# Michigan's MiDAS fraud algorithm wrongly accused tens of thousands: 40,195 cases decided by machine, a 93% error rate, and a $20 million settlement

> From October 2013 to September 2015 Michigan's Unemployment Insurance Agency let an automated system, MiDAS, adjudicate unemployment fraud with little human review: of 62,784 penalty cases the agency later re-examined, 40,195 had been resolved by the computer program alone and 85% of those findings were reversed on review, and an internal review of more than 20,000 determinations found a 93% error rate. MiDAS imposed 400% penalties and seized wages and tax refunds without a hearing. The Michigan Supreme Court ruled in 2022 that the falsely accused could sue the state for the due-process violation, and a $20 million class settlement was approved by the Court of Claims in January 2024. Figures here are quoted from the state's own published review, independent journalism, and the primary court record.

- Verification status: pending
- Case type: deployment
- Provider: Michigan Unemployment Insurance Agency (UIA) automated fraud-detection system, the Michigan Integrated Data Automated System (MiDAS), built by contractor Fast Enterprises
- Client: State of Michigan / Unemployment Insurance Agency (self-deployment), government (named)
- Sector: government / US / ops
- Canonical URL: https://theinternetninja.com/stories/michigan-midas-unemployment-fraud-algorithm-40195-cases-93pct-error-20m-settlement-2024/
- Source: The Internet Ninja (theinternetninja.com), independent verified-proof platform

## Outcomes

| Metric | Before | After |
| --- | --- | --- |
| Of 62,784 re-examined penalty cases, 40,195 were resolved by the computer program alone and 85% of those fraud findings were reversed on review (State of Michigan UIA completed review, 2017; IEEE Spectrum, 2017; Detroit Free Press / Paul Egan, 2017) |  |  |
| An internal state review of more than 20,000 fraud determinations found a 93% error rate (Bridge Michigan, 2019; The Markup, 2022) |  |  |
| MiDAS imposed a 400% penalty and seized wages and tax refunds without notice or a hearing; the Michigan Supreme Court held in 2022 the falsely accused could sue the state for money damages (Bauserman v UIA, Michigan Supreme Court, 2022) |  |  |
| A $20 million class settlement for more than 3,000 falsely accused claimants was approved by the Court of Claims in January 2024 (Bridge Michigan, 2024) |  |  |

## Verification method

Every load-bearing figure is quoted verbatim from a source fetched and grepped and saved to sources/. The 40,195 computer-resolved cases and 85% reversal rate, and the 22,589 computer-initiated-then-referred cases and 44% rate, are quoted from the State of Michigan UIA's own completed-review press release (2017), the state's official published account of re-examining these cases, and are independently corroborated by two secondaries reporting the same split: IEEE Spectrum (2017, Tier 2) and the Detroit Free Press (Paul Egan, 15 Aug 2017, Tier 2, via its governing.com/Tribune News Service republication). The 93% error rate comes from the primary Michigan Supreme Court Bauserman opinion (2022, Tier 1), corroborated by Bridge Michigan (2019, Tier 2) and The Markup (2022, Tier 2). The 400% penalty is from IEEE Spectrum; the due-process holding from the Bauserman opinion and The Markup; the $20 million settlement from Bridge Michigan (2024, Tier 2). The Bauserman opinion PDF was machine-extracted (it has a full text layer), which both quote-verifies the due-process holding and records the state review's ~93% error figure in the primary court record. The Michigan Auditor General's 2016 MiDAS audit was examined and decrypted but is a MiDAS security/appeals/performance audit that does NOT contain the 40,195/85% or 22,589/44% figures, so it is not cited for them. The 40,195/85% split is corroborated beyond the state's own review by two independent secondaries (IEEE Spectrum and the Detroit Free Press); the state review itself is the first party, an admission against its own interest, used only as an already-public record, not a chased confirmation.

