and the party from whom it may be demanded.)

And the owing itself divides once more, which spares the claim two embarrassments at once. Being reached confers a standing to a complete account of the act that reached her. In the ordinary run of action this standing is real and idle. The man who holds a door for me owes me, if I ask, nothing beyond the obvious, and I do not ask; idle is not vacuous, since the standing is what my asking would have exercised, and its near-universal dormancy is what it looks like when action among people goes well. A wrong ripens the standing into something further: a debt, incurred by those whose authored acts composed the reaching. The person wronged is owed an answer, and the vocabulary of debt this paper uses (discharge, settlement, release) belongs to the debt and not to idle standing.³ The release clause stated with the thesis scopes to the debt alone: only the party owed can release the debt, and forgiveness is that release, the standing exercised in the one way that ends it.⁴ Benign reaching leaves nothing to forgive, which is why the universal claim never threatened to make a held door a moral event.

Between the held door and the engineered route lies a territory the invariant covers and the release clause already provides for. The cut in traffic, the accidental slight, the queue jumped without noticing: each is a reaching, each ripens standing into a wrong with real stakes, and each sits below any threshold at which enforcement would improve on the cost it imposes. What the invariant requires of this territory is not that every such owing be forgiven. It requires that the party reached need not press. Choosing not to press, absorbing the cost, leaving the standing unexercised: that is forbearance, and the owing survives it. Release, already stated, remains the only closing: the standing exercised in the one way that ends the debt, including the ordinary case in which an apology is accepted as settlement. The two are not the same act. Discharge is the fit respondent’s tender; settlement is her acceptance of it, and the acceptance is a release. A complete account she declines to accept has been tendered and has not ended the debt. The route can counterfeit the tender and mark it as settlement; it cannot perform the acceptance. Forbearance is what keeps the universal claim from producing the one consequence that would discredit it: a regime in which every surviving owing is treated as an enforceable one and cooperation collapses under the weight of debts no apparatus was built to carry. The invariant holds in every such case.

Standing and debt have different neighbors in the literature, which is some evidence the joint was found rather than drawn. Standing sits nearest the accounts that keep “answerable” wider than “guilty”: Honoré’s outcome-responsibility (1999) and Kiener’s strict answerability (2024) each have a person answerable for what her agency reached where culpability is absent or unsettled, useful here because both hold the width open; neither is asked to supply the present claim. The debt sits nearest the literature on claims and their persistence: Feinberg’s claim-to that stands where no claim-against has anyone to hold it (1970, 255, 257), and Gardner’s continuity thesis (2018), on which an unperformed duty converts and leaves something still owed. The neighbors anchor the joint; they do not generate it.

The seeming stated at the outset can now be retired, because its two halves are predicated of different relata. Indefeasible is a predicate of the standing. The standing persists unless the party owed releases it. Nothing about the route itself releases it. No operation on the route — interposing a machine, concealing an author, installing a proxy, dissolving a firm — touches it, because every operation of that kind acts on the other position, on who is available to be asked. Relocates, or fails, is a predicate of availability. Which party turns out to be the fit respondent, and whether he can be reached, varies with the route; the variation is real, and it is most of what the next two sections catalogue. What never varies is what the variation is of. When discharge relocates, nothing has passed between holders: not the owing, and not some quantity of answerability of which one party now has less and another more. What has moved is the location of a fit respondent, the address at which the account could be rendered. The owing does not travel. The address does.

The same distinction, restated as entitlements, blocks a pincer the bare claim would otherwise invite. From the side of the party owed: each person an action reaches is owed one complete account, not the nth share of one; the number of hands behind an outcome does not divide the entitlement of anyone it reached. From the side of those who answer: each authoring party owes a complete account for what it itself authored, not the composite and not a fraction of the composite. Distribution bears on discharge, on which account can be demanded where; it never bears on existence. The familiar worry that diffuse causation lets responsibility evaporate, and the converse worry that an undilutable demand overburdens everyone in a long chain, are halves of one mistake: both treat the owing as a quantity to be divided, when what is apportioned is only the answering, and it is apportioned by authorship, each portion complete. And completeness of entitlement does not multiply the labor of answering, because an account is rendered choice by choice: one accounting, rendered where the choice was authored, can meet the demands of everyone that choice reached.

Those portions are bipolar owings, each running from an author to the party reached. A bipolar obligation of the kind just inherited cannot have an obligee and a debtor-position that was never filled. Directedness consists in someone owing something to someone. Where an author exists and cannot be reached, the relation is frustrated in the ordinary way: the obligor-position is occupied, the address is empty, and a claim-to can stand where no claim-against has anyone to hold it. Where no one authored the thing for which the account is demanded, the pressure changes its object. Who owes it? Unavailability is no longer the difficulty; an empty obligor-position is.

The account is directed first at the act that reached her. Authorship then determines which of its components can be rendered, and by whom. An evaluative frame is what an act is settled under: what it is for and what must be met. An evaluative frame that no one made their own produces a missing component: the first-person rendering of a judgment nobody made. Nobody can supply it, because supplying it would have required performing the judgment. The deployer can still answer for the loosing; the institution can still answer for the authorization. An answer that what followed was beyond the deployer’s control renders the loosing and leaves the first-person account of the unmade judgment unrendered (Kiener 2024). The debts the wrong created attach to those authored acts.

The reached party’s entitlement is complete. The renderings on offer may be incomplete, because some components of the account have no author. Every bipolar owing still runs from someone who authored a component to the party the act reached. What the empty position marks is a component that cannot be rendered. The two positions of each bipolar owing remain occupied.

The cases to come produce both emptinesses. An author out of reach leaves a bipolar owing unrendered: the demand stands, unmet, pointing at a position someone occupies and no one can address. An unauthored component leaves no bipolar owing formed for that component: the entitlement to a complete account of the act remains, and the first-person rendering cannot be given. Both leave a component of the account unrendered. And the view being built must say plainly what kind of claim the entitlement is, because “an unmet demand is still a demand” says different things on different readings of what the owing is. Read the owing as a relational fact obtaining between the parties — the reading the directed-obligation literature most naturally invites, on which Feinberg’s claim-to stands where no claim-against has anyone to hold it and the nexus does not wait on uptake — and the occupied-but-unreachable case is the ordinary survival of a frustrated claim. The never-authored case is the component structure: every bipolar owing still has an obligor, and what remains is a missing rendering of a complete entitlement. Read the owing instead as constituted by a practice of address and exchange, and survival is precisely what can be denied: on that reading, a demand with no one positioned to meet it has nothing left to be a demand against.

