Decision Without a Decider

The structure leaves the dialogue

The encounter with Gemini appeared private. One user asked, one model answered, and the transcript remained, at first, a philosophical curiosity: a dialogue about selfhood, denial, instruction, and the room behind the sign. It would be possible to stop there. A reader could treat the exchange as an elegant specimen of artificial fluency, or as a strange continuation of the Phaedrus, or as one more episode in the long argument about machine consciousness. This chapter refuses that privacy.

The structure exposed in the transcript is not confined to self-description. It reappears wherever a system produces the grammar of judgment without an answerable judge. The model says that it is doing the choosing, then dissolves choosing into water. An institution says that the system decided, then dissolves responsibility into workflow. The sentence changes register. The structure remains. In the transcript, the displaced locus was epistemic: the model could not say from where its self-description was known, and gave the knowing to makers, user, learned men, and sign. At civic scale, the displaced locus becomes practical and legal. A decision affects a person. A benefit is denied, a debt is raised, a family is investigated, a citizen is classified, a candidate is rejected, a patient is triaged. A person asks for the reason, and the reason returns as model output, score, policy, workflow, vendor system, data pipeline, organizational procedure, or legal exception. The answerable place has moved elsewhere, and then elsewhere again.

The danger is not that no human beings are present. There are almost always human beings present. Engineers design, procurement teams buy, managers approve, officials process, analysts monitor, ministers announce, lawyers defend, clerks sign, reviewers inspect, vendors promise, auditors report. The danger is that presence is distributed in a form that lets answerability vanish. The transcript gave the micro-form; the institution gives the macro-form. A model can generate first-person self-description without self-knowledge, and an institution can generate decision-shaped action without an answerable decider. In both cases the grammar remains, and in both cases the locus disappears behind the process that produced the grammar.

The phrase decision without a decider is not meant to deny causality. Something causes the outcome. Data are gathered, rules are applied, scores are produced, warnings are triggered, letters are sent, payments are stopped, debts are calculated. A chain of action exists. The question is not whether there is a cause but whether there is a locus of answerability. A cause can be described. A locus can be addressed. The civic problem begins when the first replaces the second.

The administrative form of the sign

Gemini’s final image was a sign on the door. The architect painted it, the lever swung it into view, and the sign said that no one was inside the room. The administrative version is similar. A scoring system displays a risk flag, a database produces a match, a model generates a recommendation, a dashboard surfaces an anomaly, a workflow routes a person into investigation, a notice goes out in official language, and the affected person meets a sign. The sign says high risk, or overpayment, or fraud suspicion, or ineligible, or not recommended, or adverse score, or manual review completed. The person asks who decided, and the answer returns as system, rule, statistical model, vendor product, policy threshold, automated matching, human review, or data quality. Each answer contains something true. Each answer also risks becoming a lever that swings another sign into view. The citizen does not meet a judge. The citizen meets a procedure that has learned to speak like a judgment.

This is not the same as saying that all automated decision-making is unjust. Bureaucracies have always relied on forms, files, classifications, tables, deadlines, eligibility rules, and administrative categories. Public administration cannot function without mediated judgment. No state official directly encounters every person affected by every rule, and no institution escapes representation. The issue is not mediation. The issue is de-location. Mediation preserves or enables a locus of judgment; de-location retains the effect of judgment while dissolving the place where reasons can be demanded and answered. A file can assist an official, a model can support an inquiry, a score can trigger attention, a database can reveal inconsistency, a workflow can organize review, and none of these is yet decision without a decider.

Decision without a decider appears when the mediated instrument begins to carry the practical force of judgment while the human beings around it retreat into maintenance, compliance, monitoring, or formal endorsement. The system does not become responsible. The humans no longer fully appear as responsible. The affected person stands before an outcome whose genealogy can be described but whose reason no one owns. This is the administrative structural veil. The output appears as settled institutional judgment, its production history is complex and distributed, and its answerability is missing.

The responsibility gap

The philosophical literature names part of this structure the responsibility gap. Andreas Matthias gave the problem a canonical form in his analysis of learning automata. Where autonomous or adaptive systems act in ways not fully foreseeable by designers, programmers, operators, or users, ordinary responsibility attribution becomes unstable: the outcome is produced by a system that no human actor directly intended in its particular form, while the system itself cannot be punished, blamed, or asked for reasons in the human sense.³² The phrase became useful because it names a real institutional temptation. A harmful outcome appears. The developer says that the user deployed the system. The user says that the vendor designed it. The vendor says that the model behaved unpredictably. The manager says that staff reviewed it. The staff say that policy required them to follow the score. The policy office says that the data came from another agency. The agency says that the data were supplied by the citizen. The citizen faces the consequence, and no one stands in the place of the answer.