## FAQ

**How many fraud cases did Michigan's MiDAS decide by algorithm, and how often was it wrong?**

The agency's own completed review of penalty cases from October 2013 to August 2015 reported that of 62,784 cases, 40,195 'were originally resolved by way of computer program' and '85 percent of these original fraud findings were reversed', while a further 22,589 were computer-initiated then referred to an investigator and 44 percent of those were reversed. IEEE Spectrum reported the same figures, describing the 40,195 as adjudicated 'by algorithm alone' with 85 percent 'resulting in incorrect fraud determinations'. A separate internal state review of more than 20,000 determinations found a 93% error rate.

**What did MiDAS do to people it flagged?**

MiDAS disqualified claimants, imposed a 400% penalty on the alleged fraud amount, and seized wages and tax refunds without notice or an opportunity to be heard. In July 2022 the Michigan Supreme Court held in Bauserman v Unemployment Insurance Agency that the falsely accused had stated a cognizable constitutional-tort claim for which they could recover monetary damages.

**Was there a settlement, and does it cover everyone affected?**

In January 2024 the Michigan Court of Claims approved a $20 million settlement shared by more than 3,000 class members (the Bauserman class). That is distinct from an earlier roughly $20.8 million the agency said it would refund in penalties; the two figures should not be merged, and the class does not cover every one of the tens of thousands wrongly accused.

## Full case file

**Verification status: CHECKING — every figure is corroborated against the state's own published review, independent journalism, and the primary court record. The exact 40,195/85% auto-adjudication split is quoted from the state's own completed review and independently corroborated by two secondaries (IEEE Spectrum and the Detroit Free Press). Not yet a green badge; the checker grades it.**

## The problem
Facing budget pressure, Michigan's Unemployment Insurance Agency automated the detection and adjudication of unemployment-insurance fraud through a system called MiDAS, the Michigan Integrated Data Automated System. The design let the machine decide fraud with little or no human review, and it did so at scale: the system "initiates an automated process that can result in recipients being disqualified from benefits and subjected to penalties and criminal prosecution, all without notice or an opportunity to be heard" ([source](https://www.courts.michigan.gov/4a1c19/siteassets/case-documents/opinions-orders/msc-term-opinions-(manually-curated)/21-22/bauserman-op.pdf)). A fraud finding was not a warning letter; it triggered a <span class="kpi">400%</span> penalty on the alleged amount and the seizure of wages and tax refunds ([source](https://spectrum.ieee.org/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold)).

## What was built
MiDAS was an automated rules system that flagged and adjudicated fraud without a person confirming its conclusions. Between October 2013 and August 2015 it ran largely unsupervised, and when the agency itself later re-examined the penalty cases its output was overwhelmingly wrong. Of 62,784 penalty cases the agency reviewed, "40,195 were originally resolved by way of computer program based on available information" and "85 percent of these original fraud findings were reversed" ([source](http://web.archive.org/web/20190111094548/https://www.michigan.gov/som/0,4669,7-192-47796-428651--,00.html)). Even where a human joined the loop the results were poor: the other "22,589 cases of the original 62,784 were initiated by computer program and then referred to an investigator", and "44 percent of these fraud findings were reversed" ([source](http://web.archive.org/web/20190111094548/https://www.michigan.gov/som/0,4669,7-192-47796-428651--,00.html)). Independent reporting described the same split, casting the reversals as errors: "MiDAS adjudicated—by algorithm alone—40,195 cases of fraud, with 85 percent of those resulting in incorrect fraud determinations" and "Another 22,589 cases that had some level of human interaction involved in a fraud determination found a 44 percent false fraud claim rate" ([source](https://spectrum.ieee.org/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold)). A second independent newsroom reported the same split from the agency's figures: "40,195 were robo-adjudicated by MiDAS with no human involvement and those had an 85% error rate" and "Another 22,589 cases involving some human interaction had a 44% error rate, the agency said" ([source](https://www.governing.com/topics/health-human-services/tns-michigan-unemployment-fraud-accusations.html)). The plaintiffs' attorney later summed up the scale bluntly: "MiDAS was wrong 93 percent of the time, and in about 40,000 instances it incorrectly found fraud" ([source](https://themarkup.org/newsletter/hello-world/the-seven-year-struggle-to-hold-an-out-of-control-algorithm-to-account)).

