PAN Lab example

Michigan MiDAS

After the settlements: review returns to a benefits-fraud system

Michigan's MiDAS software decided many unemployment fraud cases without human review. A 2017 lawsuit settlement made review a requirement. This case asks what sustains it.

MiDAS, Michigan's Integrated Data Automated System, is software that Michigan's Unemployment Insurance Agency rolled out in 2013 to decide unemployment-insurance fraud. From 2013 to 2015 it decided many fraud cases with no human review. In many cases it issued accusations, penalties, and collections on that basis.

Two points in time

This network shows the same deployment as the MiDAS System network, at its second documented point in time. The MiDAS System network shows the period of fully automated fraud decisions. This network shows the structure after that period ended and the settlements followed.

What went wrong, in figures

IEEE Spectrum reported a review of MiDAS from October 2013 to September 2015. In that period, MiDAS decided 40,195 fraud cases by algorithm alone, and 85 percent of those determinations were wrong. Another 22,589 cases had some human involvement, and 44 percent of those fraud claims were false.

In 2016, the Michigan Auditor General reviewed 22,000 MiDAS fraud determinations and found that 93 percent did not involve fraud. The sources do not explain how these counts relate.

MiDAS wrongly accused roughly 40,000 people. The harm the sources document includes seized tax refunds, garnished wages, quadruple penalties, and bankruptcies.

What changed, and when

In September 2015, the agency decided to stop using MiDAS for purely automated fraud assessment. That followed pressure from the federal government and a federal lawsuit filed that month.

In a 2017 settlement of the Zynda lawsuit, the agency agreed to stop using MiDAS's automated functions without human review. The settlement also made the agency reverse and refund certain fraud determinations.

Advocates filed Bauserman, a class action, in 2015. The parties settled it for $20 million in October 2022. The Michigan Attorney General's office described it as a civil-rights class action alleging false accusations of unemployment fraud.

Who acted

The controls came from outside the agency: federal pressure, courts, and litigation, years after MiDAS was deployed.

What this case asks

This case asks what holds the reviewed system in place. The review runs because a lawsuit settlement required it.

In this case, the agency's governance means whoever at the agency can change MiDAS and how it is used. The sources do not name them. They can improve MiDAS and add a check as MiDAS makes each determination. They can raise the staff's capacity to correct records. They can also limit how much MiDAS output staff act on without review.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. A tool closes a pathway when mistakes stop passing along it.

This case has a budget of 9 units and offers four tools. Upgrade model costs 3 units, or 5 at its stronger setting. Verify output costs 4. Train the staff costs 2, or 4 at its stronger setting. Pause AI on alarms costs 4. These prices are the same at every target level.

Explore (No Targets) sets no targets. Under Service Targets Only, the targets can be met. The cheapest way is Train the staff alone, at 2 of the 9 units. No combination that includes Pause AI on alarms meets them. Halting MiDAS also halts the determinations sent for human review, and the service target is missed.

Under Service and Safety Targets and All Governance Targets, this case is not fully addressable with the available tools. Both levels ask you to close every failure pathway, among other targets. Every combination within the budget was checked, and none meets them.

Money is not what stands in the way. With every tool at its strongest setting, costing 17 units, three failure pathways stay open. They are staff reading the determinations on file, staff writing corrections into the claimant records, and the determinations read in the lawsuits.

No tool offered here acts on any of them. The first two are the correction work this point in time exists for. The third is the courts' reading of the records. That is a finding about the deployment, not a gap in your approach.

Stylized model of a documented deploymentPublic benefits & eligibility

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the MiDAS-class review network after the settlements network: 5 components and 8 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 6 assumed · 3 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

Show all 9 assumptions

assumed
This network models the same deployment as the MiDAS System network, Michigan's MiDAS, at a later point in time. It shows the structure after the agency stopped purely automated fraud assessment in 2015, and after the settlements that followed. It is a simplified model of that structure, not a copy of the real system.
baseline
The point in time shown here comes from the settlements. MiDAS decided many fraud cases with no human review from 2013 to 2015. In 2017, the Zynda settlement made human review a requirement. Bauserman, a class action alleging false accusations of unemployment fraud, settled for $20 million in 2022. The network shows the human review and the review check as working, because the sources describe review as a settlement term. It shows the correction work at a scale of tens of thousands of people, because roughly 40,000 people were wrongly accused.
baseline
The network treats the human review as review that engages with the determinations, not a formality. That comes from this case's own contrast, from October 2013 to September 2015. Cases MiDAS decided by algorithm alone were wrong 85 percent of the time. In cases with some human involvement, 44 percent of the fraud claims were false. Review is still no guarantee, so the network does not treat it as catching every mistake.
assumed
Pathways the sources show ended with fully automated fraud decisions are left off this network. MiDAS writing determinations straight into claimant records is one. Copying fraud flags into enforcement and collections is another. The others are one set of rules applied to every case, and earlier determinations used in later ones. All four belong to the fully automated period of 2013 to 2015. The MiDAS System network and the case file describe them.
baseline
The pathway from MiDAS straight to staff stays on the network, though the agency stopped purely automated fraud assessment in 2015. The agency decides how much MiDAS output staff may act on without review. That is the question this point in time poses.
assumed
MiDAS writing determinations into the records is left off this network. The paired Robodebt Remediation network keeps its own writing pathway, because that reform can govern how much is written. The MiDAS reform, as the sources describe it, gives the agency no such control. So an ended pathway with nothing governed behind it is left off.
assumed
Two things the agency can do after the settlements have no tool in this case. One is cleaning the determination records by checking what each one says. The other is raising the share of determinations that can be checked against the facts they rest on. So the cleaning of wrongful determinations appears in the corrections staff write, not in a tool. The Lab offers no tool that audits stored records by their content. A Lab tool that deletes records without reading them would not do this cleaning.
assumed
Workload and staffing follow the same reasoning as the MiDAS System network, applied to this later point in time. The standing workload is the correction and settlement work itself. Without MiDAS, the alternative is the working manual process for deciding fraud that the automation replaced. That manual process is what the review requirement brought back.
assumed
MiDAS's harm fell on unemployment claimants. The sources document seized tax refunds, garnished wages, quadruple penalties, and bankruptcies. This network traces how mistakes pass between the agency's parts, not that harm. Claimants are not a part of the network. The network counts the settlement sum as part of the institution's work, not as an outcome for any claimant. It does not estimate what anyone the agency serves gained or lost.