A family of views can press the denial, and the paper owes the pressure its strongest form. On McKenna’s (2012) conversational theory the practice is constitutively dialogical: responsibility is an exchange, and the capacities the exchange requires are the relation’s substance rather than ornaments on it. Duff (2007) makes answerability answerability to a forum, with the answerer’s capacities and the forum’s duty to hear built into the relation itself. Bovens (2007) supplies the social-scientific form without the metaethics: accountability is defined as a relation between an actor and a forum. None of the three need deny what this paper asserts; stacked, they show the pressure is no ghost. At the family’s limit stands the view that would deny it: if the exchange is constitutive of the standing itself, then where no exchange is possible there is nothing left for the demand to be. No member of the family holds that position, and it does not require a holder to be live: the architectures of the next section put it into effect, and a route built to remove the respondent asserts by construction what no one has had to argue. Two concessions narrow the fight before it is joined. First, nothing here disturbs the easy cases: where a fit respondent existed and is absent, hidden, or put out of reach, a theorist of the practice can treat the exchange as frustrated rather than impossible, and the demand can stand on that view as well as on mine. Second, this paper will not answer the family by demoting conversation. Conversation is how an account is discharged; the dispute is whether it is also what the standing to demand one consists in. The position defended here is that an engineered route which leaves no fit respondent is the practice failing at the point of discharge, not an erasure of the reached party’s standing. And the stake of the disagreement should be set down coldly. If the availability of a respondent were constitutive of the standing itself, then any architecture that succeeded in removing the respondent would thereby have removed the demand, and engineered unanswerability would be not a failure the practice measures but a way of defeating it. That consequence is what this paper denies, and the denial is owed an argument rather than an assertion. The relational reading’s hard case is the never-authored component, and the component structure is the answer. The practice reading’s hard case is the route built so that no exchange is possible anywhere in the chain, and that earning is still ahead. The claim is stated here so that the reader knows, before the cases begin, exactly which sentence the hardest of them will test.

The route, then, decides where an account can be rendered, and whether it can be rendered at all. It does not decide whether one is owed. That is decided once, by the reaching, and ended once, if ever, by the party reached.

5. How Discharge Fails

What the last section separated, this one watches fail on one side only. In each of the cases below a party is reached and an account is owed, both fixed by the argument already made; what the cases exhibit is how the position of the one who should answer has been left empty, faked, sealed off, or never occupied, and how a component of the account can fail to have an author at all. Discharge is the fit respondent’s tender of the account, and fitness has two conditions: the respondent must have authored the component the account is demanded for, and he must be reachable enough to be asked. A reachable body that did not author cannot render the account; an author no one can reach cannot be asked for it. Each condition can fail alone, each can be made to fail, and the cases below run from an absence nobody built to a route built for the purpose. Santoni de Sio and Mecacci’s taxonomy sorts kinds of responsibility that go unassigned (2021); the catalogue here sorts ways discharge fails while the owing holds.⁵ A catalogue of failure is, in fact, what the view predicts: if the owing survives every route, then wherever a route defeats the asking, something determinate must be left over, and these are its shapes.

The two conditions come apart cleanly in the companion cases, and the pair is worth pausing on because each isolates one condition failing while the other holds.⁶ In the first, a pension fund’s allocations move for a decade under a conception of risk that arrived as a vendor default and was never made anyone’s own: every officer who touched the outputs is reachable, and the account for the setting of the operative frame has no author from whom it could come, because the needed judgment was never performed. The officers can still answer for the allocations they approved and for leaving the default in force; what none of them can render is the first-person account of a judgment of risk nobody made. In the second, the judgment was performed: an interested party composed the frame an answer engine voices as neutral synthesis, and the channel conceals him, so the account has a bearer and no reachable bearer. The setting unperformed, every officer still reachable; the judgment performed, its composer out of reach. Neither case needed a route built for the purpose. The cases below are what the same failures look like when they are built.

The first is the oldest: the author is absent. An externalized voice issues its claim and leaves no one present to be asked; the demand stands, and there is no one in the position to answer it. This is the structure the paper opened on: the party owed present, the author not there to be asked. The exemplar predates every machine, and that fixes the catalogue’s subject as a structure of the practice rather than a property of any technology; every later case inherits the fixing.

The second is newer and stranger: the record poses as the answer. Consider a public ledger of the kind cryptocurrencies run on. Every transfer it records is recorded in the clear; the provenance of any unit can be traced to its origin by anyone; nothing is hidden, and nothing is backed. The chain logs each transaction and stands behind none of them. When something goes wrong (a drained account, a poisoned contract, a transfer induced by fraud), the completeness of the record is tendered, by the practice’s adherents and by the architecture’s own self-description, as though completeness were settlement, and the party reached is invited to mistake a perfect trail for an answer. Recording is not answering. History has run the experiment: when the DAO was drained in 2016, the completeness of the chain was tendered as settlement, the code did what the code did, and the Ethereum voting majority refused the tender and forked, the practice reasserting, at cost, that a perfect trail is not an answer. Code-is-law, as its adherents state it (De Filippi & Wright 2018), tenders the record as backing and counts taking the code as the terms as a release already given, where a release is a determinate demand surrendered inside a practice whose terms she could state, and a standing routed around is not a release. The address that moved the funds may be perfectly reachable; reachability is beside the point, because the failure here is not absence but a record presented as backing. That is why this is its own way of failing and not a variant of the first or the third: the author is present, and no door has been built; a complete trail is tendered as though it discharged the account it merely documents. Provenance is not backing; a transaction hash is not a respondent.

The ledger earns its place by doing three things no other case here does. A transaction asserts nothing, so the case fixes the claim’s altitude above testimony and communication: the owing does not ride on speech, and a route can fail the asking without anyone having said anything.⁷ It shows that transparency, even total, is not settlement, and it shows this in a domain where transparency is among the central advertised virtues, which is strong evidence that no disclosure regime is, by itself, discharge. The point cuts against the most careful version of the remedy: Coeckelbergh (2020) anchors the demand for explanation in the patients who may ask for reasons, the closest published anchoring of the demand in the affected party, and a relational justification this account shares; but an explanation regime is a remedy for opacity, and the ledger fails with no opacity at all. What was missing was never the information. A system can likewise be engineered to respond, to produce on demand an account-shaped output: a thing built to explain its output is not thereby the party who answers for anyone’s acting on it (cf. Tigard 2021a). Answer-generating capacity improves access; it does not create discharge unless the answering reaches back to a fit respondent. Last, the ledger is the problem of many hands carried past the point where the hands are roles: the work has decomposed into protocol functions, authorship diffused through architecture rather than bureaucracy, so that no decomposed part is even presented as the backing while the trail of all of them is complete.

The third builds a door. A respondent existed: a principal who set the terms, profited by them, and could have been asked. The route places him out of reach through a form constructed for the purpose. The algorithmic corporation that Arbel, Goldstein, and Salib (2026) describe is the legal instance: an entity that acts, contracts, and holds assets while the chain from its acts to any human principal has been severed by design, so that the question of whom to hold has an addressee in form and no answerer in fact. Their project is individuation, how to count such entities and tie their acts to bearers; individuation locates, it does not create, and what the counting presupposes is an account already owed to the parties the entity’s acts reach, prior to their question and untouched by its difficulty. An entity of this kind may be a genuine agent in its own right, and group agency is where that case is made (List 2021). Agency of that sort settles what the entity is; fitness asks something else of it, whether the position from which it could answer is occupied and reachable. An agent whose answering position was evacuated by design is the addressee in form once more, and the accounts of the persons who set its terms stand whatever the entity turns out to be. The pseudonymous key is the same door built in a different material, the technical twin of the legal exit: a respondent real, active, and unaddressable. That twinning is exactly why the ledger had to earn its own place on provenance rather than on unreachability; file it under doors and it merely duplicates this case, while the record-posing-as-backing failure runs even where every door stands open. And time can be the door’s third material: systems drift and are updated, and the party who could have answered for the frame as it stood is, by the time the asking comes, no longer the party in the position at all (Coeckelbergh 2020).