## The outcome
The failure was not a modelling curiosity; it fell on real people who lost paychecks and refunds. The scale of the error is recorded in the primary court record: the Michigan Supreme Court's Bauserman opinion states that a review by the state's own agency concluded "approximately 93% of the automated system's fraud determinations were incorrect" ([source](https://www.courts.michigan.gov/4a1c19/siteassets/case-documents/opinions-orders/msc-term-opinions-(manually-curated)/21-22/bauserman-op.pdf)). Independent reporting on that same review put it identically: "An internal state review of more than 20,000 fraud claims showed a <span class="kpi">93 percent</span> computer error rate" ([source](https://bridgemi.com/michigan-government/broken-human-toll-michigans-unemployment-fraud-saga/)). Because the penalties and seizures happened with no hearing, the case turned into a landmark due-process fight. In July 2022 the Michigan Supreme Court ruled the falsely accused could sue the state: "In July of this year, the Michigan Supreme Court finally ruled that workers falsely accused of unemployment fraud can sue the state" ([source](https://themarkup.org/newsletter/hello-world/the-seven-year-struggle-to-hold-an-out-of-control-algorithm-to-account)). The court held "plaintiffs have alleged a cognizable constitutional-tort claim for which they may recover monetary damages" ([source](https://www.courts.michigan.gov/4a1c19/siteassets/case-documents/opinions-orders/msc-term-opinions-(manually-curated)/21-22/bauserman-op.pdf)).

The accountability arrived slowly and only partly. In January 2024 a settlement closed part of the ledger: "More than <span class="kpi">3,000</span> Michiganders falsely accused of unemployment fraud are in line for a share of a <span class="kpi">$20 million</span> settlement finalized this week in the state Court of Claims" ([source](https://bridgemi.com/michigan-government/michiganders-falsely-accused-jobless-fraud-share-20m-settlement/)). That figure should not be conflated with an earlier, separate remediation: the agency at one point "said it would refund $20.8 million to residents who were wrongfully accused of fraud" ([source](https://themarkup.org/newsletter/hello-world/the-seven-year-struggle-to-hold-an-out-of-control-algorithm-to-account)). The two are different amounts for different purposes, and the 3,000-member settlement class is far smaller than the tens of thousands the system wrongly accused.

**Weakest load-bearing source.** The exact 40,195/85% split comes first from the State of Michigan's own completed review, which is the first party to the story ([source](http://web.archive.org/web/20190111094548/https://www.michigan.gov/som/0,4669,7-192-47796-428651--,00.html)). That is an admission against the agency's own interest rather than a chased or self-serving confirmation, but it is still the subject, so its weight rests on independent corroboration. Two independent secondaries report the same precise split: IEEE Spectrum, which describes the 40,195 as adjudicated "by algorithm alone" with 85 percent "incorrect" ([source](https://spectrum.ieee.org/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold)), and the Detroit Free Press (Paul Egan, 15 August 2017), which reports "40,195 were robo-adjudicated by MiDAS with no human involvement and those had an 85% error rate" ([source](https://www.governing.com/topics/health-human-services/tns-michigan-unemployment-fraud-accusations.html)). The Detroit Free Press original is behind a paywall, so it is quoted from its Tribune News Service republication on governing.com, which carries the wire copy in full. The Markup's "~40,000 wrongly / 93% wrong" and Bridge Michigan's "more than 20,000 / 93%" describe a different statistic and are not counted toward this split. The 93% error rate, separately, is quote-verified against a primary court record: the Michigan Supreme Court's Bauserman opinion records the state review's ~93% figure in its own text ([source](https://www.courts.michigan.gov/4a1c19/siteassets/case-documents/opinions-orders/msc-term-opinions-(manually-curated)/21-22/bauserman-op.pdf)). No confirmation was sought from the State of Michigan; the badge never depends on the subject, and the state review is used only as an already-public official record.

