Public administration is built from general rules. A welfare office needs eligibility criteria, a tax authority needs classifications, and a court system needs procedures that do not change with every official who opens a file. Software seems to promise a cleaner version of that ambition: translate a rule into a model, apply it consistently, and remove the unpredictable human element.
Yet the human element is not only a source of bias. It is also the place where an institution can notice an exceptional fact, explain a decision, correct a record, and decline to turn a general rule into an absurd result. Algorithmic government becomes dangerous when it treats the exception as noise and mercy as a defect. The issue is not whether artificial intelligence is a person, nor merely that code cannot forgive. The issue is how public power should exercise discretion when a rule meets a life that the rule cannot fully describe.
Mercy in Christian teaching is not a synonym for administrative discretion. It is gratuitous compassion toward a person in misery and, ultimately, the free mercy of God. It is not earned by a successful application or dispensed as a private official's mood. In Misericordiae Vultus 20–21, Pope Francis presents mercy as neither a denial of justice nor a license to ignore truth: it expresses God's way of meeting human misery and brings justice to its fullness. Public administration cannot manufacture that divine gift. It can, however, make its lawful mechanisms capable of responding to human need in ways informed by mercy.
Rules Need Judgment
A rule is an abstraction. It identifies facts that normally matter and supplies a presumptive response. Real cases contain facts that are missing, ambiguous, unusual, or morally relevant but not represented in the form. A missed deadline may reflect carelessness, hospitalization, displacement, an inaccessible notice, or an agency error. A benefits record may say that income crossed a threshold while omitting that the money was a one-time payment needed to prevent eviction. A risk score may identify a pattern without showing whether the underlying data is accurate.
Discretion is the institutional capacity to ask what the rule is for, whether its conditions are satisfied, and what response is proportionate in this case. It is not the opposite of law. In mature administrative systems, discretion is bounded by statutory purpose, evidence, reasons, review, consistency, and non-discrimination. The challenge is to govern it rather than pretend it can be eliminated by putting a formula on a server.
Catholic social teaching supplies a moral vocabulary for this challenge. The Catechism calls prudence the virtue that disposes practical reason to discern the true good in particular circumstances and choose the right means (CCC 1806). Justice gives each person what is due, while solidarity refuses indifference to a neighbor's concrete condition (CCC 1807, 1939–1942). These principles do not require an official to waive every obligation. They require an institution to distinguish faithful application from a process that cannot see the person before it.
What Algorithmic Administration Changes
An algorithm can help officials sort applications, identify missing information, detect duplicate claims, or prioritize inspections. The administrative danger begins when a recommendation becomes a determination without an intelligible route to challenge it. A worker may technically click “approve,” yet lack time, authority, or incentives to depart from the score. In that situation, the system has transferred discretion from a visible official to an opaque rule, vendor, or training dataset.
Automation can also harden old judgments. Historical data records who was investigated, denied, sanctioned, or served under earlier practices; it does not simply record need. A model trained on those outcomes can reproduce unequal attention while appearing neutral. An accurate model may also optimize a proxy—payment probability, predicted fraud, or expected attendance—that is not the program's legal purpose. The question is whether the institution pursues a lawful objective by relevant means.
Public administration has a special duty here because a citizen cannot always take her business elsewhere. A person denied a license, benefit, permit, or release is not a consumer choosing among competing services. The state exercises authority backed by law. Administrative law traditions therefore emphasize lawful authority, fair procedure, relevant evidence, reasons, impartiality, and review. The doctrines differ across jurisdictions, but the demand is recognizable: power must be answerable to the person it affects.
Prudence, Equity, and Mercy
Several administrative ideas must be kept distinct. Prudence is a virtue of practical reason: the Catechism describes it as discerning the true good in particular circumstances and choosing the right means (CCC 1806). Proportionality is a public-law and policy test asking whether a measure is suitable, necessary, and not excessive in relation to a legitimate aim; its formulation varies by jurisdiction. Equity is a legal idea, present in different forms across legal systems, that can attend to relevant particulars when a general rule would otherwise produce an unjust fit. Lawful discretion is authority delegated by law within defined limits. Clemency is typically a competent authority's remission or reduction of a penalty, usually after a finding of responsibility and according to the legal system's rules. None of these concepts is identical to Christian mercy.
These mechanisms can be informed by mercy without being renamed mercy. Prudence may lead an official to seek the relevant facts. Equity may support a lawful exception or a cure period. Proportionality may rule out a sanction that is needlessly destructive. Lawful discretion may authorize a different remedy, and clemency may remit a penalty after responsibility has been established. Each still requires a source of authority, fair criteria, reasons, and review. Mercy supplies a disposition toward the suffering person; it does not suspend the institution's duties to truth, equal protection, accountability, or legitimate public order.