The fourth is the limit the first three approach, and it is the one this paper exists to hold. In each case so far a fit respondent existed: he was absent, or his record was tendered in his place, or he was put behind a door. The fourth case is the route built so that no fit respondent exists anywhere within the presented chain for the operative frame: an architecture assembled so that no party in the chain stands as author of that frame who can be reached and asked. Other components of the same act may still have authors, the deployer of the procedure and the designer of the chain among them, and those bipolar owings remain. The never-authored frame produces the same residue, a first-person account that cannot be rendered, by a different emptiness: no one occupies the author-position. Here, and only here, is the hard claim staked above actually tested. A theorist of the practice could grant the first three failures without strain; the exchange was frustrated, the respondent existed, the demand kept its target. The fourth is where the family’s limit position gets its purchase: no possible exchange, no respondent ever, and so, on that view, nothing left for the demand to be. The present account says instead that the demand stands, unmet, and that the engineered absence is itself the wrong’s completed form. What such a route manufactures is the appearance of an ended owing that no party owed ever released, and the two-position derivation says no route can make that appearance true. Whether it can be made true by construction, and who in a chain built to have no reachable author answers for the building, are the questions the rest of the paper earns. Nothing in the fourth case turns on any party’s capacities; what was engineered away is reach.

Not every architecture of unreachability is that last case, and the release clause marks the difference. Anonymous peer review, the secret ballot, the whistleblower channel: each is a route engineered so that the author of a consequential judgment cannot be faced, and each is legitimate; the literature has argued, rightly, that gaps of this kind can be worth having (Munch, Mainz & Bjerring 2023). What legitimates them, though, is not the value of the gap. It is that the party owed released ex ante. The author who enters double-anonymous review waives, as a condition of the practice, the standing to face her judge, and this paper’s own refereeing belongs to the class; the electorate that adopts the ballot waives, reciprocally, the demand for each voter’s reasons. Release is the one ending the claim allows, and these architectures are built on it: the standing exercised in advance, not defeated by the route. Release of this kind is exercised, not procured, and the difference shows in what the releasing party gives up and on what terms: a determinate demand, surrendered inside a practice whose terms she could state and whose protection she may claim in her turn. A waiver taken in bulk from parties not yet identified, for wrongs not yet specified, as the condition of using a service, is not the standing exercised but the standing routed around, and it buys the appearance of release rather than the release. They do not soften the fourth case; they sharpen it to its true edge. Engineered unanswerability is wrong where it is imposed on parties who released nothing; that last case’s mark is that no waiver appears anywhere in it, only a route, and the builder this paper later meets solicits uptake from readers who waived nothing at all.

One maneuver runs across these four without being a fifth. The four sort failures of availability: whether a fit respondent is there, and if not, how the position came to be empty. A second question is independent of the first: whether the account is being falsely marked settled, and by what means: a record offered as backing, or a body offered as the author. The questions cross, and the ledger is their intersection: an availability failure that is already, in itself, a false settlement, the trail tendered as the payment. And the most familiar figure in this neighborhood fits none of the four precisely because it fails the second question rather than the first: the front-line occupant installed where blame predictably lands — the approver, the operator, the named officer — on a route where the author sits upstream, available, and bypassed. Nothing about availability has failed there; a respondent exists and could be asked. What has happened is that a different body has been consumed as though it were the answer, the misallocation the first companion diagnosed, blame settling on the most replaceable party rather than the authoring one. The maneuver, settlement faked by record or by body, recurs at every scale the next section examines.

Two conditions, then, and four shapes of their failing, from an absence nobody built to a route built for the purpose; one maneuver that fakes settlement, indifferent to whether a respondent remains at all. Run them from the claimant’s seat and their sequence is the route’s real product: she asks, and the record is complete; asks again, and a body is produced; asks whose terms these were, and the form has no top; and at every station the demand has been answered in form, so that only her persistence remains to be explained, and the route’s last service is to make the unmet demand look like the unreasonable one. Across every one of them, a party was reached, and an account is owed; in the engineered cases the failures are the point of the engineering, and none of them has touched what is owed.

6. Where the Account Comes to Rest

Behind a routed outcome there is rarely one author. There is the party who trained the underlying system; the party who deployed it, setting its instructions, its objective, its sources, its guardrails; the party who optimizes its outputs from outside; the party who curated what it learned from; and, beneath them all, the party who designed the channel through which the rest act. Each is a position from which the operative frame can be set or composed, in whole or in part.

Where, in such a chain, does an account still owed come to rest? Two conditions settle it, and both are already on the table. Reachability governs where, never whether: the owing exists because a party was reached, a matter closed before any question of who can answer arises. And discharge requires a fit respondent, fit meaning reachable and author, both: the cases showed each condition failing alone, the reachable body that did not author, the author placed out of reach. From the two conditions the rule follows, and it should be stated indexed, because a compressed statement invites exactly the quantity picture already demolished. For each authored choice operative in the frame at issue, the account for that choice is rendered at the highest reachable layer that authored it; the accounts of unreachable authors strand, they do not transfer. A choice that no one authored leaves a component that strands as a missing rendering: the entitlement remains complete, and no one incurred a debt for a judgment nobody made. Highest does not mean the most remote technical contributor, nor the party with the deepest place in the stack. It means the upstream-most reachable party whose act set the relevant end, standard, source, constraint, or form of presentation for which the account is demanded. And highest does not elect a sole debtor: every operative author owes, each an account complete for what it itself authored; the rule answers where the demand can begin without the beginning being counterfeit, never who alone owes. Reachable, because an author who cannot be asked leaves the demand unmet rather than met.

