AI and the Poor

What the preferential option actually requires

Begin with the principle, because the principle is not always understood even by Catholics. The preferential option for the poor is not a sentimental commitment to caring about poor people. It is a substantive claim about how moral evaluation is supposed to work.

The principle holds that any social, economic, or political arrangement is to be judged primarily by what it does to the poor and vulnerable. Not on average. Not at the median. Specifically at the bottom. A policy that benefits most people while severely harming the most marginalized is, in Catholic teaching, a policy that fails. A system that produces good outcomes for the well-resourced and catastrophic outcomes for the poor is a system that has been designed to do exactly that, whether or not its designers intended it.

This framing is not abstract. The Latin American bishops articulated it at Medellin in 1968 and Puebla in 1979. Pope John Paul II made it central to Sollicitudo Rei Socialis in 1987, calling it a Christian duty that admits of no exceptions. Pope Benedict XVI reaffirmed it in Caritas in Veritate. Pope Francis returned to it constantly. The Catechism states it directly. It is part of the irreducible core of Catholic social teaching, and it is one of the few principles in that tradition that applies universally to every social question without modification.

What it requires in practice is twofold. First, that the effects of any arrangement on the poor be made visible and measurable, rather than averaged away. Second, that when those effects are bad, the arrangement itself is morally suspect, regardless of how well it works for others.

Applied to AI, this is a demanding standard. It is also exactly the standard that is missing from most policy discussion about AI fairness, which tends to settle for statistical parity across groups rather than asking the harder question of whether the system should exist at all.

What AI is actually doing to the poor

Five patterns recur across the documented cases. None of them is hypothetical.

Algorithmic denial of public benefits. Welfare, unemployment, food assistance, and disability programs across the United States, the United Kingdom, the Netherlands, and Australia have been automated in ways that produce systematic wrongful denials. The Michigan unemployment fraud-detection system, the Dutch childcare benefits scandal, and the Australian Robodebt program are the best-known cases. In each, an algorithm was deployed against welfare claimants, accused tens or hundreds of thousands of fraud or overpayment based on erroneous logic, suspended benefits before the claimant could respond, and forced an appeals process that the affected people were poorly equipped to navigate. The harms were measured in suicides, foreclosures, deportations, and lives derailed.

Predictive policing. Tools that predict where crime will occur or which individuals are likely to offend have been deployed in police departments across the United States and Europe since the mid-2010s. The systems are trained on historical arrest data, which reflects historical policing patterns more than it reflects underlying crime. The result is that the same neighborhoods that were over-policed in the past are flagged by the algorithm as high-priority for the future, the same individuals appear repeatedly on lists they cannot see and cannot challenge, and the people most subject to police attention are the least able to do anything about it.

Algorithmic immigration enforcement. Risk-assessment tools at borders, in asylum processing, and in detention decisions reduce human beings to scores. The scores draw on data the affected person cannot inspect and weights the affected person cannot challenge. People are deported, detained, or denied entry based on these scores. The category of people most likely to be subjected to them is also the category least likely to have legal representation or to speak the language in which the system was built.

Tenant screening and credit scoring. Private AI services aggregate court records, eviction histories, and consumer data to decide who can rent which apartments. The systems are riddled with errors that disproportionately affect low-income and minority renters. The same is true of the algorithmic credit-scoring tools that decide who can access loans, what interest rates they pay, and what insurance premiums they face. Errors are systematic and biased; the cost of correcting them falls on the people least able to pay.

Algorithmic scheduling and worker surveillance. Hourly workers at large retailers, restaurants, and warehouses now have their schedules set by algorithms that optimize for the employer's labor costs at the expense of the worker's stability. The same systems surveil productivity in ways that intensify the pace of work and penalize workers who fall behind. The workers affected are disproportionately low-wage, often immigrants, often supporting families on incomes that have not kept pace with inflation. They cannot easily quit. They cannot easily contest. The algorithms operate on them, not for them.

