Every consequential automated decision should leave a path to a human being who can understand the case, hear a challenge, and change the result. This is not a demand that a person supervise every spam filter or approve every low-risk recommendation. It is a positive design principle for decisions that can alter a person's liberty, livelihood, housing, education, health, family life, or access to essential services.
Call it the Human Appeal Principle. Its claim is simple but stronger than the familiar promise of a “human in the loop”: when an automated system makes or materially determines a consequential decision, the person affected must have a timely, intelligible, and accessible appeal to an accountable human reviewer with authority to suspend, reverse, or correct the decision. The reviewer must consider the person's evidence and circumstances rather than merely confirm that the system ran as designed.
This principle is not a magical cure for bias, and it is not a claim that every disagreement should prevail. It is a boundary against unanswerable power. A prediction may begin an inquiry. It must not be the final court of the person whose life it affects.
Why Appeal, Not Just Explanation
Public discussion often treats transparency as the main safeguard. People are promised an explanation of the model, a list of features, or a notice that automation was involved. Such information can be useful, but explanation by itself does not give a person a remedy. A claimant may understand that a system found a high risk while having no way to correct the address, employment history, medical circumstance, or mistaken identity that produced the risk.
An appeal is different because it creates a relationship of answerability. It says that the institution does not regard its output as more authoritative than the person who must live with it. It turns a one-way classification into a contestable process. The value is procedural and moral before it is technical: a person is recognized as someone who may speak, give reasons, present evidence, and be heard by another person with the power to respond.
Catholic social thought supplies the anthropology behind this requirement. Gaudium et Spes presents the human person as possessing dignity and as called to participate responsibly in social life (GS 24, 26). The Catechism describes conscience as a judgment in which the person is involved and insists that dignity includes the exercise of freedom and responsibility (CCC 1731, 1776–1782). A consequential institution therefore cannot treat the person only as an object of prediction. It must make room for the person's reasoned participation.
A Precise Rule
The principle needs a threshold more useful than “important decisions.” A decision is consequential when it can reasonably be expected to produce a significant adverse effect on a person's rights, material security, bodily integrity, civic status, or access to a relationship or institution essential to ordinary participation in society. Denial of a loan may be consequential in one context and trivial in another; an automated recommendation for a cheaper hotel room usually is not. The assessment should consider severity, duration, reversibility, vulnerability, and the number of people affected.
The rule is triggered not only when software takes the final click. If a human official is required by policy to accept a model's recommendation except in extraordinary circumstances, the system has materially determined the decision. Calling the final step “human” does not evade the safeguard. Nor should a provider escape the rule by splitting one action into a sequence of apparently minor scores whose combined effect is severe.
For every covered decision, an institution should provide six guarantees. The first is notice: the person must know that an automated system materially contributed to the outcome and which institution is responsible for acting on the appeal. The second is intelligible reasons: the notice should identify the principal facts, rules, and uncertainties that mattered, in language appropriate to the person and the decision.
The third is a pause when immediate enforcement would cause serious and avoidable harm, unless a narrowly defined emergency requires interim action. The fourth is an accessible route to appeal, available without specialized legal or technical knowledge and without fees that make review theoretical. The fifth is a genuinely empowered reviewer. The sixth is a recorded resolution that states what evidence was considered, what was decided, why, and what further remedy remains.
What Meaningful Review Looks Like
“Human review” has become an empty phrase in systems where a worker receives a score, sees a recommended outcome, and is measured for agreement with the machine. Meaningful review requires a different institutional arrangement. The reviewer must have enough time and training to understand the relevant facts, access to the inputs and policy governing the result, and explicit permission to depart from the recommendation without retaliation for exercising judgment.
The reviewer must also be able to distinguish three questions that automated systems often collapse: whether the input is accurate, whether the rule was applied as specified, and whether applying that rule in this circumstance is justified. A person might lose assistance because the database contains an old address, because the software misread a document, or because the rule itself fails to account for a lawful exception. Each requires a different response. Re-running the same model cannot solve all three.