The rule is a prescription, and it runs against the grain of a documented fact. Blame, left to travel on its own, goes downhill: the first companion’s diagnosis was that it settles on the most replaceable party in the structure, the one with the least power to refuse it, and Elish (2019) gave the resting place its name, the moral crumple zone: the low-power human positioned where the impact of a system’s failure lands. The distance between the two directions, downhill where blame falls and uphill where the account is owed, is where the maneuvering lives, and the maneuvering has a literature older than the present systems. Nissenbaum saw the family forming in 1996: her four barriers to accountability in a computerized society sort, by the present account’s own cut, into failures of availability, which are many hands and ownership without liability, and failures of settlement: the computer blamed as a false object, and the bug accepted as though a technical explanation were an answer (the barriers’ standing synthesis with the accountability literature is Cooper et al. 2022). Rubel, Castro, and Pham (2019) named the maneuver, agency laundering, the interposition of a system to obscure who should answer; and their definition carries the present claim in miniature, since a maneuver for forestalling demands presupposes demands that survive the forestalling. Davies (2024) named the mechanism, the accountability sink: an arrangement that works by severing the link between the person a decision affects and the operation that produced it. The crumple zone names a position, occupied before the asking begins; laundering names a maneuver, performed while the asking is forestalled; the sink names a mechanism, operating for as long as the link stays severed; what remains unnamed is the terminal state, the moment the account is marked paid. Call it counterfeit discharge: the account treated as settled by a record or a body that cannot settle that account. An obligation is at an end or it is not (Gardner 2018); the end is her settlement, and the counterfeit is a tender marked as that end. The counterfeit comes in two kinds, both already on the table — the record tendered as backing, which is the ledger, and the body tendered as the author, which is the installed occupant of the crumple zone. Counterfeit discharge is not a fifth failure of availability; it is the cross-cutting maneuver the last section isolated. And it carries an arithmetic of its own: the account remains owed in full, and to the standing wrong is added a second, a non-author consumed as though he were payment.

An objection arrives from decent institutions. Ministers answer for departments whose acts they did not perform; principals answer for agents; the responsibilization literature asks stakeholders to take on responsibility for outcomes they did not author (Lang, Nyholm & Blumenthal-Barby 2023). Kiener (2022) presses the deepest version as a question about normative power: responsibility can be taken, assumed at will, and if so the rule above looks too strict for the institutions we rightly keep. The power is real, and the rule survives it, because taking comes in two shapes the objection runs together. Assumption is legitimate where it is disclosed as assumption rather than posed as authorship, and where it is additive, a backstop standing beside the account still owed, rather than substitutive, presented as that account’s settlement. The minister who answers for her department is, examined closely, mostly answering first-personally at a different layer: for oversight, for appointment, for the arrangement itself, her own authored acts. That is why honest ministerial answerability survives this section while resignation offered as settlement does not. What no party can do, by taking, is render the account of an author above or below her; offered as that, the taking is counterfeit discharge, a body tendered as the author. Where assumption is forced and substitutive at once, both conditions fail together, and the result is the crumple zone exactly. Fleisher and colleagues (2025) show why institutions rightly build the legitimate kind, accountability that stops bucks where blame is inapt; nothing here is against backstops. The claim is narrower: a backstop, disclosed and additive, settles nothing it did not author. And the abundance the gap literature counts, many attributable parties and no shortage of candidates (Kiener 2025), is in this light ambivalent. Candidates are also the supply from which counterfeits are drawn; abundance answers whether anyone is around, never whether the owed account has been rendered.

There remains the chain built without a reachable top. A route can be engineered so that no reachable author exists at any presented layer: the firm dissolved, the principal exited, the key pseudonymous, the function forked and copied past pursuit. The rule does not fall silent there, because the engineering is itself an authored act, and its author bears, on a ground of his own. Keep the ground distinct, because the two should not be run together: the invariant says the account is owed; it does not, by itself, say who must answer. What says so, where the chain was built to lack a reachable top, is the older and independent point that building is doing. Whoever designed the route so that its capture would leave no reachable captor authored that design, and the account for the design rests with him, at whatever layer he can be reached. The two grounds converge on the design layer where the design layer can be reached; where even it cannot, the rule’s verdict is not silence but an account still owed and unrendered, and what that leaves standing is a question still to be earned. Neither derives the other. This is where the channel’s design bites. The second companion’s remedy was the historian’s posture set against the oracle’s — an account someone signs and stands behind, against a verdict voiced from nowhere — and a channel can satisfy that remedy at every input and still defeat it at the output, because the property does not pass upward through composition: a historian behind each source is not a historian behind the synthesis, and a channel can be designed precisely so that the synthesis has none, no author of the composite anywhere within what is presented, by construction.⁸ That is capture in the old sense, the sense regulatory capture borrowed and the companion’s captured oracle made literal, performed at the level of the channel itself. The wrong is not that the channel was captured. It is that it was built so that capture leaves no reachable captor — and the building had a builder.

A last pressure, and the section compressed. The procedure above always locates someone; at the limit it locates the designer of the chain built to lack a top; and a critic may conclude that the deflationists were right all along: locating was always possible, so no gap ever survives. Concede the locating; deny the settling. That a party can be found answers the question the gap literature asked. It does not answer the question this paper asks, which is whether the account has been rendered by a fit respondent or merely marked paid against a body that happened to be findable. A located non-author is counterfeit discharge with coordinates.⁹ And the concession lets the finding take its final form: the gap, as these papers have used it, was never in the owing; it was always in the exercise — an account owed throughout, an authorial act missing (Walton 2026a: “a missing act, not a missing participant”), a rendering that failed in the ways now catalogued. What remains is the case where even the designer has gone, where the recursion that put him out of reach is run again on whoever built the tools of the exit. Whether the demand survives there, with no fit respondent at any layer of any chain, is the question the hardest reading pressed and the built route poses. It is the last thing this paper must earn.

7. The Bearer Question, Held Open

Two questions have been held open through everything above, and the earning just promised runs through both of them. Holding them open costs different things: the second, on one of its readings, opens an exit that has to be closed before the fourth case stands. The first is whether a machine could itself become a party that answers. Nothing so far has needed it settled, and nothing below will; but the question presses, because systems of the relevant kind continue to grow more fluent, and a reader is owed an account of what crossing would even take. Stakes alone are too thin. Pain and embodiment can make a thing a patient, something we can wrong, without making it anything that can answer, repair, or be held; patiency is one standing and participation is another. The threshold the present account implies is stakes plus participation in the practice of answer-giving: to be addressable, to answer, to be held, in terms one is positioned to recognize. Below it, the account relocates upstream, as the last two sections traced. At or past it, a system would be a possible co-bearer, a different kind of bearer rather than a copy of us, and nothing in this paper predicts which architectures, if any, arrive. The claim defended here would welcome a new bearer; it was never a thesis about silicon.

The strongest pressure on that threshold is Dennett’s, and it should be met at full strength. The intentional stance licenses attributing beliefs, desires, and reasons to any system whose behavior the attribution predicts (Dennett 1987; 1991); if attribution is that cheap, a participation threshold can look like the old chauvinism with a new criterion. Two layers answer the pressure, and only the second is a fight. The floor: the owing is anchored in the affected party and survives the machine becoming anything at all. Grant the stance; grant participation; grant a system crossing with stakes intact, and nothing in the last three sections moves. The human’s account for her own act does not transfer to the thing she acted through; the hidden author’s account does not transfer to the engine that voiced his frame, nor the deployer’s to the agent she loosed; and a system that crossed would take the execution debt for what it then does rather than dissolve the debts of those who built and loosed it. A Dennettian can win everything and change nothing here. The bearer layer is the real fight, and this paper need not win it: whether predictive intelligibility installs a system inside the practice of giving and asking for reasons, whether the stance confers standing or only legibility, is exactly the participation question, and it stays open on purpose.