The common structure across all five patterns is the same: an algorithmic system is deployed in a context where the affected person has limited power, limited information, and limited recourse. The system shifts decision-making away from accountable human actors and toward instruments the affected person cannot see. The errors that result fall hardest on the people least equipped to absorb them. And the institutional incentives are structured so that the people designing and deploying the systems do not directly experience the harms they cause.

This is not a glitch. It is what these systems are for.

What the Vatican has already said about this

Catholic engagement with algorithmic harm to the poor did not begin in 2025. Pope Francis spoke repeatedly about what he called the throwaway culture, the economy of exclusion, and the technocratic paradigm. Laudato Si in 2015 connected environmental degradation and the exploitation of the poor to a common logic of treating both the earth and human beings as instruments. Fratelli Tutti in 2020 extended the analysis to algorithmic systems explicitly, warning that digital tools were being used to fragment human solidarity and surveil populations in ways that diminished freedom.

The most direct Vatican engagement to date is Antiqua et Nova, the January 2025 doctrinal note from the Dicastery for the Doctrine of the Faith. The note addressed AI's effects on the poor across several sections, warning that AI is being deployed in ways that increase inequality, that the benefits of AI accrue disproportionately to those who already have power, and that the harms accrue disproportionately to those who do not. The note called on public authorities to ensure that AI development serves the common good rather than narrow private interests, and that the most vulnerable have meaningful protection from systems they cannot control.

Magnifica Humanitas, Pope Leo XIV's first encyclical, releases May 25, 2026. The encyclical is expected to develop these themes with the magisterial weight that Antiqua et Nova as a doctrinal note could not carry. Pope Leo XIV chose the name Leo XIV in conscious continuity with Leo XIII, whose 1891 encyclical Rerum Novarum addressed the industrial economy's effects on workers. The new encyclical is expected to be the AI-era counterpart, with the preferential option for the poor as one of its anchors.

Why algorithmic harm is structurally different

It is worth pausing on what makes AI-mediated harm to the poor different from older forms of structural injustice, because the difference is part of what the Catholic framework has to grapple with.

Older forms of discrimination, even structural ones, typically involved identifiable human actors making identifiable decisions. A landlord refused to rent to a tenant. A police officer stopped a driver. A loan officer denied an application. The harm was real and the chain of responsibility, while often obscured, ran through human persons who could in principle be held accountable.

Algorithmic systems break that chain. The landlord did not refuse the tenant; the screening service flagged them. The officer did not target the driver; the predictive system identified the area as high-priority. The loan officer did not deny the application; the model returned a score below the threshold. At every link in the chain, the question of who is responsible becomes harder to answer. The data scientist who built the model can point to the product manager who specified it. The product manager can point to the regulators who approved it. The regulators can point to the lack of legislation governing this specific use case. The legislators can point to the lack of consensus among experts. By the time the harmed person tries to find someone accountable, the chain has dissolved into a fog of distributed responsibility in which no single actor sees themselves as the cause.

This is one reason the Catholic framework is unusually well-suited to this moment. The Catholic tradition has thought carefully about structural sin, about social structures that produce harm even when no individual within them is acting with malice, and about the moral responsibility that attaches to participation in such structures. The framework was developed in the context of unjust labor systems, racial segregation, and exploitative international economic arrangements. It applies with full force to algorithmic systems that distribute responsibility in ways that protect everyone except the people they hurt.

The Catholic move is to refuse the dissolution. Even when the chain of responsibility is hard to trace, the harm is real, and someone has to answer for it. The institutions deploying the system carry responsibility. The systems themselves carry the imprint of moral choices that were made by the people who built them. The patterns of harm are not natural facts; they are the predictable result of decisions made by identifiable institutional actors. The work of moral accountability is to refuse to let those decisions disappear into the technical complexity that produced them.

What Catholic teaching actually requires

Applied to AI and the poor, the preferential option for the poor produces a set of demands that are more specific than general gestures toward fairness.

