“The algorithm decided” sounds like an explanation, but it is often an evasion. An algorithm has no office, conscience, or liability. It does not sign a warrant, meet a debtor, dismiss an employee, or visit a family whose benefits have been stopped. Yet institutions increasingly describe consequential decisions as if the machine were a small, impersonal official whose judgment must simply be accepted.
The moral danger is not only that an automated system may be inaccurate. It is that automation can rearrange the chain of responsibility until no one appears to have chosen what happened. A vendor blames the customer who configured the tool. The customer blames a model that it cannot explain. A caseworker blames policy. A manager points to the dashboard. At the end of the chain stands a person who has lost a job, a home, access to medicine, or liberty, while every institution involved can say that the outcome was merely produced.
The focus here is institutional rather than ontological. Whether a model counts as a causal or functional agent does not settle whether a bureaucracy has preserved identifiable authority, review, and remedy. Nor does this essay attempt to assign personal guilt in each incident. Its question is whether an organization has arranged responsibility so that someone can still answer for what the organization does.
Distance Is Not Innocence
Bureaucracy has always made moral distance possible. Forms replace faces, categories replace stories, and rules distribute a decision across offices. This division can serve justice: a public procedure is often fairer than an official's whim, and a written standard can protect a person from favoritism. The problem begins when procedure ceases to be a way for people to exercise responsibility and becomes a way for everyone to disclaim it.
Artificial intelligence intensifies this old temptation. A model can turn a judgment into a score, a score into a queue, and a queue into an action taken at a scale no individual could administer. The output arrives with statistical language and a technical vocabulary that discourages ordinary questions. Because the system is complex, those who rely on it can portray themselves as unable to intervene. Complexity then becomes a kind of alibi.
But complexity does not dissolve agency. Someone selected the objective, determined which errors were tolerable, gathered or purchased the data, chose a threshold, approved deployment, and decided what remedy would be available. Other people decided to keep using the system after complaints appeared. Responsibility may be shared, and its exact allocation may require investigation, but it does not disappear merely because many hands touched the process.
The Person Behind the Record
Catholic anthropology gives this problem a precise center. The human being is not an interchangeable case, a risk estimate, or a bundle of behavioral signals. Gaudium et Spes teaches that the human person is willed for its own sake and called to a free, responsible gift of self in relation to God and neighbor (GS 24, 26). This vocation expresses human dignity; it does not create it or make dignity depend on a person's present capacities. The Catechism locates the dignity of the person in creation in the image of God (CCC 355, 1700), and treats conscience as a judgment in which the person is implicated, not as a preference generated by a system (CCC 1776–1782).
This does not mean that every decision must be made from intuition or that institutions may ignore evidence. It means that evidence must remain ordered to a judgment made for a human purpose. A probability can inform a decision; it cannot itself bear the moral meaning of the decision. A model may identify a pattern of missed appointments, but it cannot know from that pattern whether a parent lacked transport, a patient was frightened, or an agency made its own requirements impossible to meet. Context is not noise around the person. In many cases it is what makes the person's action intelligible.
The distinction matters even when the prediction is accurate. A person may be statistically likely to default, reoffend, resign, or require care. The institution must still ask what it is entitled to do with that information, what burden it is placing on the person, and whether a proportionate path to correction exists. Prediction is not permission. Risk management is not a moral verdict.
How Accountability Gets Laundered
Responsibility tends to disappear through several familiar moves. First, an institution changes a normative question into a technical one. Instead of asking whether a family should lose assistance, it asks whether the model crossed its fraud threshold. The threshold appears objective because its value is numerical, although choosing it expressed a judgment about false positives, false negatives, and whose hardship counts.
Second, an organization treats procurement as delegation. It buys a system described as validated, compliant, or explainable and assumes that the vendor has assumed the ethical burden. But purchasing a tool does not transfer the duty to protect the people over whom the tool is used. A hospital remains answerable for how it incorporates a clinical prediction. An employer remains answerable for a hiring practice. A public agency remains answerable for the lawful and humane administration of a benefit.