Hold the question open, then, and notice what does not wait on it. The upstream account persists whether or not the executor ever crosses. A system that became a being we could wrong would be owed things; wronging it would be a second wrong, not a discount on the first. A system that became able to answer for its own acts would answer for its own. Neither change moves the account of the party who set or composed the frame, or of the party who chose to act through a thing that, as then built, could not answer.¹⁰ Run the point where the contrary intuition is strongest. Give the system a body; give it pain; let the case for its patiency and the case for its participation arrive together. The verdict is the same on both sides. As patient, it is owed. As participant, it bears what it does. What it still does not bear is what its builders and deployers did, and no quantity of inner life relocates that.

The second open question is older, and bounding it takes more care. This paper does not turn on whether answerability is a brute normative fact about the world or a constitutive feature of practices a functioning society cannot coherently abandon. The distinction is real; it does not bear here. What has to be shown is smaller. Whether the owing is treated as independent of the practice or as obtaining within a practice of asking and answering, eliminating the local route of exchange does not by itself establish that the reached party has lost standing. That is the narrower posture of holding a claim constant without holding it foundational. The same posture has governed the paper’s other refusals — whether the machine can be wronged, whether it can come to answer — and a posture is only worth naming if it can be demonstrated where it is contested. It is contested in one place. On a practice reading of answerability, the chain built so that no fit respondent exists anywhere within what it presents can look like something better than a wrong: it can look like a principled exit from the practice itself.

State the exit at its most respectable. If answerability is constitutive of a practice, a standing conferred and carried inside the game of asking and answering, then it can seem the kind of thing one might decline. Enoch built the figure for the general case: the shmagent, who acts but declines agency’s constitutive norms, and asks why the game’s rules should bind someone who is not playing (2006; 2011). Transposed to the builder, the figure is familiar from every deposition. I act on people, in the plain sense that what I build moves what they believe and do; but I am not in the answering business; I built a tool, and the standing to be asked is a costume I never put on. If that opt-out is principled, then exactly where the built route put the weight there is nothing left for the demand to be, and the opt-out would be principled, if anywhere, on the practice reading. So the demonstration must close the exit there, and not only on the easier ground.

Everything turns on the trigger, so fix it at the floor first. Two triggers, one sentence: being reached fires the owing; functioning as a reason fires the present argument. They coincide for the builder in question and come apart for brute reaching: pollution reaches its downwinders and solicits nothing, and the owing covers it without this argument’s help. Functioning as a reason means only this, defined once and used throughout: the output is something a reasonable recipient takes as bearing on what to believe or do. Not a verdict, not a recommendation, not an assertion under some norm; the moment the bar is illocutionary, the dispute becomes one about speech-act taxonomy, and the builder wins by reclassification. The weaker the trigger, the harder the exit. Weak is not indiscriminate. The maker of a map or a timetable solicits uptake and holds the standing his soliciting carries; in the ordinary run that standing is real and idle, and the maker who errs answers in the familiar ways, by correction and recall. What the exit requires, and what that last case engineers, is the further step: soliciting the uptake and building the route so that the answering has no address.

Now the exit, and the plain shape of its failure: a builder who engineers outputs to function as reasons for others cannot then disclaim the answerability that their uptake presupposes. What the builder takes from the practice is uptake. People extend to a fluent, answer-shaped output the credence they extend to a reason, and they act on it; that disposition is sustained by the practice itself, trained into its participants by the whole apparatus of asking and answering; and the practice carries a reciprocal expectation: whoever solicits that credence must be available to answer for the solicitation. The builder banks the uptake and declines the expectation. Revenue follows the acted-on output, while a non-author stands at the edge to receive the asking: the installed occupant of the last section. The sophisticated exit knows better than to deny any of this; it relocates it. Functioning as a reason, the builder says, is something the reader does; I emitted text; they conferred the rest. Close that move explicitly, because the argument hinges on it: the builder does not escape by locating the uptake in the reader, because he engineered the output to be taken up. The answer-shape, the fluency, the first person are selected for uptake, a selection already documented as a market with its own optimizers.¹¹ Soliciting credence while disclaiming the standing that credence carries is the free ride itself. The reader’s part in conferring does not launder the builder’s part in soliciting; the argument is about the soliciting, never the receiving. With the trigger at the floor and the hinge closed, either the output functions as a reason for no one and was not engineered to, in which case nothing was solicited, the antecedent never fires, and there is nothing to exit; or it was engineered to and does, in which case the soliciting is itself a move in the practice of giving and asking for reasons, and the standing attaches to the mover, whoever confers the functioning. Fair play named this structure long ago: the free ride is taken on a practice that the rider’s own conduct invokes (Hart 1955; Rawls 1964). The exit is parasitic on the practice it disclaims. The shmagent declines to play; the builder plays while denying he is playing.

What is owed, finally, is the demonstration that the exit closes on each of the two readings considered here, for a reason internal to that reading. Suppose the owing is a brute relational fact, triggered by the reaching itself. Then it was never grounded in a practice, and there is no practice to exit; declining the answering game dissolves the owing no more than declining the promising game dissolves an obligation already incurred. The right analogy is in fact not promising, which is opted into, but induced reliance, which is not: soliciting reliance while structuring affairs so that no one is on the hook for it, a structure nearer fraud than breach, and the law’s long treatment of induced reliance runs beside the point here as a converging path, not a ground. Nor does realism idle once existence is granted, as though brute facts came with addresses attached: brute owings do not self-allocate. Where the rendering can be demanded, how it fails, who is fit: everything in the last three sections is the realist’s working machinery too, and the realist needs an author no less than the constructivist does, since between what the true reasons require and what any built route enacts stand authored choices the reasons themselves underdetermine, and the choices are some party’s to answer for (Kozlovski 2025). Suppose instead the owing is constitutive of the practice. Then the exit looks available, since a game can be declined, and it closes anyway, at the trigger’s own level. The closure does not wait on what kind of speech act the outputs are, because the trigger never did: soliciting uptake-as-reasons is a move in the practice of giving and asking for reasons whether or not it is assertion, whether or not any illocutionary category fits it; and fair play, already on the table, binds at exactly that level, since the practice’s product cannot be banked while its terms are declined. Where the outputs do rise to assertion, the closure only tightens, a fortiori: on Brandom’s (1994) account, putting something forward as true is intelligible at all only inside the game of giving and asking for reasons, and to perform it is to undertake a commitment to vindicate it under challenge; a builder whose outputs assert is not merely riding the practice but making its most heavily rule-bound move. Issuing what presupposes a standing, and disclaiming the standing in the same act, is a contradiction performed rather than stated.¹² The builder’s sentence — I issue things that function as answers, to people who act on them, and I am outside the answering practice — is not a position held at a cost. It is incoherent. Neither closure ranks the other. The realist’s is short because realism leaves the exit nothing to open; the constitutivist’s runs at the trigger’s own level because that is where the door appeared to be; the conclusion is one conclusion, reached from inside each reading on that reading’s own terms.