Visibility of impact on the most affected. Any AI system deployed in a context affecting the poor should be evaluated on what it does to the worst-off people who interact with it, not on aggregate performance metrics. False denial rates for welfare claimants, error rates for low-income tenant applicants, false-positive rates in predictive policing in specific neighborhoods. The numbers that matter are the ones at the tails of the distribution, not the averages.

Meaningful human oversight and appeal. Decisions that materially affect a person's access to housing, benefits, employment, or liberty cannot be made by an algorithm acting alone. There must be a human in the decision loop, an accessible appeal process, and a real possibility that the human will override the algorithm. Cosmetic human oversight, where a human signs off on the algorithmic decision without actually reviewing it, does not satisfy this requirement.

Refusal of systems that cannot be made just. Some AI systems are not reformable. A predictive-policing system trained on historical arrest data will always reproduce historical patterns of over-policing, no matter how it is tuned. Catholic teaching does not require infinite reform of such systems. It permits, and sometimes requires, the conclusion that the system itself should not be deployed.

Structural accountability. The institutions deploying AI systems against the poor must bear the costs of the systems' failures, not the affected individuals. If a welfare algorithm wrongly denies benefits, the burden of correction should fall on the agency that deployed it, not on the claimant. If a tenant-screening service errs, the cost of remediation should fall on the service, not on the renter. The current arrangement, in which the affected person bears the cost of contesting the algorithm, is morally unsustainable.

Solidarity, not pity. The Catholic framework is not satisfied with charity in response to algorithmic harm. The principle of solidarity requires that those harmed by these systems be treated as agents with standing to demand reform, not as objects of compassion. The voices that should be loudest in the policy debates about algorithmic accountability are the voices of the people the algorithms are operating on. Catholic institutions can amplify those voices. They cannot substitute for them.

What this asks of Catholics

The page closes with what it asks of readers, because the preferential option for the poor is not a contemplative principle. It demands action.

For Catholic readers who work in technology, the question is whether the systems you are building serve the poor or extract from them. Most Catholic engineers will not personally build welfare algorithms. Many will build the foundational systems on which such algorithms are constructed: large language models, computer vision systems, cloud platforms, data infrastructure. The argument that the technology is neutral and only the applications carry moral weight is not, in Catholic teaching, an argument that survives serious scrutiny. The systems you build are deployed in particular ways for particular ends, and the responsibility for those deployments runs back through the technology to the people who built it.

For Catholics in policy, civil society, and journalism, the question is whether the regulatory debates about AI in your jurisdiction adequately address the systems being deployed against the poor. The EU AI Act, the U.S. state-level laws being drafted, and the emerging international frameworks all need pressure from voices that take the preferential option seriously. Catholic institutions have unusual standing in these debates, and have not yet used it as fully as they could.

For Catholics in Catholic institutions, the question is whether the AI tools your institution is deploying conform to the standards Catholic teaching demands. Catholic Charities, Catholic schools, Catholic hospitals, dioceses, and religious orders are themselves now deploying AI systems for hiring, fundraising, beneficiary screening, and administrative work. The same standards that the Catholic framework demands of public agencies should apply to Catholic institutions deploying analogous tools. Anything less is hypocrisy.

For ordinary Catholic readers, the question is harder and more diffuse. Most Catholics will not personally deploy AI systems against the poor. Most Catholics will, however, vote on policies that constrain such systems, support institutions that deploy them, and live in societies where these systems are increasingly determining the life chances of millions of people. The minimum that Catholic teaching asks is that Catholics not pretend the systems are neutral, not accept the convenient framing that algorithmic harms are nobody's fault, and not look away from the documented patterns of who pays the cost of automation.

The preferential option for the poor was articulated for a different technological era, but it was articulated as a principle that applies to every era. The current technological era is producing exactly the kind of harm the principle was developed to identify. The question is whether Catholic readers will use the framework, or whether the framework will sit unused in the tradition while the harm it was built to address continues to compound.