Third, scale is used to excuse impersonality. A human reviewer may say that there are too many cases to hear individually, so the model is needed. There is truth in the resource constraint. But the answer to too many claims cannot be to make the affected person unable to be heard at all. If a system can produce thousands of adverse decisions, the institution must budget for proportionate review, not treat the number of decisions as proof that no one can be responsible for them.
Finally, the institution hides behind secrecy. Trade secrets, security concerns, privacy law, and technical complexity can all justify limiting disclosure in particular circumstances. None automatically justifies an unanswerable decision. A person can be told the material reasons for an adverse action without receiving every line of source code or another person's private data. Secrecy must be specific and proportionate, not a general license to deny explanation and appeal.
What Responsibility Requires
Responsibility is more than assigning blame after damage. It includes foresight, competence, supervision, attention, and repair. A director who deploys an unreliable system without testing is responsible for the foreseeable consequences. A professional who accepts an output without performing the review required by the role is not excused by the output's authoritative appearance. An operator who notices a systematic error and continues to invoke the system may acquire a different responsibility from the person who made the original design.
Catholic moral theology has long resisted the idea that an act becomes innocent merely because it was mediated by an institution. The Catechism's treatment of conscience, prudence, and responsibility assumes that persons must seek what is good and apply general principles to concrete circumstances (CCC 1786–1789, 1806). Prudence is not indecision. It is practical reason attentive to the real goods and harms at stake. A screen can support prudence; it cannot perform the whole act of prudential judgment while excusing the human being who relies on it.
This also explains why an audit log, although valuable, is not enough. A log can show which model version ran, which inputs were supplied, and which employee approved an action. That evidence makes accountability more possible, but it does not make the outcome just. Recordkeeping should support an intelligible answer to three questions: who had authority, what reasons were considered, and what remedy is available to the person affected?
The Strongest Objection: Humans Fail Too
The strongest objection is obvious. Human decision-makers are biased, inconsistent, exhausted, and sometimes corrupt. A model may reduce arbitrary favoritism, detect fraud that a reviewer would miss, and extend scarce expertise to people who would otherwise receive no answer. If a human appeal is required for every decision, will the result not be slower, more expensive, and merely another form of discretion?
This objection should be accepted where it is true. The alternative to automation is not a golden age of face-to-face justice. Human review can itself discriminate, and an institution should measure whether its safeguards actually improve outcomes. The answer, however, is not to replace visible human fallibility with invisible institutional fallibility. A model can standardize a process while making its errors harder to contest and its underlying values harder to see.
The right comparison is not “machine or human,” but which division of labor gives a person the best chance of a truthful, proportionate, and revisable decision. Machines are often excellent at searching records, detecting anomalies, translating documents, and applying clearly specified rules. Humans are indispensable when facts conflict, reasons are incomplete, exceptions matter, rights are at stake, or the institution must explain and repair an error. A human reviewer should not be a rubber stamp. Review must have authority, time, information, and independence enough to change the result.
Capacity Is Part of the Safeguard
A promise of answerability is empty if the institution has no capacity to deliver it. Review should therefore be tiered: reversible, low-severity actions can receive a fast correction channel; serious effects should trigger a pause or temporary relief where feasible; decisions affecting liberty, housing, health, or essential income require a reviewer with authority, time, and access to the relevant evidence. Small organizations may share an independent review service rather than pretend that the original operator is independent.
The organization should measure whether this division of labor works, not assume that a visible person is humane. Useful signals include response time, reversal rate, disparate delay, correction of bad input versus disagreement with the rule, repeat complaints, and the burden placed on claimants. A reviewer who cannot change an outcome is not oversight, and an appeal channel that resolves after the harm is irreversible may be only retrospective documentation. These tests preserve the essay's claim about answerability while conceding that unstructured human discretion can also be biased, fatigued, or corrupt.
Designing Against Moral Distance
A responsible organization should map an automated decision as a chain of accountable acts. Before deployment, it should name the decision owner, the affected population, the legitimate purpose, the foreseeable harms, and the conditions under which the system must not be used. It should test performance across relevant groups and real operating conditions, not only on a vendor's summary metric. It should establish a stop rule for material failures rather than treating continued operation as the default.
