When Does AI Help Social Workers—and When Does It Automate Away Their Ethics?

Artificial intelligence is quietly reshaping the machinery of the welfare state. Predictive risk models flag children who might come to harm, large language models draft case notes, algorithmic systems process benefit claims, and digital-care devices monitor elderly people in their homes. A new open-access perspective published in Discover Social Science and Health argues that this wave of deployment has outpaced the ethical vocabulary social work uses to evaluate it. Kwangseon Hwang of the Department of Public Management and Policy at Gachon University in South Korea proposes a structured framework for deciding when a given AI system functions as augmented intelligence—technology that extends a practitioner’s relational judgment—and when it slides into automated welfare, a mode of governance that entrenches exclusion and erodes moral responsibility. The paper reframes a debate that has often been polarized between techno-optimism and outright resistance into a more precise question: under what conditions, if any, can specific forms of AI be ethically defensible in social work?

The core of the framework is a tri-lens analytical matrix that crosses three dimensions. The first dimension consists of three moral traditions: utilitarian ethics, which weighs aggregate outcomes and costs; deontological ethics, which focuses on duties, rights, and rules that must not be violated regardless of consequences; and virtue ethics, which asks what a good practitioner with good character would do in a given situation. The second dimension distinguishes AI’s two operational arenas. Frontstage systems are those clients directly encounter—chatbots, screening tools that shape decisions about their families, monitoring devices in their homes. Backstage systems are the algorithmic administration that clients rarely see: automated eligibility checks, data-matching pipelines, and risk-scoring infrastructures that quietly sort populations. The third dimension consists of four constitutive social-work values: privacy, self-determination, social justice, and human dignity. Every AI deployment in social welfare, the paper contends, can be located within this twenty-four-cell matrix, and each cell generates distinct ethical questions that generic AI-principle frameworks tend to blur.

This specificity matters because social work is not a generic domain for AI ethics. The discipline is anchored in relational practice—the idea that change happens through sustained, trusting relationships between practitioners and the people they serve—alongside care ethics and anti-oppressive praxis. A framework borrowed wholesale from medicine or finance risks missing what is distinctive about welfare encounters: they involve people in situations of dependency, often coerced or involuntary, where the state itself is both the service provider and the surveillance authority. Hwang argues that existing scholarship documents many harms of welfare AI—bias, opacity, surveillance, and the erosion of professional discretion—but remains conceptually fragmented, producing case-by-case critiques without a shared diagnostic language. The matrix is an attempt to supply that language, and to make explicit that different AI modalities pose different risks requiring different safeguards.

To demonstrate the framework’s reach, the paper applies it to three purposively selected cases whose heterogeneity spans continents and technologies. The first is the Allegheny Family Screening Tool in the United States, a predictive risk model used in child welfare screening in Allegheny County, Pennsylvania. The second cluster concerns algorithmic welfare administration: the Dutch SyRI system, which mined government data to detect suspected welfare fraud until a court curtailed it; the Australian Robodebt scheme, which automated debt recovery against benefit recipients and became one of the country’s worst public-administration scandals; and South Korean crisis-household detection, which uses administrative data to identify families at risk. The third case examines digital-care technologies in Korea and OECD comparator countries, where sensors, wearables, and remote-monitoring platforms mediate care for older adults. Together these cases show that a child-welfare risk score, a fraud-detection pipeline, and a care robot are not ethically interchangeable, even though all are labeled AI.

The frontstage–backstage distinction does much of the analytical work here. A frontstage tool like the Allegheny screening model directly shapes what a caseworker sees and does at the moment of first contact with a family. Its errors are consequential and intimate: a falsely high risk score can trigger intrusive investigation, while a falsely low one can leave a child unprotected. Because the tool operates at the point of professional judgment, the framework treats it as ethically permissible only if it augments that judgment—informing the worker’s assessment without displacing relational authority. A backstage system like SyRI or Robodebt, by contrast, operates on populations at scale, often without individual knowledge or meaningful consent. Its harms are structural: wrongful debt notices, discriminatory targeting, and the conversion of welfare applicants into suspects. The framework suggests that backstage automation faces a higher ethical bar precisely because no relational check exists to catch its errors before they land on people’s lives.