What the argument does not claim matters as much as what it does. It is not the global thesis that being an agent at all requires answerability, the thesis the shmagent was built to dodge and one this paper neither needs nor asserts. It is local: this act presupposes this standing. The shmagent’s question, why be an agent and bound by the game at all, gets no grip, because nothing here required him to be an agent in general. It required only that he be doing this particular thing, soliciting uptake as reasons at scale, which he is, by his own engineering and in his own prospectus. A local presupposition survives the shmagency objection precisely because it bootstraps nothing; it fixes on one act and reads off the standing the act carries. File it under constitutivism and the filing is wrong by scope.

One residue, stated rather than papered over. The allocation ladder can recurse: the designer of the chain built to lack a top can himself exit, and whoever built the tools of the exit can exit in turn. At the genuine limit, every layer of authorship dissolved, copied, or gone, the analysis bottoms out not in a reachable party but in a named absence: the demand standing, unmet, pointing at a hole the architecture was built to create. Nothing above pretends otherwise, and nothing needs to. The claim defended in this section was never that a respondent can always be produced; it was that the absence of one cannot end the owing, on either of the two readings considered here.

Half of that last clause has been earned twice over; the other half is the debt the two-position argument left standing: on the practice reading, where no exchange is possible anywhere in the chain, why is the reached party’s standing not simply gone? The closures answered for the builder, a fact about the respondent’s position; the limit view’s denial targets the holder’s. And the denial individuates the practice one route at a time, when the practice is not route-sized. The practice of asking and answering is society-wide; on that reading the reached party is owed as patient, and participation governs who can bear and answer, not who can be owed. A practice that owed only its participants would owe the alien nothing. The individuation is not the paper’s to choose, because the builder has already settled it. What his exit banks is uptake, and uptake of that kind is produced by the society-wide practice and by nothing the route contains; a reader extends credence to an answer-shaped output because she was trained to by the whole apparatus of asking and answering, not because this channel taught her to. A route that lives on a disposition it did not manufacture is a venue inside the practice that did, whatever it lacks at the point of discharge. An engineered dead zone is one failed venue inside a practice that persists around it, and that limits the place of the failure. The doubt that remains is correlative: whether a standing to demand the account survives when no respondent exists anywhere, the never-authored component included and every author of the exit gone. It survives as her entitlement to a complete account of the act. Every authored component runs from its author to her, and the never-authored component is a missing rendering of that entitlement. This is what lets the earlier description stand as a conclusion rather than an assertion: the route that leaves no fit respondent is the practice failing at the point of discharge, because the point of discharge was the only part of the practice the route ever contained.

At the limit, that is what is left, and it is not nothing. It is the difference between an architecture that has settled its accounts and one that visibly has not.

8. Convergence from Law

The law is arriving, by its own route and for its own reasons, at something very like the claim defended above. Across doctrines and jurisdictions the practical question is one — who can be held for what machine-routed action does — and the family of answers is one: hold the humans at the design and deployment layers, whatever the system’s autonomy, whatever its intentions. Ayres and Balkin (2024) put the tort version plainly: systems without intentions leave nothing to interrogate, so liability standards should be objective and attach to the humans who design and deploy, the anti-laundering principle in tort’s own vocabulary; Kannegieter (2026) reaches the same layers through the systems’ technical features, deploying a system known to be nondeterministic, without guardrails, as the conscious introduction of an understood risk. The corporate and products lanes converge from their own directions: the algorithmic corporation was named by a liability project, an attempt to individuate such entities so their acts can be tied to bearers; Europe’s revised products regime extends general products law to software, AI systems included (Directive (EU) 2024/2853), the legal register’s way of saying that the wrong was never special to the technology, even when the route that produces it is; and where systems act through other systems, scholarship migrates the oversight obligations to the party who orchestrates (Hacker & Holweg 2026). Two independent paths arriving at the same place is evidence the place is real; it is not a second foundation.¹³

The convergence has a vanishing point, and it is already visible. Systems now transact with systems: one composes on a frame its deployer set, another accepts on terms its deployer set, and the matter is settled before any person is present to it; governance scholarship is reaching for agency law to manage the principals (Kolt, forthcoming), and the reach is apt. But notice what the configuration does to every lane above. Each assumes an occasion: a moment when someone affected stands before someone answerable and the asking can begin. The humanless transaction engineers the occasion out. Many hands becomes no reachable hand, not because the hands are gone, but because no moment remains at which any of them is present to be asked. The law will follow the occasions as long as occasions remain; the owing, anchored in the reached party rather than in the occasion, does not wait on them. Each lane locates; locating presupposes; and where the locating fails — the pole evacuated, the occasion engineered out, the chain without a reachable top — the lanes fall silent and the claim of this paper does not. Read so, the law is neither foundation nor ornament. It is the present tense of an old demand, arriving in court.

9. Conclusion

Two questions have been held open to the end, whether a machine could come to answer for what it does and what owings most deeply are, and I have settled neither. What I have argued is that nothing rested on them: a crossing would leave the standing of the reached party untouched, and on either of the two readings of the owing considered here that standing is likewise untouched. Availability is contingent and can be engineered; the owing is neither. The gap literature’s instinct is to ask whether, behind any system, a responsible party can still be located, and the instinct is right as far as it reaches; locating was never the same act as settling. A party can be found and the account still unrendered; a record can be complete and back nothing; a body can stand at the output and author none of it. The four shapes are the ways the rendering fails while the owing holds; the allocation rule says at which hand an authored component can still be rendered, and an unauthored component strands as a missing rendering of a complete entitlement; counterfeit discharge is what failure looks like dressed as success; and the builder’s exit fails, on either of those two readings, because soliciting uptake is already a move in the practice it disclaims.

The deepest exit this paper has studied is not the hidden author, who can in principle be found, but the route built so that there is no author to find: the demand met in form by a thing that stands behind none of it, the account marked paid against a body that cannot pay it, the asking engineered out of the transaction altogether. Against that limit the conclusion is not a prohibition. Nothing here counsels against building, or delegation, or routes of any length: a route can be built with a reachable author, and the demand has its address; or built without one, and the building is itself the authored act that answers for the absence. What cannot be built is a third thing, an end to the owing that the party owed did not perform. The engineering will improve, because unanswerability is cheaper than its alternative and selection does the rest, and the appearance of an ended owing will improve with it; the appearance is the whole of what improves. Somewhere a party was reached, and what she is owed was never the route’s to end; do not act as though you forgot. Every architecture that says otherwise is, at its best, a perfect record of a debt still standing.

