Skip to content
AltonHenley

The Knowledge Barrier – Failures

The Knowledge Barrier

Documented AI Failures in Government

Appendix C — Detailed documentation of government AI failure cases analyzed in Chapter 9, with PEARS framework analysis for each case.

Back to The Knowledge Barrier

The PEARS Framework

The Tony Blair Institute for Global Change (2024) proposes that AI decisions in government should be:

Predictable — Consistent and foreseeable outcomes Explainable — Decision logic understandable to affected parties Accountable — Clear responsibility for AI outcomes Reversible — Ability to correct errors Sensitive — Appropriate to the stakes and context

COMPAS Recidivism Algorithm (2016)

Criminal Justice & Public Safety

What happened

ProPublica’s 2016 analysis of over 7,000 defendants in Broward County, Florida revealed racially disparate outcomes from the COMPAS risk assessment tool. Black defendants were nearly twice as likely as white defendants to be incorrectly flagged as high-risk (false positive rate: 44.9% vs. 23.5%). White defendants were more likely to be incorrectly flagged as low-risk (false negative rate: 47.7% vs. 28.0%).

Institutional failure pattern

Deployed without pre-deployment testing for racial disparities. Methodology protected as a trade secret, preventing independent validation. Jurisdictions adopted the tool based on vendor representations without technical capacity to evaluate those claims.

P — FailedE — FailedA — FailedR — PartialS — Failed
Angwin et al. (2016); Cummings (2025)

Detroit Facial Recognition Wrongful Arrest (2020)

Criminal Justice & Public Safety

What happened

Robert Williams, a Black man, was wrongfully arrested after Detroit police relied on a faulty facial recognition match. Williams was held for 30 hours and interrogated. The technology had documented accuracy disparities across demographic groups—Buolamwini and Gebru (2018) found error rates up to 34 percent higher for dark-skinned faces.

Institutional failure pattern

Police treated a low-confidence algorithmic match as sufficient probable cause without independent corroboration. No protocol existed for verifying facial recognition results before making arrests.

P — FailedE — PartialA — FailedR — PartialS — Failed
Hill (2020); Buolamwini & Gebru (2018); Cummings (2025)

LAPD PredPol Predictive Policing (2013–2019)

Criminal Justice & Public Safety

What happened

PredPol generated geographic predictions of likely crime locations for the LAPD from 2013 to 2019. The system led to over-surveillance of Black and Latino communities, directing patrol resources based on historical arrest rates rather than actual crime distribution. An LAPD internal audit found “insufficient data to determine effectiveness.” Discontinued in 2019.

Institutional failure pattern

Trained on arrest data reflecting police deployment decisions, creating a feedback loop: more arrests generated more predicted crime, which directed more patrol. Six years of use without independent effectiveness evaluation.

P — FailedE — PartialA — FailedR — FailedS — Failed
Puente (2019); Brayne (2020); Lum & Isaac (2016)

Indiana Welfare Modernization (2007–2009)

Benefits & Social Services

What happened

IBM’s $1.3 billion automated welfare eligibility system replaced in-person caseworker interviews with document-based processing. Application denials increased 54 percent in the first year. Approximately one million applications were denied during the contract period, many for minor technical errors. Indiana terminated the contract.

Institutional failure pattern

Designed to increase efficiency without adequate consideration of accuracy or human consequences of errors. Individual caseworker discretion—the safety valve for system errors—was eliminated.

P — FailedE — FailedA — FailedR — PartialS — Failed
Eubanks (2018); Shorey (2025)

Allegheny County Child Welfare Algorithm (2016–present)

Benefits & Social Services

What happened

The Allegheny Family Screening Tool (AFST) generates risk scores from county data to assess families reported to the child abuse hotline. Despite published methodology and independent ethical review—Eubanks calls it “the best-case scenario”—the system systematically disadvantaged poor families and families of color. It measures visibility to government rather than actual risk. The system calculated mandatory-investigation scores for 32% of Black children referred for neglect vs. 21% of white children.

Institutional failure pattern

Demonstrates that the proxy-variable problem is structural, not fixable through better governance. Government data measures government contact, which correlates with poverty and race.

P — PartialE — MetA — PartialR — FailedS — Partial
Eubanks (2018); Ho et al. (2022)

Idaho Medicaid Budget Tool (2011)

Benefits & Social Services

What happened

An algorithm used to determine home-care hours for Medicaid recipients with developmental disabilities produced drastic, unexplained cuts—some by more than 40 percent. Affected individuals could not understand why their benefits changed. A federal court ruled the system violated due process rights.

Institutional failure pattern

Implemented for consequential benefit determinations without ensuring outputs could be explained to affected individuals or independently reviewed.

P — FailedE — FailedA — FailedR — PartialS — Failed
Lecher (2018); K.W. v. Armstrong (D. Idaho 2016); Cummings (2025)

Arkansas Medicaid Algorithm (2016)

Benefits & Social Services

What happened

An algorithm assessing home-care needs produced unexplained cuts for disabled residents. One plaintiff saw weekly hours cut from 56 to 32 with no explanation. Legal advocates discovered coding errors producing inconsistent results. A federal court ruled Arkansas violated due process rights.

Institutional failure pattern

Deployed for consequential decisions without ensuring explainability, accuracy, or meaningful appeal processes. Coding errors found only through litigation.

P — FailedE — FailedA — FailedR — PartialS — Failed
Lecher (2018); Jacobs v. Gillespie (E.D. Ark. 2018)

NYC MyCity Chatbot (2023–2024)

Resident-Facing AI

What happened

New York City’s generative AI chatbot, launched in October 2023, provided illegal advice on multiple occasions. When asked “Can I take a cut of my worker’s tips?” it answered “yes”—violating New York labor law. Other errors included incorrect housing regulation and business licensing information.

Institutional failure pattern

Deployed a generative AI system for public-facing use without adequate verification mechanisms. Unlike rule-based chatbots, generative systems produce novel responses that may be factually wrong despite sounding authoritative.

P — FailedE — PartialA — PartialR — FailedS — Failed
Caldera (2024); Chaurasia et al. (2025); Prinvil et al. (2025)

Common Failure Patterns

Across all documented cases, five institutional failure patterns recur.

PatternCases
Inadequate pre-deployment testingCOMPAS, Detroit facial recognition, NYC MyCity
Insufficient ongoing monitoringLAPD PredPol (6 years), Idaho/Arkansas Medicaid
Opacity about system operationsIdaho Medicaid, Arkansas Medicaid, COMPAS (trade secret), Allegheny AFST
Diffuse accountabilityAll cases
Disparate impact on vulnerable populationsCOMPAS (Black defendants), Detroit (Black residents), PredPol (Black and Latino communities), Allegheny (poor families, families of color), Indiana (eligible applicants in poverty), Idaho/Arkansas (disabled Medicaid recipients)

Table C.1: Recurring Institutional Failure Patterns in Government AI Deployments

The consistency of these patterns across different jurisdictions, domains, and time periods suggests they are structural features of how governments deploy algorithmic systems.

© 2026 Alton Henley. The Knowledge Barrier.
altonhenley.com · LinkedIn