The three moral traditions also pull in different directions, and the matrix makes those tensions visible rather than hiding them. A utilitarian assessment of a predictive screening tool might emphasize aggregate benefits: earlier identification of at-risk children, more efficient allocation of scarce caseworker time, and prevention of tragedies. A deontological assessment asks whether the system respects the rights of the people it scores—whether they know they are being scored, whether they can contest the score, and whether consent is meaningful at all. A virtue-ethical assessment asks what the technology does to practitioners: does it cultivate attentiveness, humility, and care, or does it train workers to defer to a number and abandon the interpretive habits that relational practice requires? The paper’s position is that an AI system must survive scrutiny from all three lenses, not merely the consequentialist one, because welfare systems have historically justified harms to individuals by appeals to aggregate efficiency.

The four social-work values function as the normative anchors of the analysis. Privacy is threatened not only by data collection but by inference: a risk model can reveal intimate patterns of family life that no one deliberately disclosed. Self-determination is eroded when algorithmic outputs narrow the options a client is offered or the discretion a worker exercises. Social justice, the paper stresses, must be substantive rather than merely formal—a system can satisfy formal nondiscrimination while still distributing burdens unevenly across already-marginalized groups, because the underlying data reflect historical inequities. Human dignity is at stake whenever a person is reduced to a risk score or a fraud probability, processed by machinery they cannot see or question. The framework asks of every deployment: does this system advance these four values, or does it quietly trade them away for administrative convenience?

From the case analyses, the paper distills a set of conditions under which AI in social work is ethically defensible. First, AI must augment practitioner judgment without displacing relational authority—the human relationship remains the locus of decision, and the algorithm serves as one input among many. Second, AI must advance substantive rather than merely formal equity, which requires auditing outcomes across groups and correcting for the inequities embedded in historical data. Third, AI must be embedded in contestable, auditable institutions: affected people must be able to challenge algorithmic decisions, independent bodies must be able to inspect the systems, and the terms of deployment must remain politically revisable rather than locked in by technical opacity. Where these conditions fail, the same technology crosses from augmentation into automation, and the paper’s analysis suggests that resistance is then not Luddism but an ethical obligation grounded in the profession’s core commitments.

The Robodebt and SyRI episodes illustrate what failure looks like. Robodebt automated income-matching and debt issuance at scale, shifting the burden of proof onto vulnerable recipients and generating unlawful debts before a royal commission exposed the scheme’s flaws. SyRI’s opaque data-mining targeted low-income neighborhoods and was ruled in violation of the European Convention on Human Rights by a Dutch court in 2020. Both systems were backstage, uncontestable in practice, and structurally biased toward suspicion of the poor—precisely the configuration the framework flags as automated welfare. By contrast, the paper’s treatment of digital-care technologies suggests a more promising frontier, provided that monitoring remains genuinely consensual, that data flows serve the person being cared for rather than only the institution, and that human contact is not replaced by sensor readouts. The Korean crisis-household detection case sits between these poles, showing how administrative data can surface families in need—but only if detection triggers supportive outreach rather than punitive intervention.

The paper’s contribution is threefold, and its implications reach beyond social work. It reframes the field’s debate from replacement versus resistance to augmentation versus automation, a shift that turns a stale binary into a designable and governable distinction. It supplies a reusable diagnostic matrix that practitioners, agencies, and regulators can apply to concrete systems before deployment rather than after harm. And it integrates Korean, European, and North American evidence within a single analytic lens, demonstrating that the dilemmas of the digital welfare state are global even as their legal and institutional responses remain local. As governments worldwide race to embed AI in public services, the framework’s central warning is worth stating plainly: the question is never whether AI can make welfare administration faster or cheaper, but whether it leaves the human relationship—the actual substance of social work—intact. The research was supported by the Asan Foundation and a Gachon University research fund, and was presented at the Asan Foundation Symposium in Seoul in May 2026.

Subject of Research: An ethical framework for evaluating artificial intelligence as augmentation or automation in social work

Article Title: An ethical framework for assessing artificial intelligence as augmentation or automation in social work

Article References: Hwang, K. (2026). An ethical framework for assessing artificial intelligence as augmentation or automation in social work. Discover Social Science and Health. https://doi.org/10.1007/s44155-026-00463-x

Image Credits: AI Generated

DOI: 10.1007/s44155-026-00463-x

Keywords: artificial intelligence, social work ethics, augmented intelligence, predictive risk modelling, algorithmic decision-making, digital welfare state, relational practice, care ethics, anti-oppressive practice, Allegheny Family Screening Tool, Robodebt, SyRI