Review should be independent enough to be credible. The original decision-maker may be well placed to correct a clerical error, but an appeal cannot be meaningful if the only reviewer is rewarded for preserving the original result or lacks authority over the department that caused the harm. Independence need not mean a distant tribunal in every case. It can mean separation of duties, escalation rights, supervision by a responsible officer, or an external route for serious disputes.
Accessibility is part of justice, not a customer-service extra. An appeal channel should accommodate disability, language, limited connectivity, and the fact that a person under financial or emotional pressure may not be able to compose a perfect technical objection. Institutions should accept ordinary explanations and help identify what evidence is relevant. A process that technically accepts appeals but requires a person to know the model's vocabulary has moved the burden of institutional competence onto the vulnerable.
The Standard for a Decision
A reviewer needs a standard, not merely a power to be sympathetic. The decision should be tested for lawful purpose, evidentiary reliability, proportionality, and consistency with the person's rights and the institution's stated obligations. The reviewer should ask whether the result is supported by relevant information, whether contrary evidence was ignored, whether the response is excessive in relation to the risk, and whether a less harmful alternative can protect the legitimate interest.
These questions do not abolish rules. They prevent rules from becoming a substitute for judgment where circumstances matter. Catholic teaching calls prudence the virtue that disposes practical reason to discern the true good in particular circumstances and choose the right means (CCC 1806). An appeal process institutionalizes a modest form of that work. It acknowledges that a general policy can be necessary while its application still requires attention to reality.
Objection: The Cost Will Be Impossible
The strongest objection is practical. Large institutions make millions of decisions, and a human appeal for each one could be ruinously expensive. Delays could harm people who need a quick answer. Smaller organizations may not be able to staff a specialized review office. If the principle is absolute, the objection says, it will either be ignored or produce a formal process too thin to matter.
The answer is to make the principle risk-proportionate without making it optional. Not every covered decision requires a hearing before temporary action. A low-severity decision can have a fast, simple correction channel; a decision affecting liberty, housing, medical treatment, or essential income requires stronger pause and review protections. Institutions can use triage, shared review services, ombuds offices, and random audits of apparently favorable decisions. They can also reserve full automation for tasks whose errors are readily reversible and do not impose serious burdens.
Cost should be counted honestly. If a system saves money by shifting the expense of error onto people who cannot buy counsel, the system is not efficient in any morally meaningful sense. Catholic social teaching asks whether social arrangements serve the common good, understood as conditions that allow persons and groups to reach fulfillment more fully and easily (Gaudium et Spes 26). A lower administrative budget purchased through unremedied hardship is not a neutral gain.
Objection: Appeals Will Be Abused
A second objection is that determined applicants will flood the system with repetitive claims, manufacture sympathetic stories, or use appeals to defeat rules designed to protect the public. This risk is real. A humane procedure still needs evidence standards, deadlines, anti-fraud measures, and the ability to deny a claim after genuine consideration.
But the possibility of abuse does not justify eliminating the innocent person's avenue of correction. Institutions already distinguish good-faith disputes from fraud in courts, schools, hospitals, and benefit systems. They can do so here as well. A reviewer can require relevant evidence, limit duplicative submissions, and refer suspected fraud for separate investigation without treating every claimant as an adversary. The appeal itself should not become a new automated suspicion score that silently closes the door.
There is also a symmetry often missed in this objection. Automated systems can make institutional abuse scalable: a flawed rule can deny thousands of people before anyone notices. A review process gives the institution evidence about its own failure modes. Appeals are not only individual escape hatches; they are sensors for systemic injustice. Patterns in overturned cases should trigger model revision, policy review, and, where necessary, restitution to others who were affected.