During use, the institution should preserve the input and model version relevant to each decision, train staff to recognize uncertainty, and prevent automation from quietly becoming a presumption of guilt. The affected person should receive timely notice, an intelligible statement of the main reasons, and a practical route to submit correcting evidence. When the stakes are high, the route should lead to a person who can exercise judgment rather than merely restate the system's output.
After an adverse outcome, review should ask more than whether the software followed its specification. It should ask whether the specification served a legitimate end, whether the evidence was reliable, whether the burden placed on the person was proportionate, and whether the institution owes restoration. A benefits error may require back payment and an apology, not only a corrected database. A wrongful disciplinary action may require clearing a record, not merely allowing the employee to apply again.
Keeping the Face in the Decision
The phrase “human in the loop” is too weak if it means that a person clicks approve after glancing at a recommendation. What matters is a human being with a role, a reason, and the power to say no. The person who reviews a case must be able to encounter the claimant as more than an exception code. That does not require sentimentality. It requires the disciplined recognition that justice is done to someone, not to a row in a table.
Subsidiarity offers a useful institutional test. Decisions should be made at the lowest competent level that can see the relevant context, while higher levels supply resources, standards, and correction. An automated platform should not pull every judgment upward into an opaque central model if a local professional can responsibly resolve the matter. Nor should local discretion become a license for unreviewed domination. The principle asks both for proximity and for structures that protect the common good.
Artificial intelligence need not make responsibility disappear. It can make responsibilities clearer if institutions disclose who uses it, preserve evidence of how it was used, and provide a real path from prediction to conversation and from error to repair. The goal is not to prohibit every automated aid. It is to refuse a system in which the person harmed by a decision can find only a machine, a contract, and a chain of employees who insist that no one decided.
The moral question is therefore not whether a model is autonomous enough to be blamed. It is whether the human community has arranged its work so that someone remains answerable. Catholic social thought begins with persons, relationships, and the common good rather than with the convenience of an apparatus. A humane use of AI will do the same. It will let computation widen human competence without allowing technical distance to become moral distance, or institutional complexity to become an excuse for abandoning the neighbor at the point of greatest vulnerability.
Sources and qualifications
The theological framework draws on Gaudium et Spes 16, 24, and 26 for conscience, dignity, and the common good; and on theCatechism of the Catholic Church 355, 1700, 1731, 1776–1789, 1806, and 1905–1912 for the human person, freedom, conscience, prudence, and social responsibility. The appeal to subsidiarity follows Quadragesimo Anno 79–80 and the Catechism's account of social life. These texts do not discuss contemporary machine-learning systems directly; their principles are applied by analogy.
The practical recommendations are consistent with the accountability, transparency, human oversight, and remedy principles found in the OECD AI Principles, the NIST AI Risk Management Framework, and contemporary data-protection and administrative-law practice. Those frameworks differ in legal force and scope. No audit, explanation, or human-review procedure guarantees justice by itself. The central claim is narrower: institutions remain morally and legally answerable for consequential uses of systems they select, deploy, supervise, and continue to operate.
Source Notes
A Catholic account of conscience, social responsibility, solidarity, and institutions ordered toward the dignity of the person.
The Church’s teaching on conscience, prudence, responsibility, subsidiarity, social life, and the common good.
A legal framework that makes organizational roles and lifecycle obligations visible rather than treating an AI system as an autonomous bearer of responsibility.
International guidance on human oversight, auditability, responsibility, and remedies for AI-assisted decisions.
A practical governance framework for assigning roles, monitoring risks, documenting decisions, and managing harms across the AI lifecycle.
Intergovernmental principles for accountable, transparent, robust, and human-centered AI.
Editorial revision note
This essay completed one self-editing, cross-essay consistency, and strongest-criticism pass. Its controlling thesis was preserved while objections, qualifications, evidence limits, and practical consequences were strengthened.