Consider a housing benefit with a reporting deadline. A system that automatically terminates assistance after one missed form may be easy to administer, but it can create homelessness for a person who was hospitalized or never received an accessible notice. Prudence asks what happened and what the program is for. Equity or a lawful discretion may support a grace period, a chance to cure, or temporary continuation during review. Proportionality tests whether termination is excessive, while an authorized clemency power would concern a penalty rather than eligibility itself. A response informed by mercy takes the person's misery seriously; it does not pretend that the deadline, the evidence, or the public purpose does not exist.
Reasons, Appeal, and the Exceptional Case
A person cannot challenge a decision that has no stated reason. “The model found you ineligible” is a result, not an explanation. A usable reason identifies the governing rule, the material facts, the important uncertainty, and the evidence that would change the outcome. It need not reveal trade secrets or dump technical details on a claimant. It must make the decision intelligible enough for a person and a reviewer to contest its premises.
Appeal must include a real possibility of correction. A reviewer should be able to fix inaccurate data, consider lawful evidence omitted by the system, suspend immediate harm where appropriate, and depart from a recommendation. The reviewer should know whether the issue concerns an input, the model, the policy, or the fit between the policy and the circumstances. Sending the same facts through the same automated pipeline is not an appeal.
Exceptional cases need not receive a completely unstructured hearing. Institutions can define triggers for human escalation: a severe adverse effect, a disputed identity, a disability accommodation, a credible allegation of agency error, an unusually high confidence interval, or evidence that the case falls outside the data used to design the system. A person should also be able to request review in plain language, since those harmed by rigid administration may be least able to name its technical failure.
Strong Objection: Discretion Is Biased
The strongest objection is that human discretion has a terrible history. Officials may favor people who resemble them, punish those who complain, demand deference from the poor, or treat mercy as a reward for confidence and fluent speech. A formula at least appears to apply the same condition to everyone. Reintroducing judgment, the objection says, reintroduces arbitrary power.
This objection is correct about the risk and mistaken about the remedy. Removing visible discretion does not remove judgment; it relocates judgment into data selection, target definitions, thresholds, missing-data rules, vendor contracts, and the official who decides whether to follow the output. A rigid system can be discriminatory while making its discrimination harder to contest. The appropriate response is accountable discretion: written criteria, structured reasons, training, conflict rules, demographic and outcome monitoring, supervisory review, records of departures, and an external complaint route.
Institutions should examine both false positives and false negatives. A model that looks fair among approved cases may deny help to people who never make it through the first screen. Appeal outcomes should be analyzed for patterns, and repeated reversals should trigger policy or system change rather than being treated as isolated worker error. Discretion should be visible enough to learn from and constrained enough to avoid becoming a private favor system.
Strong Objection: Consistency Is Justice
Another objection says that equality demands identical treatment. If one applicant receives an extension or a lower sanction, others may reasonably ask why they did not. A system full of exceptions can be slow, expensive, and impossible to predict. People need rules they can plan around, not an official's compassion at the end of a queue.
Consistency matters, but there are two kinds. Mechanical consistency gives the same output to the same coded inputs. Just consistency gives like cases like treatment while recognizing relevant differences. Public guidance can make the second form more reliable by describing the factors that justify relief, publishing aggregate decisions, setting service deadlines, and allowing a person to invoke prior precedents. An exception that is available to everyone who meets its conditions is not favoritism; it is a lawful part of the rule's structure.
Predictability should also include the prediction that a person will be heard when the stakes are high. A fast automated denial may be procedurally consistent and substantively reckless. A slower system with notice, reasons, interim protection, and appeal can be more legitimate because it recognizes error and circumstance.
Discretion as a Reviewable Entitlement
“Human review” is too thin a safeguard if a claimant must persuade an overworked caseworker from scratch. An exception should be a reviewable entitlement when published conditions are met: accessible exception triggers, a plain-language request, protected interim benefits where delay threatens subsistence, reasons for granting or refusing relief, and a reviewer with authority independent of the original recommendation. Random or rotating assignment, conflict rules, and records of departures can reduce the chance that fluent or well-connected applicants receive all the mercy.
Agencies should compare structured human review with the automated and ordinary human baselines rather than assuming either is humane. They can examine reversal rates, disparate delays, false positives and negatives, claimant burden, and whether reviewers actually change outcomes. A categorical rule may be safer when the interest is routine, the facts are stable, and exceptions would invite coercion or favoritism; a grace period or individualized review is more defensible when a rigid result risks homelessness, loss of care, or an uncorrectable error. Existing administrative and disability-access principles support these questions, but the evidence of improvement must come from the particular program or pilot.
