Arkansas Cut 18,000 Eligible People From Medicaid Over Paperwork, and the 2027 Fix Is AI Its Own Authors Say Will Reproduce the Bias

We have run this experiment already, and the record is not ambiguous. In June 2018 Arkansas became the first state in the country to bolt a work requirement onto Medicaid, and within seven months the state stripped more than 18,000 people of their coverage. The rule itself sounded like common sense, the kind of thing most working people would nod along to: log 80 hours a month of work or community engagement, report it through a state website, keep your insurance. Then the results came in. More than 95 percent of the people subject to the rule already met it or qualified for an exemption. They were working, disabled, caregiving, or otherwise carved out by the rule’s own exceptions. They lost coverage anyway.

Harvard researchers tracking the rollout in the New England Journal of Medicine documented that the overwhelming majority of the target population should have sailed through and lost insurance regardless, because they never heard about the reporting mandate, could not work the website, or got tangled in the documentation. A follow-up in Health Affairs checked the other half of the promise and found no increase in employment two years on. The rule did not move people into jobs. It moved eligible people off insurance. The problem was never the idea of work. It was the paperwork engine bolted on top of it.

That is the history sitting behind a paper published August 7 in JAMA Health Forum, and it is the history the paper is trying to head off. The 2025 budget reconciliation law, HR 1, revives Medicaid work requirements nationwide, and beginning January 1, 2027, adults covered under the ACA expansion in 44 states will have to document 80 hours a month of qualifying activity or lose coverage. Same rule as Arkansas, now at national scale. The authors know precisely how Arkansas ended, and they are proposing artificial intelligence as the way to keep the national version from ending the same way.

Point the machine at the paperwork

The pitch is concrete. Use machine learning to link Medicaid rolls with payroll, tax, and other program data so a state can confirm someone is working without making them prove it a second time. Point generative models at the code that screens for compliance. Build predictive tools to flag eligible people at risk of getting cut for procedural reasons. Lean on chatbots, already running in about a quarter of state Medicaid programs, to walk applicants through the maze, and have language models read call-center transcripts to find where people get stuck. The authors sort all of it into five domains: execution, targeting, communication, monitoring, and adaptation. The logic is clean enough. If the Arkansas failure was administrative friction, a system built to strip out friction should rescue the eligible people the old system dropped.

It is a serious proposal, and the seriousness is the reason to look harder, not to wave it through. These are among the more cited health-services researchers in the country, and they are not selling a fantasy of hands-off automation. They are trying to fix a real trap. The question is whether the tool they reach for can carry the weight they want to put on it.

What the builders admit in their own paper

Start with who is building it. The lead author is Emma McGinty of Weill Cornell Medicine, writing with colleagues at Weill Cornell and the University of Michigan. Their disclosure statement runs long: grants and personal fees touching the National Institutes of Health, the CDC, the Centers for Medicare and Medicaid Services, the Novartis Foundation, the Commonwealth Fund, the Robert Wood Johnson Foundation, and Arnold Ventures, among others, with one co-author reporting personal fees from JAMA Network itself. This is the health-policy establishment proposing to route a slice of eligibility determination through software, and it is asking for that trust on the far side of a pandemic in which its judgment did not exactly earn a blank check. The deeper worry is not the letterhead. It is public agencies handing the gate to a system that the person on the wrong side of it cannot see into or argue with.

The strongest case against the plan is the authors’ own writing. Asked about the approach, McGinty told a reporter that “there are biases baked into our data that AI implementation will 100% reproduce here, and so having a human in the loop and really careful monitoring and oversight of AI is needed.” The paper says the same thing in colder language: Medicaid enrollees with high needs but poor access to care “may appear to have low use and costs in Medicaid claims data,” so a predictive model trained on that data “would reproduce this bias” and quietly write off the very people it is meant to protect. The same paper warns that AI carries “opaque internal workings” and “generation of hallucinated false information,” that “humans will need to conduct bias checks to identify false matches and nonmatches” when records are linked, and that, in a line most authors would bury, “AI tools are unlikely to engender trust in public programs.”

Then the cruelest detail, stated plainly in the text: “tech literacy is often lowest in the groups most vulnerable to implementation failure, such as people with disabilities.” The population a chatbot-first system is most likely to lose is the same population the rule is most likely to cut.

So the remedy for a paperwork machine that dropped more than 18,000 people, at least 95 percent of whom already met the rule or qualified out, is a prediction machine its own designers say will reproduce the bias in its data, cannot fully explain its own decisions, and will struggle most with the most vulnerable. To their credit, the authors hide none of this. They insist AI must “augment, but not replace” human judgment, that “this cannot be a ‘hand it over to the bots’ solution,” and that rigorous testing with humans in the loop has to come first. That caution is exactly right. It is also the first line item cut when a state agency is understaffed, on deadline, and handed a tool sold as a way to do the work of people it no longer has to pay.

None of this is an argument against expecting people to work. It is an argument about who runs the gate, and whether the person on the far side of it meets a human who can hear “I lost my job in March, my hours are in the tax file, please look,” or an interface that already scored them and moved on.

The paper is not the test. The disenrollment numbers coming out of the states in the first months of 2027 are, and the question is whether CMS and the states publish them straight or let them blur the way Arkansas’s confusion blurred for months. If the system works, coverage losses will track the number of people actually refusing to work. If it fails, the losses will run far past that number, the same tell Arkansas showed in 2018. Watch the gap between the two. A good machine closes it. Arkansas with a faster processor widens it, and this time it does so in every expansion state at once.

Sources

1. JAMA Health Forum – McGinty et al., “Closing the Health Policy Implementation Gap With Artificial Intelligence” (Special Communication, Aug 7, 2026)
2. Arkansas Advocates for Children and Families – 18,000 Arkansans stripped of coverage under work reporting requirements in 2018
3. New England Journal of Medicine – Sommers et al., Medicaid work requirements, first-year results from Arkansas
4. Health Affairs – two-year impacts of Arkansas Medicaid work requirements on coverage and employment
5. KFF – A Closer Look at the Work Requirement Provisions in the 2025 Federal Budget Reconciliation Law
6. News-Medical – Weill Cornell summary of the JAMA Health Forum paper, with McGinty’s “100% reproduce” comment