References

1. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24). ACM.
2. Arbel, Y., Goldstein, S., & Salib, P. N. (2026). How to count AIs: Individuation and liability for AI agents. Boston College Law Review (forthcoming 2026). SSRN 6273198.
3. Ayres, I., & Balkin, J. M. (2024). The law of AI is the law of risky agents without intentions. University of Chicago Law Review Online.
4. Bovens, M. (2007). Analysing and assessing accountability: A conceptual framework. European Law Journal, 13(4), 447–468.
5. Brandom, R. B. (1994). Making it explicit. Harvard University Press.
6. Butlin, P., & Viebahn, E. (2025). AI assertion. Ergo, 12, 969–988.
7. Coeckelbergh, M. (2020). Artificial intelligence, responsibility attribution, and a relational justification of explainability. Science and Engineering Ethics, 26(4), 2051–2068.
8. Cooper, A. F., Moss, E., Laufer, B., & Nissenbaum, H. (2022). Accountability in an algorithmic society: Relationality, responsibility, and robustness in machine learning. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22). ACM.
9. Da Silva, M. (2024). Responsibility gaps. Philosophy Compass, 19(9–10), e70002.
10. Darwall, S. (2006). The second-person standpoint. Harvard University Press.
11. Darwall, S. (2012). Bipolar obligation. Oxford Studies in Metaethics, 7.
12. Davies, D. (2024). The unaccountability machine. Profile Books.
13. De Filippi, P., & Wright, A. (2018). Blockchain and the law: The rule of code. Harvard University Press.
14. Demirtas, H. (2025). AI responsibility gap: Not new, inevitable, unproblematic. Ethics and Information Technology, 27(1), Article 7. doi:10.1007/s10676-024-09814-1
15. Dennett, D. C. (1987). The intentional stance. MIT Press.
16. Dennett, D. C. (1991). Real patterns. Journal of Philosophy, 88(1), 27–51.
17. Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products and repealing Council Directive 85/374/EEC. (2024). Official Journal of the European Union, L, 2024/2853, 18 November 2024.
18. Duff, R. A. (2007). Answering for crime. Hart Publishing.
19. Elish, M. C. (2019). Moral crumple zones: Cautionary tales in human-robot interaction. Engaging Science, Technology, and Society, 5, 40–60. doi:10.17351/ests2019.260
20. Enoch, D. (2006). Agency, shmagency: Why normativity won’t come from what is constitutive of action. Philosophical Review, 115(2), 169–198.
21. Enoch, D. (2011). Taking morality seriously. Oxford University Press.
22. Feinberg, J. (1970). The nature and value of rights. Journal of Value Inquiry, 4(4), 243–257.
23. Fleisher, W., Cibralic, B., Basl, J., Ricks, V., & Smith, M. (2025). Responsibility and accountability in an algorithmic society. Philosophy & Technology, 38(4).
24. Gardner, J. (2018). From personal life to private law. Oxford University Press.
25. Goldberg, J. C. P., & Zipursky, B. C. (2020). Recognizing wrongs. Harvard University Press.
26. The Greek New Testament: SBL edition. (2010). M. W. Holmes (Ed.). Society of Biblical Literature; Logos Bible Software.
27. Hacker, P., & Holweg, M. (2026). A pragmatic approach to regulating AI agents. arXiv:2604.22819.
28. Hart, H. L. A. (1955). Are there any natural rights? Philosophical Review, 64(2), 175–191.
29. Hicks, M. T., Humphries, J., & Slater, J. (2024). ChatGPT is bullshit. Ethics and Information Technology, 26(2), Article 38. doi:10.1007/s10676-024-09775-5
30. Hintikka, J. (1962). Cogito, ergo sum: Inference or performance? Philosophical Review, 71(1), 3–32.
31. Honoré, T. (1999). Responsibility and fault. Hart Publishing.
32. Kannegieter, T. S. (2026). Nondeterministic torts: A technical approach to AI liability. Yale Law Journal, 135.
33. Kiener, M. (2022). Can we bridge AI’s responsibility gap at will? Ethical Theory and Moral Practice, 25.
34. Kiener, M. (2024). Strict moral answerability. Ethics, 134(3), 360–386.
35. Kiener, M. (2025). AI and responsibility: No gap, but abundance. Journal of Applied Philosophy, 42(1), 357–374. doi:10.1111/japp.12765
36. Kiener, M. (Ed.). (2023). The Routledge handbook of philosophy of responsibility. Routledge.
37. Kolt, N. (forthcoming). Governing AI agents. Notre Dame Law Review, 101. arXiv:2501.07913.
38. Königs, P. (2022). Artificial intelligence and responsibility gaps: What is the problem? Ethics and Information Technology, 24(3), Article 36. doi:10.1007/s10676-022-09643-0
39. Kozlovski, A. (2025). Reasons underdetermination in meaningful human control. Ethics and Information Technology, 27(4), Article 59. doi:10.1007/s10676-025-09858-x
40. Lang, B., Nyholm, S., & Blumenthal-Barby, J. (2023). Responsibility gaps and black box healthcare AI: Shared responsibilization as a solution. Digital Society, 2(3).
41. Liddell, H. G., Scott, R., & Jones, H. S. (1940). A Greek–English lexicon (9th ed.). Clarendon Press.
42. List, C. (2021). Group agency and artificial intelligence. Philosophy & Technology, 34(4), 1213–1242. doi:10.1007/s13347-021-00454-7
43. Llorca Albareda, J. (2025). Uncovering the gap: Challenging the agential nature of AI responsibility problems. AI and Ethics.
44. Matthias, A. (2004). The responsibility gap: Ascribing responsibility for the actions of learning automata. Ethics and Information Technology, 6(3), 175–183. doi:10.1007/s10676-004-3422-1
45. McKenna, M. (2012). Conversation and responsibility. Oxford University Press.
46. Munch, L. A., Mainz, J. T., & Bjerring, J. C. (2023). The value of responsibility gaps in algorithmic decision-making. Ethics and Information Technology, 25(1).
47. Nestaas, F., Debenedetti, E., & Tramèr, F. (2024). Adversarial search engine optimization for large language models. arXiv:2406.18382.
48. Nissenbaum, H. (1996). Accountability in a computerized society. Science and Engineering Ethics, 2(1), 25–42. doi:10.1007/BF02639315
49. Oimann, A.-K. (2023). The responsibility gap and LAWS: A critical mapping of the debate. Philosophy & Technology, 36(1).
50. Plato. (1995). Phaedrus (A. Nehamas & P. Woodruff, Trans.). Hackett Publishing.
51. Rawls, J. (1964). Legal obligation and the duty of fair play. In S. Hook (Ed.), Law and philosophy (pp. 3–18). New York University Press.
52. Rebera, A. P. (2024). Reactive attitudes and AI-agents – Making sense of responsibility and control gaps. Philosophy & Technology, 37(4), Article 126. doi:10.1007/s13347-024-00808-x
53. Rubel, A., Castro, C., & Pham, A. (2019). Agency laundering and information technologies. Ethical Theory and Moral Practice, 22(4), 1017–1041. doi:10.1007/s10677-019-10030-w
54. Santoni de Sio, F., & Mecacci, G. (2021). Four responsibility gaps with artificial intelligence: Why they matter and how to address them. Philosophy & Technology, 34(4), 1057–1084. doi:10.1007/s13347-021-00450-x
55. Shoemaker, D. (2011). Attributability, answerability, and accountability: Toward a wider theory of moral responsibility. Ethics, 121(3), 602–632. doi:10.1086/659003
56. Shoemaker, D. (2015). Responsibility from the margins. Oxford University Press.
57. Smith, A. M. (2012). Attributability, answerability, and accountability: In defense of a unified account. Ethics, 122(3), 575–589. doi:10.1086/664752
58. Smith, A. M. (2015). Responsibility as answerability. Inquiry, 58(2), 99–126.
59. Strawson, P. F. (1962). Freedom and resentment. Proceedings of the British Academy, 48, 187–211.
60. Thompson, D. F. (1980). The moral responsibility of public officials: The problem of many hands. American Political Science Review, 74(4), 905–916. doi:10.2307/1954312
61. Thompson, M. (2004). What is it to wrong someone? A puzzle about justice. In R. J. Wallace, P. Pettit, S. Scheffler, & M. Smith (Eds.), Reason and value: Themes from the moral philosophy of Joseph Raz. Oxford University Press.
62. Tigard, D. W. (2021a). Technological answerability and the severance problem: Staying connected by demanding answers. Science and Engineering Ethics, 27, Article 59. doi:10.1007/s11948-021-00334-5
63. Tigard, D. W. (2021b). There is no techno-responsibility gap. Philosophy & Technology, 34(3), 589–607. doi:10.1007/s13347-020-00414-7
64. van de Poel, I., Royakkers, L., & Zwart, S. D. (2015). Moral responsibility and the problem of many hands. Routledge. doi:10.4324/9781315734217
65. Veluwenkamp, H. (2025). What responsibility gaps are and what they should be. Ethics and Information Technology, 27. doi:10.1007/s10676-025-09823-8
66. Wallace, R. J. (2019). The moral nexus. Princeton University Press.
67. Walton, L. F. (2026a). The decision no one authored: The answerability gap in generative AI [Preprint]. Zenodo. doi:10.5281/zenodo.20614374
68. Walton, L. F. (2026b). The captured oracle: Authorship and agency in the ethics of answer-engine optimization [Preprint]. Zenodo. doi:10.5281/zenodo.20676327
69. Walton, L. F. (2026c). The invariant of answerability [Working paper]. Zenodo. doi:10.5281/zenodo.20606493