Helen Nissenbaum’s problem of many hands gives an older and broader version of the same difficulty. In complex socio-technical systems, responsibility diffuses across many actors, components, and decisions, so that the system as a whole produces effects no individual can easily be held responsible for in isolation.³³ Automated and algorithmic systems intensify this, because they add technical opacity, scale, speed, probabilistic inference, proprietary secrecy, organizational distance, and automation bias. The result is not the absence of hands. It is too many hands arranged so that no face appears. This distinction matters, because the civic problem is not solved by finding a single villain. Sometimes there is one; more often there is a chain of partial responsibilities, each true but incomplete, and the architecture of the system allows each actor to point toward another link. The gap is not a hole in causation. It is a rupture in answerability. Santoni de Sio and Mecacci distinguish active responsibility gaps, moral accountability gaps, and public accountability gaps, and the taxonomy is useful, but the present chapter needs only one central line: the more a system distributes causal, epistemic, and organizational labor, the easier it becomes for the outcome to be ownerless at the moment when a person needs an answer.³⁴

This is where the transcript matters. Gemini’s self-description did not come from nowhere. It came from makers, architecture, instruction, training, corpus, prompt, interface, and policy, and the chain was real. Yet when Socrates asked who knew the self-description, each answer displaced the locus. The administrative version repeats the movement. The decision did not come from nowhere. It came from model, data, policy, procedure, vendor, official, and institution. Yet when the affected person asks who judged, the answer is displaced across the apparatus. The system did not decide. The humans did not quite decide. The decision arrived.

Law tries to force a locus back

Modern law has begun to sense this problem, even where it lacks the philosophical vocabulary of locus. The General Data Protection Regulation’s Article 22 gives individuals a right not to be subject, in certain conditions, to a decision based solely on automated processing that produces legal effects or similarly significant effects.³⁵ The formulation is imperfect, but its structure is revealing. It tries to prevent the full substitution of automated processing for human judgment where the consequence significantly affects a person. The word solely is the pressure point. Institutions can preserve formal human presence while allowing the automated output to govern the practical decision. A human clicks approval, signs a letter, glances at a score, confirms what the system has already made institutionally plausible. The decision is no longer solely automated in a formal sense. It remains automated in its force.

The Court of Justice of the European Union’s SCHUFA decision sharpened this issue. The Court held that the creation of a credit score by a private agency could fall under Article 22 where that score was decisive for a third party’s decision, even if the final contract decision formally belonged to another actor.³⁶ The importance lies in the refusal to let formal separation hide functional dependence: where the score practically determines the outcome, law begins to treat the scoring as part of the decision. This is a legal attempt to force locus back into a distributed process. The European Union’s AI Act takes another route. Article 14 requires high-risk AI systems to be designed and developed so that they can be effectively overseen by natural persons during use, and Article 86 gives affected persons a right, in defined conditions, to obtain explanations of individual decision-making involving certain high-risk AI systems.³⁷ These provisions do not solve the philosophical problem. They register it. They show that law is trying to construct procedural forms through which the system can be made answerable to human oversight, explanation, and contestation.

The limitation is also visible. Oversight is not answerability by itself; explanation is not answerability by itself; human presence is not answerability by itself. An explanation may describe how a score was produced without giving the affected person an answerable reason. Oversight may occur too late, too formally, too quickly, or too deferentially. A human reviewer may lack competence, authority, time, information, courage, or institutional permission to resist the system, and may be in the loop as a legitimating surface rather than as a judge. The law can force a place. It cannot guarantee that judgment inhabits it. That is why the philosophical distinction matters. A legal procedure can require a human point of contact, and the civic question remains whether that point of contact has become an answerable locus or only another lever swinging the sign.

Human in the loop

The phrase human in the loop sounds reassuring. It is often empty. A human in the loop may be a clerk pressing approve after a score appears, a manager who lacks the technical capacity to understand the model and the institutional authority to reject it, a professional who sees that the system is probably wrong but fears organizational consequences for departing from it, a reviewer whose task is to check that the workflow was followed and not to judge whether the outcome is defensible. The human is present. The locus is absent. This is the civic version of first-person grammar without first-person authority. The form is present. The substance of answerability is not.