> **How this was verified.** Method: every load-bearing figure is quoted verbatim from a source fetched and grepped this session and saved to sources/. The 40,195 computer-resolved cases and 85% reversal rate, and the 22,589 computer-initiated-then-referred cases and 44% rate, are quoted from the State of Michigan UIA's own completed-review press release ("Michigan's unemployment agency completes review of fraud determination cases", 2017; the state's official published account, first party, an admission against its own interest; Wayback 20190111094548, saved to sources/michigan-uia-completed-review-2017.html) and are independently corroborated by two secondaries reporting the same split: IEEE Spectrum ("Michigan's MiDAS Unemployment System: Algorithm Alchemy Created Lead, Not Gold", Robert N. Charette, 2017, updated 2018, Tier 2; Wayback 20260620103624, saved to sources/ieee-spectrum-midas.html), which also carries the 400% penalty, and the Detroit Free Press (Paul Egan, "After Falsely Accusing Thousands of Unemployment Fraud and Wrongly Taking Their Money, Michigan Makes Amends. But Is It Enough?", 15 August 2017, Tier 2), quoted from its Tribune News Service republication on governing.com (the freep.com original is paywalled; saved to sources/detroit-free-press-egan-2017-governing.html and archived Wayback 20190808013428, both quotes verbatim). The 93% error rate is quote-verified against a primary: the captured Michigan Supreme Court Bauserman opinion PDF was machine-extracted (it has a full text layer, 147 content streams) and records "approximately 93% of the automated system's fraud determinations were incorrect" (26 July 2022, Tier 1; Wayback 20260713223317, saved to sources/bauserman-msc-opinion.pdf); it is corroborated by Bridge Michigan ("Broken: The human toll of Michigan's unemployment fraud saga", 2019, Tier 2; Wayback 20260215151947, saved to sources/bridgemi-broken-human-toll.html) and The Markup ("The Seven-Year Struggle to Hold an Out-of-Control Algorithm to Account", 20 August 2022, Tier 2; Wayback 20260817231330, saved to sources/themarkup-midas.html). The due-process holding is quoted from the same Bauserman opinion. The $20 million settlement is from Bridge Michigan (31 January 2024, Tier 2; Wayback 20260801091826, saved to sources/bridgemi-20m-settlement.html). The Michigan Auditor General's 2016 MiDAS audit (sources/michigan-auditor-general-midas-2016.pdf) was decrypted and read but is a MiDAS security/appeals/performance audit that does not contain the 40,195/85% or 22,589/44% figures, so it is not cited for them. The 40,195/85% split is corroborated by two independent secondaries (IEEE Spectrum and the Detroit Free Press) beyond the state's own review. Checked 2026-08-31 (checking round 4).

## Related case files
- [Australia's Robodebt automated income-averaging scheme was ruled unlawful and settled for A$1.8 billion](/stories/robodebt-australia-automated-income-averaging-unlawful-475m-2026-class-action-settlement/) — the closest analogue: a government automated debt-and-fraud system that generated false liabilities at scale, was found unlawful, and ended in a mass settlement.
- [Amsterdam halted its Smart Check welfare-fraud AI after bias testing](/stories/amsterdam-slimme-check-smart-check-welfare-fraud-ai-halted-bias-1600-applications-2023/) — a government fraud-detection algorithm stopped before it could repeat MiDAS-style harm, useful for contrasting prevention with Michigan's after-the-fact remedy.
- [Denmark's Udbetaling Danmark automated welfare-fraud algorithms drew a mass-surveillance warning from Amnesty](/stories/denmark-udbetaling-danmark-atp-automated-welfare-fraud-algorithms-amnesty-mass-surveillance-2024/) — another national automated welfare-fraud system, showing the pattern is not unique to Michigan.

