PAN Lab example

Michigan MiDAS

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

The same automated fraud-determination deployment as the MiDAS System network, at its second documented instant. Litigation changed the structure: the Zynda settlement required human review of fraud determinations to return in 2017, and the Bauserman class action over false fraud accusations settled for $20M in 2022. The no-review mode that issued 60,000+ determinations at a documented ~93% error rate is over. The live question this instant poses is what holds the reviewed system in place: the review runs because a court ordered it, and the reform's remit — a better engine, a check at generation, correction capacity, and the adoption rate — is what the institution can now build for itself.

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.

assumed
This models the same deployment as the MiDAS-class automated determination network, drawn at its second documented instant: after the settlements that ended the no-review mode. It is a stylized model of the documented post-settlement structure, not a reconstruction of the actual system.
baseline
The instant drawn here is documented by the settlements: the no-review adjudication mode ran from 2013 to 2015, the Zynda settlement required reinstatement of human review in 2017, and the Bauserman class action over false fraud accusations settled for $20M in 2022. The review stream and the examiner check are drawn live because the record documents the review mandate as a completed settlement term, and the correction write-back is drawn at documented tens-of-thousands scale — about 40,000 wrongful accusations behind the redetermination work.
baseline
The review intensity is derived from this case's own documented contrast: determinations erred at about 85% without human review versus 44% with it, so the reinstated review is drawn as a stage that substantively engages with determinations rather than a formality — and still not a guarantee, which is why the review is drawn as substantive rather than complete.
assumed
Pathways the record shows ended with the discontinued mode are not drawn: the automatic write of determinations into claimant records, the replication of flags into enforcement and collections, the one-rule application across every case, and the re-feed of prior determinations into later ones all belong to the 2013–2015 mode, and are recorded in the paired network and the case file.
baseline
The adoption pathway stays on the map carrying nothing, because the documented post-settlement governance actor explicitly holds its rate among the things it can change — how much engine output may reach staff without review is the governed question this instant poses.
assumed
The engine's write pathway is not drawn at this instant, unlike the paired Robodebt post instant which draws its write pathway empty: there, the reform actor's remit governs the write rate; here, the documented post-settlement governance holds no write-rate knob, so an ended pathway with no governed quantity behind it is simply not on this map. That difference is derived from the two reform records, not authored for separation.
assumed
Two of the documented reform capabilities have no lever on this board: content-aware cleaning of the determination records (decontaminate) and raising the share of determinations that can be verified against source facts (frac_verifiable). The cleaning of wrongful determinations therefore lives in the drawn correction flow and this record, not in a lever — the Lab offers no content-aware store-audit lever, and a content-blind purge is its documented counter-example.
assumed
Demand and capacity carry the same derivation as the pre instant, re-read against this instant's record: the standing workload is the redetermination and settlement administration itself, and the no-AI counterfactual is the working manual adjudication process the automation displaced — which is what the review mandate ordered back.
assumed
MiDAS's harm fell on unemployment claimants — wage garnishment, tax-refund seizure, quadruple penalties, bankruptcies are in the documented record. This Lab models institutional propagation, not that harm; claimants are not represented as a node, the settlement sum is an institutional flow rather than an outcome for any served person, and no benefit change for any served person is estimated.

What this example does not show

MiDAS's harm fell on unemployment claimants — garnished wages, seized refunds, quadruple penalties, bankruptcies are in the documented record — and a settlement is not repair of that harm. This Lab models institutional propagation only; the settlement sum is an institutional flow, claimants are not represented in the diagram, and no benefit change for any served person is estimated.
Two documented reform capabilities have no lever on this board: content-aware cleaning of the determination records and raising the share of determinations verifiable against source facts. They are recorded in the network's assumptions; the cleaning of wrongful determinations lives in the drawn correction flow, not in a lever.

Sources and evidence

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

MiDAS's no-review adjudication mode ran from 2013 to 2015, issuing 60,000+ determinations and wrongly accusing roughly 40,000 people; litigation ended it — 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, error among no-review auto-adjudications ran roughly 93%, and determinations erred at about 85% without human review versus 44% with it.
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
Pause AI on alarms — Deployment circuit-breaker