The Caseworker in the Age of AI: Protecting Discretion While Embracing Automation

The views expressed are those of the author and do not necessarily reflect the views of ASPA as an organization.

By Albert Nii Noi Okwei
July 10, 2026

Between 2013 and 2015, Michigan ran much of its unemployment fraud system on autopilot. The Michigan Integrated Data Automated System, or MiDAS, auto-adjudicated more than 53,000 fraud determinations with little or no human review. Of the roughly 22,000 cases the state later examined, about 93 percent were overturned. The system flagged income discrepancies as fraud, mailed questionnaires to dormant online accounts, then cut off benefits and seized wages and tax refunds before many people knew a decision had been made. The class-action settlement finalized in 2024 cost the state $20 million.

The technology is back now, faster and better funded. In April 2025, the Office of Management and Budget issued Memorandum M-25-21, which encourages agencies to accelerate AI adoption while maintaining risk-management practices for high-impact systems. Benefits eligibility and fraud determinations fall squarely within that category. Most guidance on governing these tools is written for agency directors and procurement officers. Much less guidance reaches the caseworker who must act on an algorithmic score while managing dozens of open cases.

Discretion Is the Job, Not the Obstacle

Michael Lipsky described frontline public servants as street-level bureaucrats: the people who turn statutes into decisions about individual circumstances. A caseworker reads a file that no law can fully anticipate, weighs factors no form captures and makes judgment calls. That discretion is not a defect to be eliminated. In a benefits office, it is the work.

When an algorithm quietly absorbs that judgment, the law’s promise of individual consideration can disappear with it.

The constitutional requirements surrounding due process remain. Goldberg v. Kelly still requires timely notice and a fair hearing before the government terminates someone’s benefits. No algorithm meets that standard on its own, and no vendor contract can eliminate those obligations.

Automation Bias Is the Real Risk

The challenge is not simply blind obedience to technology. Saar Alon-Barkat and Madalina Busuioc found that public officials practice “selective adherence”: they accept an algorithm’s recommendation most readily when it confirms what they already believe, often about individuals with less power to challenge decisions.

The tool does not remove human judgment. It can influence how that judgment is applied.

The consequences are documented. When Arkansas replaced nurse assessments with an algorithm that determined Medicaid home-care hours, disabled residents whose conditions had not changed experienced significant reductions in care. In the Netherlands, a child-benefit algorithm that used nationality as a fraud indicator wrongly accused thousands of families and contributed to a government crisis. In Allegheny County, researchers found that caseworkers who examined cases beyond a risk score reduced a projected racial gap in child-welfare screenings.

Human review worked, but only when workers had the time, authority and information necessary to challenge the system.

Five Rules for Caseworker-Centered AI Governance

Keeping a human in the loop means little if that person lacks the ability to act. Five principles can make the difference.

1. AI informs; caseworkers decide. No algorithm should issue a final denial, reduction or fraud finding. Every adverse action should receive review from a designated human decision-maker.

2. The right to override has to be real. Agencies should track override rates. A caseworker who never disagrees with a system is not meaningfully reviewing it. Workers should not be penalized for questioning algorithmic recommendations.

3. Every recommendation requires an explanation. Before taking action, workers should understand the factors behind an algorithmic recommendation in clear language. Agencies should build this requirement into vendor contracts.

4. Audit for disparate impact before and after deployment. Agencies should test AI tools for bias across race, disability, age and geography before implementation and continue monitoring after deployment. The National Institute of Standards and Technology’s AI Risk Management Framework provides a useful structure.

5. Residents maintain rights to notice, explanation and appeal. Individuals affected by an automated decision should know when AI was used, understand the reasons behind the decision and have access to meaningful human review.

The Future of Public Service Is Still Human

The strongest argument for AI in benefits offices is the honest one. Workers are buried, and tools that handle routine lookups can hand them back time for the cases that need a person. The Roosevelt Institute found, though, that when workers have little say over those tools, errors and workloads tend to climb rather than fall. Good design is possible. Code for America and Anthropic recently built a SNAP policy navigator that answers a caseworker’s policy questions but deliberately leaves the eligibility call with the worker.

That distinction matters: support, not substitution.

The people who determine eligibility and connect residents to services are not obstacles between citizens and technology. They are the human infrastructure of accountable government. AI can make their work more efficient and identify issues a person may miss. It cannot provide the judgment on which public service depends, and agencies should not ask it to.

Government agencies will continue adopting these tools. Whether they establish safeguards before the next MiDAS or explain failures afterward remains a decision within their control.

---

Author: Albert Nii Noi Okwei is a doctoral researcher in public policy and administration at Virginia Commonwealth University’s Research Institute for Social Equity (RISE), where his work focuses on digital governance, administrative burden, and equity-centered AI adoption in public service delivery. He can be reached at [email protected].

(No Ratings Yet)
Loading...