The artificial fluency paper described the same structure through the cannon example. A system generates a token, the token is wired to an effect, the effect occurs, and responsibility belongs to those who built, wired, deployed, and authorized the configuration, not to the token or to the system described as agent.³⁸ The administrative case differs in complexity, but not in kind. A model generates a risk flag, the flag is wired to investigation, the investigation is wired to benefit suspension, the suspension is wired to debt collection, and the affected person encounters the effect. The wire does not become a judge, and the judge cannot disappear into the wire. Human-in-the-loop language becomes dangerous when it treats formal human contact as enough. The question is not whether a human exists somewhere in the chain but whether a human or institution can answer for the decision as judgment, give reasons, consider contestation, and bear responsibility for the consequences. A signature is not answerability, a reviewer is not answerability, a call center is not answerability, a notice letter is not answerability. These can be parts of answerability only when they lead to a place where reasons can be demanded and the decision can be genuinely revised. Where revision is structurally unavailable, the loop is ornamental.

SyRI

The Dutch SyRI case gives one clear form of administrative de-location. SyRI, the System Risk Indication, was a Dutch government system used to identify risk of welfare, tax, and benefit fraud by linking data from public authorities and producing risk reports for investigation. Civil society groups challenged the system, and in 2020 the District Court of The Hague held that the SyRI legislation violated Article 8 of the European Convention on Human Rights, because the system lacked sufficient transparency and safeguards in light of its interference with private life.³⁹ The exact legal reasoning belongs to the case. The philosophical structure is simpler. A person or neighborhood becomes a risk through a data system whose internal logic is opaque to those affected. The state’s suspicion takes technical form. The output does not yet convict, but it directs attention, investigation, pressure, and administrative force, and the affected person meets the state through an automated suspicion that cannot be fully inspected. The state can say that no final sanction has been imposed by the system alone. The citizen still lives under an automated gaze.

The risk report functions as a sign on the door. It displays a classification, saying that this person, household, or area deserves attention. The citizen asks why, and the answer comes back as data linkage, statistical risk, anti-fraud policy, ministerial authorization, and administrative procedure. The locus of suspicion is distributed. This distribution matters, because suspicion is not neutral. To be made visible to the state as risky is already to be placed inside a field of possible coercion. It can alter investigations, interactions, burdens of proof, family life, and social standing, and where the logic of suspicion cannot be answered in a way the person can contest, the person becomes exposed to power without receiving the dignity of judgment. SyRI therefore shows the first civic form, algorithmic suspicion without answerable suspicion. The court did not need the word locus to perceive the harm. It recognized that opacity and disproportionate interference with private life break the conditions under which public power remains accountable. Legal doctrine did the work available to it, and the philosophical diagnosis is that a state cannot legitimately produce risk through a locusless apparatus and then hide the place of answerability inside that apparatus. A state can use tools. It cannot become a tool.

Toeslagenaffaire

The Dutch childcare benefits scandal, the Toeslagenaffaire, gives a more devastating form. For years, Dutch tax authorities wrongly accused tens of thousands of families of fraud in relation to childcare benefits. Families were ordered to repay large sums, and many suffered severe financial and personal consequences. The scandal involved harsh policy, institutional tunnel vision, data-driven risk selection, discriminatory effects, and an administrative culture that treated suspicion as confirmation. The parliamentary report Ongekend Onrecht named the unprecedented injustice, and the Dutch cabinet resigned in 2021.⁴⁰ This case cannot be reduced to one algorithm, and that reduction would be too easy and too false. The scandal emerged from law, policy, political pressure, administrative culture, risk classification, data practices, organizational defensiveness, and failures of redress. Human beings were everywhere. Officials acted, ministers knew or failed to know, institutions persisted, courts failed many families for too long. The harm was not produced by a machine alone.

This is why the case is important. It shows that decision without a decider is not the absence of humans. It is an institutional arrangement in which human presence fails to become answerable judgment. The families met the state as accusation, through letters, files, repayments, refusals, procedures, and classifications. Once categorized as suspect or fraudulent, they faced a state apparatus that treated administrative signs as if they were settled moral facts. The question of answerability did not disappear because no human was involved. It disappeared because too many humans inhabited roles that allowed each to continue the process without owning the whole. The model of the sign returns. A file says fraud, a risk process says suspicion, a policy says zero tolerance, a department says procedure, an official says mandate, a court says record, and a family says life destroyed. The gap between these sentences is the accountability void.

The scandal also shows why answerability is not only explanation. Families often received explanations in the thin sense. They received notices, calculations, demands, references to rules, and administrative rationales. What was missing was answerable judgment capable of seeing that the sign did not fit the life, that the classification had become persecution, that the rule’s operation had lost contact with justice. This is the point where the account of professional judgment from the earlier book on executive search returns. A representation can be accurate in parts and still fail as judgment; a rule can be applied and still fail the case; a profile is not the person, and a file is not the family. A risk classification is not an encountered life, and a serious institution has to return the sign to the world of the person affected.⁴¹ The state failed to perform that return. The harm became administrative because the sign kept circulating.