What this example does not show

Show all 2 limitations

MiDAS's harm fell on unemployment claimants. The sources document seized tax refunds, garnished wages, quadruple penalties, and bankruptcies. A settlement does not repair that harm. This example traces only how mistakes pass inside the institution. It counts the settlement sum as part of the institution's work, and it leaves claimants out of the diagram. It does not estimate what anyone the agency serves gained or lost.
Two things the agency can do after the settlements have no tool in this case. They are cleaning the determination records by checking what each one says, and raising the share of determinations that can be checked against source facts. The network's assumptions describe both. The cleaning of wrongful determinations appears in the corrections staff write, not in a tool.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

From 2013 to 2015 MiDAS decided many fraud cases with no human review, issuing 60,000+ determinations and wrongly accusing roughly 40,000 people; the agency stopped purely automated fraud assessment in September 2015, and litigation followed — the Zynda settlement forced reinstatement of human review of fraud determinations (2017), and the Bauserman class action over false fraud accusations settled for $20M (2022).
empirical
- Government Michigan AG, settlement of civil-rights class action (Bauserman, 2022) https://www.michigan.gov/ag/news/press-releases/2022/10/20/som-settlement-of-civil-rights-class-action-alleging-false-accusations-of-unemployment-fraud
- Investigative AI Incident Database, Incident 373 (MiDAS false fraud claims) https://incidentdatabase.ai/cite/373/
- Advocacy Benefits Tech Advocacy Hub, Michigan UI False Fraud Determinations https://www.btah.org/case-study/michigan-unemployment-insurance-false-fraud-determinations.html
- Investigative IEEE Spectrum, Michigan's MiDAS unemployment system: Algorithm alchemy that created lead, not gold https://spectrum.ieee.org/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold
In the documented MiDAS case, a 2016 review of 22,000 fraud determinations found 93% did not involve fraud, and from October 2013 to September 2015 determinations erred at about 85% when decided by algorithm alone versus 44% with some human involvement.
empirical
- Government Michigan AG, settlement of civil-rights class action (Bauserman, 2022) https://www.michigan.gov/ag/news/press-releases/2022/10/20/som-settlement-of-civil-rights-class-action-alleging-false-accusations-of-unemployment-fraud
- Investigative IEEE Spectrum, Michigan's MiDAS unemployment system: Algorithm alchemy that created lead, not gold https://spectrum.ieee.org/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold
- Investigative AI Incident Database, Incident 373 (MiDAS false fraud claims) https://incidentdatabase.ai/cite/373/
- Advocacy Benefits Tech Advocacy Hub, Michigan UI False Fraud Determinations https://www.btah.org/case-study/michigan-unemployment-insurance-false-fraud-determinations.html
Michigan's MiDAS system auto-adjudicated unemployment-insurance fraud with an extremely high error rate among automated determinations, wrongly accusing tens of thousands of people; litigation and court action forced review and compensation.
empirical
- Government Michigan AG, settlement of civil-rights class action (Bauserman, 2022) https://www.michigan.gov/ag/news/press-releases/2022/10/20/som-settlement-of-civil-rights-class-action-alleging-false-accusations-of-unemployment-fraud
- Investigative IEEE Spectrum, Michigan's MiDAS unemployment system: Algorithm alchemy that created lead, not gold https://spectrum.ieee.org/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold
- Investigative AI Incident Database, Incident 373 (MiDAS false fraud claims) https://incidentdatabase.ai/cite/373/
- Advocacy Benefits Tech Advocacy Hub, Michigan UI False Fraud Determinations https://www.btah.org/case-study/michigan-unemployment-insurance-false-fraud-determinations.html

Where this connects

Institutional pressures in this domain

Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.

All of them in context on the Public benefits & eligibility domain page.

Levers available here and the patterns behind them

Upgrade model — Improve the model
Verify output — Put a verifier on the agent
Train the staff — AI literacy & boundary rules
Pause AI on alarms — Deployment circuit-breaker