## Sources
1. IEEE Spectrum · *Michigan's MiDAS Unemployment System: Algorithm Alchemy Created Lead, Not Gold* · Robert N. Charette · 2017 (updated 2018) · https://spectrum.ieee.org/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold — **Tier 2** (independent technology journalism; one of two independent secondaries reporting the 40,195 auto-adjudicated cases, the 85% error rate, the 22,589 human-reviewed cases at 44%, and the 400% penalty, matching the state's own completed review at Source 6; Wayback 20260620103624, saved to sources/ieee-spectrum-midas.html).
2. Bridge Michigan · *Broken: The human toll of Michigan's unemployment fraud saga* · 2019 · https://bridgemi.com/michigan-government/broken-human-toll-michigans-unemployment-fraud-saga/ — **Tier 2** (independent nonprofit newsroom; carries the "internal state review of more than 20,000 fraud claims showed a 93 percent computer error rate" and individual harm detail; Wayback 20260215151947, saved to sources/bridgemi-broken-human-toll.html).
3. The Markup · *The Seven-Year Struggle to Hold an Out-of-Control Algorithm to Account* · 20 August 2022 · https://themarkup.org/newsletter/hello-world/the-seven-year-struggle-to-hold-an-out-of-control-algorithm-to-account — **Tier 2** (independent investigative newsroom; interview with plaintiffs' counsel corroborating the 93% error rate, the ~40,000 wrongful findings, the July 2022 Supreme Court ruling, and the separate $20.8 million penalty-refund figure shown here not merged with the $20 million settlement; Wayback 20260817231330, saved to sources/themarkup-midas.html).
4. Michigan Supreme Court · *Bauserman v Unemployment Insurance Agency* (opinion) · 26 July 2022 · https://www.courts.michigan.gov/4a1c19/siteassets/case-documents/opinions-orders/msc-term-opinions-(manually-curated)/21-22/bauserman-op.pdf — **Tier 1** (primary court record; the automated-process-without-a-hearing description, the holding that plaintiffs stated a constitutional-tort claim for monetary damages, and the recital that "approximately 93% of the automated system's fraud determinations were incorrect"; captured to sources/bauserman-msc-opinion.pdf and archived Wayback 20260713223317; the PDF has a full text layer and was machine-extracted verbatim this round).
5. Bridge Michigan · *Michiganders falsely accused of jobless fraud to share in $20M settlement* · 31 January 2024 · https://bridgemi.com/michigan-government/michiganders-falsely-accused-jobless-fraud-share-20m-settlement/ — **Tier 2** (independent nonprofit newsroom; the $20 million Court of Claims settlement and the 3,000-plus class size; Wayback 20260801091826, saved to sources/bridgemi-20m-settlement.html).
6. State of Michigan, Talent Investment Agency / Unemployment Insurance Agency · *Michigan's unemployment agency completes review of fraud determination cases; comprehensive changes underway to improve customer service and operations* (press release) · 2017 · https://www.michigan.gov/som/0,4669,7-192-47796-428651--,00.html — **Tier 1 (first party, admission against interest)** (the state's own official published account of re-examining 62,784 penalty cases: "40,195 were originally resolved by way of computer program based on available information ... 85 percent of these original fraud findings were reversed" and "22,589 ... were initiated by computer program and then referred to an investigator ... 44 percent of these fraud findings were reversed"; the live page is now access-denied, so it is quoted from the Wayback capture 20190111094548, saved to sources/michigan-uia-completed-review-2017.html; used only as an already-public record, never a chased confirmation). The Michigan Office of the Auditor General's 2016 MiDAS audit (sources/michigan-auditor-general-midas-2016.pdf) was decrypted and read and does NOT contain these figures, so it is not cited.
7. Detroit Free Press · *After Falsely Accusing Thousands of Unemployment Fraud and Wrongly Taking Their Money, Michigan Makes Amends. But Is It Enough?* · Paul Egan · 15 August 2017 · https://www.governing.com/topics/health-human-services/tns-michigan-unemployment-fraud-accusations.html — **Tier 2** (independent newspaper journalism, the second independent secondary for the exact split: "40,195 were robo-adjudicated by MiDAS with no human involvement and those had an 85% error rate" and "Another 22,589 cases involving some human interaction had a 44% error rate, the agency said"; the freep.com original is paywalled, so quoted from the Tribune News Service republication on governing.com, which carries the full wire copy; Wayback 20190808013428, saved to sources/detroit-free-press-egan-2017-governing.html).