Compliance Is Not Protection: Why Federal AI in Benefits Administration Needs Civil Rights Audits

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
June 19, 2026

A mother working irregular hours depends on Medicaid, SNAP, or unemployment assistance to keep her household stable. An automated system flags her case. A notice arrives in language she does not fully understand. Her benefits are delayed. The agency calls it a processing issue. In her life, it becomes missed groceries, skipped medication, and unpaid rent.

She may never know an automated tool shaped the outcome. She only experiences the consequences. In benefits administration, compliance is not the same as protection.

The New Federal AI Baseline

In April 2025, the Office of Management and Budget issued two memos that set new federal rules for AI use. OMB Memorandum M-25-21 covers how agencies use AI. M-25-22 covers how agencies buy it. Together, they require agencies to appoint Chief AI Officers, set up governance boards, manage risks in high-impact systems, disclose vendor information, and guard against vendor lock-in.

Federal AI use is no longer experimental. The 2025 Federal Agency AI Use Case Inventory reports 3,611 AI uses across 56 agencies, including 445 high-impact systems. The Department of Veterans Affairs reports more than 200. The Social Security Administration uses AI in many service and benefits processes.

The federal government has not ignored AI governance. The problem is that following rules does not always mean people are being treated fairly.

Why Benefits Administration Is Different

AI in benefits programs is not ordinary government technology. If an automated system fails in a maintenance request system, the result may be inconvenience. If it fails in SNAP eligibility, Medicaid care decisions, disability claims, or unemployment fraud detection, the result can affect food, medicine, and survival.

The impact is not evenly shared. Low-income families, disabled people, older adults, and people with limited English proficiency are most affected, and least able to manage delays or errors. A delay is not neutral. Confusion is not neutral.

The Cautionary Record

Past cases show what can go wrong.

In Michigan, the MiDAS unemployment system (2013–2015) wrongly accused about 40,000 people of fraud. A state audit found a 93 percent error rate. The system processed accusations quickly, but workers had to fight slowly to clear their names.

In Arkansas, the ARChoices Medicaid program used an algorithm in 2016 to set in-home care hours. Some people saw their care reduced by about 43 percent. Lawsuits showed problems with fairness, notice, and how the system was designed. When an algorithm affects whether someone can bathe or remain at home, the issue is dignity, not efficiency.

A recent Brookings analysis found that more than 85 percent of high-impact federal AI systems in 2025 lacked key risk safeguards.

Compliance Documents the System. Protection Tests the Outcome.

A system can follow rules, have policies, and include human oversight language and still harm people in practice.

Compliance asks: Did the agency follow the required steps?

Protection asks: Did the system treat people fairly in real life?

Both matter, but they are not the same.

What a Civil Rights Audit Should Test

This is not about giving preference to any group. It is about checking whether systems that appear neutral end up creating unfair outcomes or barriers.

A civil rights audit should be required before and after any high-impact AI system is used in benefits programs. It should ask five basic questions:

1. Data quality: Does the system rely on data that misrepresents the people it affects?
2. Unequal outcomes: Are denial rates, fraud flags, delays, or appeals different across race, disability, age, language, or income groups?
3. Extra burden: Does the system shift more paperwork, verification steps, or digital requirements onto applicants?
4. Clarity: Do people understand why a decision was made and what they can do next?
5. Human judgment: Can staff override the system, and are they trained and supported to do so?

If an agency cannot answer these questions, it should not claim that compliance equals protection.

What Public Administrators Should Do Now

Agencies should not wait for the next MiDAS or ARChoices-style failure.

They should require civil rights audits before deploying high-impact AI systems and repeat them after deployment. These reviews should include civil rights staff, frontline workers, and affected communities.

Procurement contracts should also require vendors to disclose how systems work, what data they use, and how they are tested for fairness.

But the bigger challenge is not technical. It is institutional. AI in benefits administration is not just an IT decision. It is a public service decision and a civil rights decision.

The Standard Already Exists

Civil rights law already sets expectations for fairness in public programs. Title VI of the Civil Rights Act, the Americans with Disabilities Act, Section 504 of the Rehabilitation Act, and constitutional due process still apply when decisions are automated.

Courts have already made this clear. In cases involving automated benefit decisions, judges have found that systems without clear notice and fair process can violate due process rights.

The real question is no longer whether agencies are following federal AI guidance. It is whether following those rules is actually protecting people.

Compliance tells agencies whether they followed procedures. Civil rights audits tell them whether people were protected.

---

Author: Albert Nii Noi Okwei is a doctoral researcher in public policy and administration at the L. Douglas Wilder School of Government and Public Affairs, Virginia Commonwealth University, where he is affiliated with the Research Institute for Social Equity. His research focuses on AI governance, administrative burden, and equity-centered policy implementation in U.S. public administration. He is a 2026 Sara Miller McCune Scholar at the ICPSR Summer Program at the University of Michigan and a member of APPAM’s 2026 Student Activities Committee. He can be reached at [email protected].

(No Ratings Yet)
Loading...