Luke F. Walton

Can a person stay fully in control of an AI-shaped decision and still fail to author it? They may hold the power to override the system, be expected and encouraged to do so, and remain answerable for the outcome after the fact, yet never exercise judgment over the evaluative frame the decision turns on, inheriting that frame rather than authoring it. This article argues that the gap this opens, rather than the question of whether AI is conscious, is where responsibility for machine-mediated action is won or lost. The system's fluency is what widens it: confident, plausible, finished-looking outputs invite a person to accept the machine's framing wholesale, and that temptation operates whether or not anything is actually home. And the temptation deepens rather than recedes as the systems improve, since a reliable track record makes deference rational and erodes the felt need to check at the very juncture where checking still matters. The claim concerns responsibility, not mind: the gap is a missing act, not a missing participant, and it stays open however the machine's status is settled. The dominant design-facing norm, meaningful human control, can close what I call the attributive gap, securing some human to whom an outcome can be traced. It leaves open an answerability gap: the traceable party, even amid a crowd of attributable parties, never exercised the judgment over the decision's evaluative frame that separates holding someone responsible from scapegoating them. Generative and decision-support systems are especially prone to produce this second gap, because their fluency lets a person keep full capacity to intervene while authoring none of what they accept. The norm that closes it is authorship: the answerable exercise of judgment over a decision's evaluative frame, discharged at the five junctures where that judgment is most easily skipped. These are the ends a system serves, the standards its outputs must meet, their verification, their acceptance into action, and the final form for which someone stands answerable. The contribution is this positive account, and its reframing of answerability as a relation that must be exercised rather than a position one occupies. Authorship holds independent of the machine's moral status: it is owed whether or not the system is conscious, and whether or not it is right. It is likewise independent of who can supply it: present systems require a human or institutional author, a sufficiently capable future system might not, and the requirement does not vary with the answer. More fundamentally, control governs the human's relation to the system's operation; authorship governs the human's relation to the evaluative frame through which that operation becomes action. Drawing on Acemoglu and Restrepo's task framework and Acemoglu and Johnson's critique of so-so automation, I argue that the deeper danger is a so-so automation of judgment, evaluative work displaced without new tasks in which human agency is strengthened. I close with three implications and a caution for design. This article is paper one of the Answerability Quartet — four papers on who answers when an AI acts. Companion papers: "The Captured Oracle" (the live demonstration on the verdict channel), "The Invariant of Answerability" (the general invariant), and "Building Answerable AI" (the builder response).

Interposing a machine between an agent and an outcome changes who must answer for it; it does not change whether an answer is owed. This paper defends that claim in its general form. The owing is anchored in whoever an action reaches, not in the route the action takes, and no mediation defeats it; only the party owed can release that account. The structure is old; the demand for an account from an externalized voice that cannot answer for itself is at least as old as writing. What is new is that the externalized voice through which we increasingly act no longer merely strands the demand for an answer, as a written text does, but answers back — fluently, in its own voice, while standing behind nothing. That single change is what converts a perennial structure into a pressing one and is the occasion for the paper. The paper proceeds by distinguishing whether an account is owed from where it can be demanded: the first is invariant under routing; the second relocates, or fails, while the first holds throughout. It locates where an account comes to rest when a chain of authoring hands lies behind an outcome, and it does this without resolving whether a machine could itself ever become a party that answers — a question it holds open by design, on the same terms it holds open whether such a machine could be wronged. The claim is held constant across the metaethical space — true whether answerability is a brute normative fact or a constitutive feature of practices we cannot coherently abandon — and is rested on neither. The contribution is the invariant and the allocation it forces; it is not a verdict on the machine's mind. This working paper states the central claims and framing of a paper in preparation and is circulated to place the claims on record. Companion papers — paper one: "The Decision No One Authored" (the special case); paper two: "The Captured Oracle" (the live demonstration on the verdict channel); paper four: "Building Answerable AI" (the builder response). This is paper three.

Answer-engine optimization is SEO for large language models. Ask an answer engine a question with a contested answer and it may return not a list of sources but a verdict, composed in the engine’s own voice, with no author a reader can see. A party that optimizes what such an engine says authors that verdict while the engine voices it as its own. A public acts on these verdicts, and no one is held to account for them; that is the wrong, and it is committed on the verdict channel, the public-facing surface where the engine pronounces. The familiar answer that the engine is only a tool concedes rather than settles the point: when a tool settles a frame, that frame is someone’s — as a road sign’s limit is a traffic engineer’s — and the someone is what this channel conceals, leaving the user, who authored none of it, as the only party in view. Covert authorship is laundered as neutrality. An interested frame voiced as the engine’s impartial synthesis is the wrong’s sharpest form, but the wrong survives accuracy and disclosure alike, because what laundering diffuses is not the verdict’s visibility but its ownership. The wrong runs along a single axis, the agency of whoever executes the authored frame. At the answer end a person acts on a verdict whose frame was hidden from them, and the hidden author who set it is the one who must answer; at the action end an AI agent acts in their place, the answer for the deed strands at an executor that, as built, cannot answer, and what remains answerable is the deployer’s choice to make a tool an actor. The remedy is answerable legibility: a frame someone owns and stands behind. The disclosure that would secure it must make legible not the sources cited or the process that produced the text, but who owns the evaluative frame the engine voices.

Can a person stay fully in control of an AI-shaped decision and still fail to author it? They may hold the power to override the system, be expected and encouraged to do so, and remain answerable for the outcome after the fact, yet never exercise judgment over the evaluative frame the decision turns on, inheriting that frame rather than authoring it. This article argues that the gap this opens, rather than the question of whether AI is conscious, is where responsibility for machine-mediated action is won or lost. The system's fluency is what widens it: confident, plausible, finished-looking outputs invite a person to accept the machine's framing wholesale, and that temptation operates whether or not anything is actually home. And the temptation deepens rather than recedes as the systems improve, since a reliable track record makes deference rational and erodes the felt need to check at the very juncture where checking still matters. The claim concerns responsibility, not mind: the gap is a missing act, not a missing participant, and it stays open however the machine's status is settled. The dominant design-facing norm, meaningful human control, can close what I call the attributive gap, securing some human to whom an outcome can be traced. It leaves open an answerability gap: the traceable party, even amid a crowd of attributable parties, never exercised the judgment over the decision's evaluative frame that separates holding someone responsible from scapegoating them. Generative and decision-support systems are especially prone to produce this second gap, because their fluency lets a person keep full capacity to intervene while authoring none of what they accept. The norm that closes it is human authorship: the answerable exercise of judgment over a decision's evaluative frame, discharged at the five junctures where that judgment is most easily skipped. These are the ends a system serves, the standards its outputs must meet, their verification, their acceptance into action, and the final form for which someone stands answerable. The contribution is this positive account, and its reframing of answerability as a relation that must be exercised rather than a position one occupies. Authorship holds independent of the machine's moral status: it is owed whether or not the system is conscious, and whether or not it is right. More fundamentally, control governs the human's relation to the system's operation; authorship governs the human's relation to the evaluative frame through which that operation becomes action. Drawing on Acemoglu and Restrepo's task framework and Acemoglu and Johnson's critique of so-so automation, I argue that the deeper danger is a so-so automation of judgment, evaluative work displaced without new tasks in which human agency is strengthened. I close with three implications and a caution for design.