Practical Policy and System Design
Before deployment, an agency should identify the legal authority, purpose, affected population, decision owner, data sources, and foreseeable harms. It should decide whether software is appropriate for the task at all. Sorting routine documents may be low risk; determining liberty, family separation, essential income, immigration status, or medical access requires a higher threshold and often a presumption against unreviewed automation. A vendor's proprietary status cannot relieve the agency of responsibility.
The system should preserve a case record: the applicable policy and version, relevant inputs, model output, confidence or uncertainty, human actions, notice sent, evidence received, and final reasons. Privacy and security controls are essential, but secrecy must not be used to defeat a person's legal remedy. Contracts should require audit access, incident reporting, data deletion, independent testing, and cooperation with appeals. Agencies should retain enough institutional knowledge to change providers or operate without a vendor if the service fails.
At the front door, notices should state that automation materially contributed, identify the responsible authority, explain the principal grounds, give a deadline, and offer accessible help. At the review door, the institution should provide a human with authority to pause, reverse, or modify the result. Emergency action may sometimes occur first, but it should receive prompt review and a remedy if the emergency premise was wrong. Children, disabled people, people with limited language access, and people without reliable connectivity need channels designed around their actual circumstances.
Governance should include affected communities, frontline workers, advocates, and independent experts before and after launch. Impact assessments should be living documents, not a one-time compliance form. Legislatures and oversight bodies should receive aggregate information about use, error, appeals, reversals, disparities, procurement, and incidents. Sunsetting or narrowing a system after evidence of harm is administrative competence, not technological failure.
Sources and qualifications
The account of dignity, prudence, solidarity, justice, and the common good draws on Gaudium et Spes 16, 24–26; Quadragesimo Anno 79–80; Pacem in Terris 63–67; and the Catechism of the Catholic Church 1731–1738, 1776–1782, 1806–1807, 1883–1885, and 1939–1942. Catholic social teaching does not prescribe a model, database, or appeals office. It supplies criteria for judging respect for persons, participation, responsibility, and the common good.
The theological account of mercy is deliberately distinct from these administrative mechanisms. Misericordiae Vultus 2, 20–21 presents mercy as central to the Christian revelation of God's love and explains that mercy is not opposed to justice but brings justice to its fullness; Dives in Misericordia 4 and 14 likewise treats mercy as God's compassionate love and relates it to human dignity. These are theological claims about God's mercy and discipleship, not instructions to officials to waive public rules. In administration, prudence, equity, proportionality, lawful discretion, and clemency remain distinct instruments that may be guided by merciful concern without abolishing truth, responsibility, or legitimate public order.
The administrative account reflects general principles found in constitutional and administrative-law traditions, including lawful authority, notice, reasons, impartial review, proportionality, and access to remedy. Their content varies by jurisdiction. The International Covenant on Civil and Political Rights, the Convention on the Rights of Persons with Disabilities, OECD AI Principles, NIST AI Risk Management Framework, and the Council of Europe Framework Convention on Artificial Intelligence offer relevant reference points, but none automatically governs every agency. This essay is normative and is not legal advice.
A Government That Can See the Person
Mercy is not an administrative synonym for equity, proportionality, discretion, or clemency. Christian mercy is gratuitous compassion that does not erase the truth of an act, the responsibility of an agent, or the legitimate order of a community. It perfects justice by refusing to treat a suffering person as nothing more than the fault, debt, or classification before the institution. Public servants cannot confer God's mercy by clicking an exception, but they can build lawful procedures that do not make compassion impossible.
Algorithms can assist public servants, but they must not become an unanswerable substitute for public judgment. A legitimate government gives reasons, hears objections, preserves an avenue for exceptional facts, and allows an authorized person to correct the result. When administration retains those capacities, rules can serve the common good rather than trap everyone whose life does not fit the training data. The point is to distinguish legal mechanisms from the Christian mercy that can animate their exercise, so public authority remains truthful, responsible, and humane.
Source Notes
The Church’s teaching on prudence, justice, solidarity, conscience, legitimate authority, and the common good.
A U.S. risk-management framework for governing, mapping, measuring, and managing AI risks, including accountable oversight.
International principles for human-centred AI, transparency, explainability, robustness, and accountability.
The international disability-rights framework for equality, accessibility, participation, and non-discrimination in public systems.
Pope Francis’s account of mercy as the fullness of justice, which does not abolish justice but brings it to completion (§§2, 20–21).
The Catholic account of mercy as a personal and social reality that complements justice without abandoning truth, responsibility, or the common good.
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.