Capacity, Emergency, and Evidence
Risk proportion is not only a slogan. A low-severity, reversible classification can receive a correction channel with a published response time. A temporary action needed for an urgent safety concern may proceed, but it should be time-limited, recorded, and followed by review as soon as practicable. Liberty, housing, health, and essential income decisions need independent review with authority to pause, reverse, and arrange interim relief. A small institution may contract for shared review or escalation; it may not satisfy the principle by appointing a reviewer who has no power to disagree.
Institutions should publish appeal volume, time to resolution, reversal rate, repeat-abuse findings, and the proportion of outcomes changed because of bad input, incorrect rule application, or a justified exception. Those measures will not prove justice, but they can reveal a rubber stamp, disparate delay, or a process whose burden falls on the claimant. They also answer the objection that appeals undermine finality: deadlines and evidence standards can limit duplicative submissions while leaving a genuine route to correction.
From Principle to Institutional Practice
Before deployment, an institution should publish an inventory of systems that can trigger consequential decisions. For each system it should identify the decision owner, the affected population, the permitted purpose, the data sources, the expected error costs, and the appeal route. It should test whether a person can actually correct a mistaken record and whether an appeal can change the outcome in the software and in the underlying policy.
At the moment of decision, the notice should state what happened, when it takes effect, how long a person has to appeal, what temporary relief is available, and how to reach help. Reasons should be case-specific enough to guide a response. “The system found you ineligible” is not a reason; it is a restatement of the result. “The record used an income figure from this source, and the threshold applied was this” gives a person somewhere to begin.
At review, the institution should preserve a record of the original inputs, model version, human actions, evidence submitted, and final reasons, subject to privacy safeguards. The record should not be treated as a defense merely because the software performed exactly as configured. Compliance with an implementation is one question; justification of the decision is another.
A Door That Remains Human
No procedure can guarantee that every human reviewer will be wise, merciful, or even fair. Yet the absence of a reviewer guarantees something worse: a person will face a power that cannot listen. The Human Appeal Principle keeps open a door through which facts, conscience, and responsibility can enter an automated institution.
The point is not to preserve human control as a matter of nostalgia. It is to preserve the conditions of justice. A person who can appeal is not reduced to a prediction. An institution that must give reasons and provide remedy cannot pretend that its objectives are natural facts. A reviewer with authority to change the answer makes it possible for rules to remain servants of the common good rather than masters of the people they were meant to serve.
Artificial intelligence can help institutions see patterns, allocate attention, and perform routine work. It should not close the last path by which a neighbor can say, “That account is wrong; this is what happened; please look again.” In consequential matters, looking again is not an indulgence granted to the unusually persuasive. It is a minimum recognition that human beings live in circumstances no model completely contains, and that authority remains accountable to the persons over whom it is exercised.
Sources and qualifications
The account of dignity, participation, and the common good draws onGaudium et Spes 16, 24, and 26; Dignitatis Humanae3 for responsible freedom; and the Catechism of the Catholic Church 1731, 1776–1789, 1806, and 1905–1912. The account of subsidiarity is informed by Quadragesimo Anno 79–80. These sources do not prescribe a particular software architecture. They provide principles for judging whether an institution respects persons, responsibility, and the common good.
The practical framework also reflects due-process and administrative justice traditions, the OECD AI Principles, the NIST AI Risk Management Framework, and human-oversight and remedy provisions found in contemporary AI regulation. Legal requirements vary by jurisdiction, and a generic appeal process cannot replace professional advice in a regulated domain. The proposal is a normative minimum: consequential automated decisions must remain understandable, contestable, and answerable to an empowered human being.
Source Notes
The European data-protection framework, including safeguards associated with certain decisions based solely on automated processing.
The European Union’s rules for high-risk AI systems, human oversight, transparency, documentation, and risk management.
The council’s account of human dignity, conscience, participation, and social institutions serving rather than absorbing the person.
The council’s account of responsible freedom, conscience, and the duty of institutions to respect the person’s search for truth.
The Church’s teaching on conscience, prudence, participation, subsidiarity, authority, and the common good.
A practical framework for human oversight, risk documentation, accountability, and ongoing evaluation of AI systems.
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.