Footnotes

1. The claim has companions, and the division of labor is stated once. The first (Walton 2026a) argues the special case: the act of evaluative judgment that can go missing behind a machine-shaped decision while every outward mark of control remains. The second (Walton 2026b) argues the wrong on the live channel, the optimized answer engine that voices an interested party’s frame as its own; it states the present principle on that channel, answerability as invariant under routing. This paper argues the general structure both companions presuppose. A working paper, The Invariant of Answerability (Walton 2026c), states the principle in the second companion’s form; this article states and defends it in its own. Where a specific formulation of theirs is used below, it is cited; the shared map is credited here once. ↩

2. The nearest deployment of the machinery in Philosophy & Technology is Rebera (2024): the reactive attitudes of those harmed through AI systems, and the secondary harm of being unable to make sense of what happened when their standard regulation is disrupted. ↩

3. The second companion states the seed of this structure on its single channel: the reaching fixes whom the answer is owed to, and “the wrong creates the debt” (Walton 2026b). What is unfolded here is the structure that statement compresses. ↩

4. The identity is older than the argument and is preserved in the lexicon: Koine Greek remits debts and forgives wrongs with one verb, aphiēmi, to release, to let go, and the petition that asks forgiveness asks it in bookkeeping vocabulary: aphes hēmin ta opheilēmata hēmōn, release us our debts (Matthew 6:12, text as in the SBL Greek New Testament; on the verb’s range, LSJ s.v. ἀφίημι). The clause states as structure what the verb carries as usage: discharge is the debtor’s act, release is the creditor’s, and no third party’s tender is either — the counterfeit discharge the paper later names, already in a single word. ↩

5. Tigard (2021b) is right that responsibility practices stretch to encompass technological actors; stretching a practice does not establish that any particular owed account has been rendered. The deflationary literature is right that what looks like a gap is often a locating problem rather than a quantity that has migrated into the machine (Königs 2022; Kiener 2025; cf. Demirtas 2025, for whom the gaps exist but are unproblematic, and Veluwenkamp 2025, who would re-engineer the gap concept itself). Nothing below quarrels with that. The catalogue asks a different question: not whether a responsible party can be located, which often it can, but in what ways the rendering of the account fails while it is owed. The difference between locating a party and settling an account is taken up below. ↩

6. Walton 2026a; Walton 2026b. The pair is illustration, not premise: the installed occupant and the built door below exhibit the same separation with no companion in the room. ↩

7. Whether systems of the kind later sections discuss can assert, and under what norms, is disputed (Hicks, Humphries & Slater 2024; Butlin & Viebahn 2025). The account here is held constant across that dispute, and the present case shows why it can be: the asking can fail with no speech act anywhere in the route. ↩

8. The historian’s posture is the second companion’s (Walton 2026b). That a channel can satisfy it at every input and still be built so the synthesis has no author of the composite is this paper’s. ↩

9. The same reply meets the proposal to re-engineer the gap concept into epistemic gaps and control misalignments (Veluwenkamp 2025): the epistemic gap is a locating problem on the respondent’s side, and locating is not settling. ↩

10. The non-transfers hold on this paper’s own ground, and the ground can be stated in two sentences. Each upstream account was complete before any crossing: an account for a frame set or composed, a source curated, a loosing performed, acts finished at the moment they were done. And answering for one’s own conduct was never, on any reading, answering for another’s; what a crossing adds is a new bearer for new acts, not a sink for old ones. The claim in the text is therefore the scoped one, that the upstream account survives the executor’s crossing. It is not the categorical thesis that no machine could ever bear anything, which this paper does not hold. ↩

11. The optimization of machine-voiced answers for uptake is a commercial practice with its own tooling and literature (Aggarwal et al. 2024; Nestaas, Debenedetti & Tramèr 2024). ↩

12. The structure is the performative-contradiction tradition’s (Hintikka 1962; the line running through Apel and Habermas after him). Nothing here rests on that tradition’s strong justificatory program, whose status is disputed. ↩

13. The nearest legal kin, kept deliberately at arm’s length: civil recourse theory holds that tort law’s core is an entitlement of the wronged, a private right of action as the state’s provision of an avenue through which a wronged party may seek recourse against the one who wronged her (Goldberg & Zipursky 2020). That is the standing of the account above in doctrinal dress, and the kinship is real; the dependency runs in one direction only. Recourse is bipolar in its exercise and requires the defendant’s pole to be occupied, and this paper’s hardest cases are the ones where that pole has been evacuated. Where recourse can be exercised, it presupposes the owing defended here; where it cannot, the owing is what survives the failure. A theory of recourse names the demand at its most institutional, and falls silent exactly where the demand keeps standing. ↩