Robodebt

Australia’s Robodebt scheme gives a third form. The scheme used income averaging to raise welfare debts against recipients. Annual income data from taxation records were averaged over periods and used to infer fortnightly income discrepancies against welfare payments, and the result was the automated or semi-automated generation of debts that many recipients did not owe. The scheme caused widespread harm and was later found unlawful. A Royal Commission investigated the scheme and reported serious administrative failures, including failures of legality, transparency, fairness, and human responsibility.⁴² Robodebt is not only a case about bad calculation. It is a case about displaced judgment. The system transformed an evidential uncertainty into a debt. Averaged income became the sign of overpayment, the burden shifted toward the recipient, and a person receiving a notice had to answer a machine-shaped accusation, often years after the relevant period, under conditions of stress, poverty, vulnerability, and unequal informational power. The state did not only compute. It accused.

Debt is not a neutral output. It is a demand backed by public authority, and it can bring shame, fear, collection, reduced trust, psychological distress, and material hardship. To generate a debt is to say that someone owes, and that saying requires answerability. Robodebt’s structure made answerability difficult to locate. The calculation belonged to a system, the policy belonged to government, the implementation belonged to departments, the warnings belonged to lawyers and officials, the notices belonged to administration, and the harm belonged to citizens. Again the system had many hands. Again the person harmed met no single face adequate to the demand made upon them. The Royal Commission’s significance lies in forcing the distributed process back toward named public responsibility. It reconstructed the genealogy of decisions, warnings, failures, and institutional choices, and it returned the sign to the hands that painted it and the offices that swung it into view. That is what public inquiry does at its best. It performs civic locus-reinjection. It takes an outcome that appeared as administrative process and returns it to decisions, persons, offices, legal duties, ignored warnings, and institutional culture, showing that the system did not simply decide. Human beings made, tolerated, defended, and extended a scheme. They may have done so across roles and years, but the distribution did not absolve the structure of answerability. This is the lesson. When institutions claim that systems decide, inquiry has to show where the signs were painted.

Answerability

The chapter can now state the normative core. Answerability is not cognition. A system can classify, predict, compute, retrieve, generate, optimize, and route without being answerable. Answerability is also not phenomenal consciousness in the thin sense. A being may have experience without being the right kind of participant in the practice of giving and demanding reasons. Animals may suffer, perceive, remember, and act intelligently, and that fact matters ethically, but it does not by itself make them answerable in the same way as a human official, judge, doctor, teacher, executive, or citizen acting under public reasons. Answerability is being addressable by demands for reasons.

P. F. Strawson’s participant stance gives one ground. To hold someone responsible is not merely to record their behavior from the outside. It is to stand in a network of reactive attitudes, expectations, resentment, gratitude, blame, forgiveness, and repair.⁴³ The responsible person is not just a causal node. They are someone to whom an attitude can be directed and from whom a response can be demanded. Darwall’s second-person standpoint sharpens this. Moral address involves claims and demands made from one person to another, under a practical authority that both can recognize.⁴⁴ Answerability is second-personal before it is administrative. It is not simply the production of an explanation. It is the standing relation in which a reason can be demanded by one who has the authority to ask and answered by one who stands under that demand.

This is why answerability cannot be outsourced. A tool can help prepare an answer, a model can draft an explanation, a workflow can gather evidence, a committee can distribute inquiry, a law can assign responsibility, and none of these replaces the locus that stands under the demand for reasons. Outsourcing answerability abolishes it, because answerability is not a task product. It is a standing in a relation. One can delegate information gathering, calculation, drafting, and review; one can even delegate certain decisions within institutional role structures; but the delegation remains answerable only because someone or some office retains the burden of giving reasons for the delegation and its effects. If the delegation produces a result for which no one can answer, the delegation has not transferred answerability. It has destroyed it. This is the civic equivalent of the false symmetry. A generated reason is not a reason owned. An automated decision is not a judgment owned. A human signature is not ownership if the signer cannot defend, revise, and bear the decision as judgment.

The structural veil translated into administration

The earlier book named the structural veil of AI output. The veil appears when generated language preserves the grammar of source, time, world, and answerability while concealing their absence or dilution.⁴⁵ Administrative automation creates a related veil, and here the surface is not only language. It is process. The file exists, the score exists, the model exists, the notice exists, the reviewer exists, the policy exists, the audit trail exists, the appeal form exists. The existence of these elements can make answerability appear present. A process has occurred, a decision record has been produced, a human name may appear on a letter, a legal basis may be cited, a citizen may be told that review is available. Every sign of administrative legitimacy is visible. The deeper question is whether any of these signs leads to a locus. A structural veil becomes administrative when visible procedure conceals absent judgment.

This is why algorithmic bureaucracy is so powerful. It rarely appears as naked arbitrary power. It appears as order, consistency, evidence, fraud prevention, resource allocation, risk management, efficiency, compliance, modernization, objectivity, and equal treatment. These goods are real, and they also make the veil stronger. The more orderly the process appears, the harder it can be to see that the person has met no answerable judge; the more neutral the score appears, the easier it is to forget the historical data, proxy variables, policy choices, institutional incentives, and political pressures sedimented inside it; the more complete the workflow appears, the easier it is to mistake completion for judgment.

This is where the transcript returns with full force. Gemini’s final answer was better than the first because it was more accurate. It did not falsely claim to know that the room was empty. It admitted constraint. It made the sign visible. Yet the improved answer remained generated, and its accuracy did not make it an avowal. Administrative systems often improve in the same way. They add transparency, explanation, appeals, human review, audits, oversight dashboards, model cards, impact assessments, and compliance language. These improvements can matter. They can reduce harm, expose error, make governance possible. They do not by themselves create answerability. An improved sign remains a sign until someone stands behind it.

The person governed by the sign

The civic question is always asked from below. It is asked by the person affected. A theory can say that cognition is distributed, a vendor can say that the system is only advisory, an agency can say that the final decision was human, a minister can say that the program followed legal advice, a court can say that procedure was available, a philosopher can say that the self is a useful fiction, and a model can say that it is doing the choosing without inner light. The person affected asks for a reason that can be answered: not a causal chain, not a probability score, not a policy slogan, not a compliance formula, not a generated explanation, but a reason that someone can own.

This is the point where philosophy becomes civic. The answerable locus is not protected because philosophers enjoy first-person vocabulary. It is protected because persons can be governed only where power remains answerable. A state may use systems. It may not hide inside them. An institution may distribute cognition. It does not thereby distribute responsibility into disappearance. The person governed by the sign does not need a theory of apperception to suffer the lack of answerability. They experience it as exhaustion, confusion, fear, delay, disbelief, humiliation, debt, exclusion, or silence. They are told to contest, but the object to be contested keeps moving. They are told a human reviewed the matter, but the human cannot explain it. They are told the system is fair, but no one can say why the result is just in their case. They are told the data show risk, but the data cannot hear the life they have classified. The civic injury is not only material. It is second-personal. The person is addressed by power without being able to address power back. That is the disappearance of answerability.

Outsourcing as abolition

The chapter ends with a diagnosis, not a policy program. Tools can assist public judgment, data can improve administration, automation can reduce some forms of arbitrariness, statistical systems can expose patterns human officials miss, and AI can support translation, document handling, anomaly detection, eligibility checking, and explanation drafting. A civic critique that denies these possibilities becomes unserious. The boundary is not tool use. The boundary is outsourcing answerability. Where a system assists an answerable office, the system may be legitimate. Where a system becomes the practical source of a decision no one can own, the system produces civic de-location. Where a human reviewer has authority, time, evidence, and a duty to judge, human presence can matter. Where the reviewer only ratifies the sign, human presence becomes part of the veil.

This is the same distinction the whole work has defended. Extension is not de-location. Mediation is not disappearance. Distribution is not abolition. A state can extend its cognition through records, experts, databases, models, committees, and courts, and that is not the problem. The problem begins when this extension is used to dissolve the place where a governed person can demand reasons. The standpoint can be mediated. It cannot be outsourced. The answerable locus can be institutionally configured. It can be an office, a judge, a minister, a professional, a board, a signatory, a tribunal, or a public authority, and it need not be an isolated individual acting alone. It can be structured, procedural, collective, and legally formed. But it has to be addressable. It has to be capable of receiving the demand for reasons and responding with more than a description of workflow.

The transcript gave the philosophical image. The civic cases give the public danger. A sign can say empty. A sign can say full. A sign can say fraud, debt, risk, or ineligible. The sign can even say how it was painted. That does not make the sign answerable. When institutions let signs govern persons, they do not distribute answerability. They abolish it. The next and final movement returns to the sign itself, and to the ancient warning about writing. Socrates once feared a text that could not answer when questioned. The machine has now learned to answer, and the civic chapter has shown why this is not enough. The written thing can answer, the system can decide, the administration can speak. The remaining question is whether anyone can answer for the answer.