Verification Laboratory for Machine-Governance Claims: Primary Sources, Quantitative Recalculation, Counterexamples, and Falsifiability

Full report

On this page

1. Executive summary

The integration of algorithmic systems into the apparatus of state and corporate governance represents a profound shift in the architecture of administrative law, criminal justice, and civic allocation. As public institutions increasingly substitute human adjudication with automated probability architectures, the foundational assumptions of due process, evidentiary standing, and constitutional protections are subjected to extreme stress. The analysis indicates that while machine-driven processing scales the enforcement capacity of the state, it simultaneously introduces systemic vulnerabilities—specifically regarding base-rate neglect, automation complacency, and the downstream multiplication of unreviewed errors. This standalone verification report independently audits twenty conceptual frameworks and twenty legal claims surrounding machine judgment, false positives, and institutional incentives. The verification architecture strictly segregates binding legal authority and peer-reviewed empirical phenomena from heuristic simulations, conceptual metrics, and normative proposals. By mapping claims against controlling jurisprudence—spanning the European Court of Justice (SCHUFA Holding)1, the Supreme Court of Canada (Ewert v. Canada)3, and United States federal and state courts (Mobley v. Workday, K.W. v. Armstrong, State v. Loomis)5—the analysis establishes the precise boundaries of current algorithmic accountability. Ultimately, this report demonstrates that sweeping, uniform declarations regarding machine governance (e.g., "algorithms inherently violate due process") are legally and factually insupportable. Courts have fractured aggressively: some have embraced proprietary risk-assessment tools by shielding them behind trade-secret privileges, while others have invalidated automated models for violating fundamental human rights to transparency and proportionality. To ensure the credibility of public content, frameworks exploring machine governance must rigorously maintain the distinction between verified institutional practice and speculative systemic risk.

2. Method and evidence hierarchy

To effectively evaluate complex, interdisciplinary claims regarding machine judgment, sources must be strictly stratified based on their binding authority, empirical rigor, and epistemological weight. Algorithmic governance currently suffers from an over-reliance on marketing literature, vendor white papers, and theoretical sociology, which are frequently conflated with established law or technical reality. The evaluation framework relies on a stringent source-quality hierarchy. Controlling legal authority—defined as binding statutory law, finalized appellate jurisprudence, and active regulatory regimes—serves as the absolute baseline for determining what the law currently enforces. Official government records, including inspector general audits, consent decrees, and administrative rulings, are utilized to document confirmed institutional practices and failures. Lower-tier sources, such as vendor claims or professional journalism, are acceptable solely for illustrating industry trends, mapping emerging controversies, or establishing the existence of a marketed capability; they are fundamentally insufficient for establishing technical accuracy, compliance with civil rights laws, or empirical realities. The application of this hierarchy ensures that conceptual mechanisms designed to simulate theoretical state behavior are not improperly elevated to the status of recognized legal doctrines.

Source-Quality Table

3. Concept verification ledger

The analysis subjects twenty conceptual frameworks regarding machine judgment to independent verification. This process requires determining whether a concept is an established legal or sociological term, a mathematically validated metric, or an illustrative heuristic designed to simulate theoretical risk. The concept of "actuarial justice," for instance, is a highly recognized and widely cited framework within sociological and criminological literature, established by Malcolm Feeley and Jonathan Simon in 199212. It effectively describes the shift from individualized, rehabilitative adjudication to the statistical, demographic management of populations based on aggregate risk profiles. Similarly, "automation bias" and "precautionary deference" are established phenomena in human-factors engineering, describing the psychological tendency of human operators to defer to machine-generated outputs, particularly in high-volume or high-stress environments19. However, the majority of the concepts advanced in the project—such as the "Judgment Velocity Ratio," "Judgment Burden Index," and "Cascading Denial Velocity"—are project-created conceptual terms. They do not exist in statutory law, appellate jurisprudence, or peer-reviewed empirical literature. They function as illustrative heuristics and simulation tools designed to model how systemic failures compound when automated logic is applied to civic infrastructure. For example, the "Cascading Denial Velocity" effectively models the theoretical danger of interconnected databases sharing unverified flags, resulting in simultaneous denials of service across housing, employment, and social welfare6. While valuable for normative risk assessment, presenting these conceptual metrics as established, observable laws of governance undermines the credibility of the analysis. They must be explicitly framed as theoretical frameworks or simulation tools. Furthermore, several concepts operate as normative proposals—arguments for how the law ought to adapt to machine governance. The "Machine Presumption of Innocence Protocol" and "Cognitive Liberty Firewall" represent advocacy positions urging the creation of new legal barriers against preemptive algorithmic surveillance and neuro-tracking20. They should not be presented as currently functioning legal protections, as contemporary jurisprudence frequently permits the integration of probabilistic modeling in law enforcement and administrative arenas, provided minimum procedural safeguards are met7.

Twenty-Concept Verification Ledger

4. Legal claim audit

The integration of opaque, automated decision-making systems into the civic sphere has generated profound jurisdictional fragmentation. An exhaustive audit of twenty high-value legal claims demonstrates that global courts are actively diverging on how to handle machine judgment. In the United States, courts have frequently permitted the use of proprietary algorithms in criminal proceedings without requiring source-code disclosure. For example, in People v. Wakefield, the New York Court of Appeals held that the TrueAllele probabilistic DNA genotyping system was generally accepted and that the software's source code did not constitute a "witness" subject to the Sixth Amendment's Confrontation Clause22. Similarly, the Wisconsin Supreme Court in State v. Loomis upheld the use of the proprietary COMPAS risk assessment tool at sentencing, provided the trial court was given warnings about its limitations and did not use the score as the sole determinative factor7. Conversely, in contexts involving administrative benefits, employment, and human rights, courts and regulators have adopted far more stringent postures against automated opacity. In K.W. v. Armstrong, the Ninth Circuit affirmed an injunction against the Idaho Department of Health and Welfare for utilizing a proprietary algorithmic tool to unilaterally reduce Medicaid budgets, ruling that such reductions without sufficient explanation and the ability to challenge the underlying formula violated procedural due process6. In Australia, the catastrophic "Robodebt" scheme—which unlawfully utilized averaged income data from the taxation office to automatically generate massive debt notices—was struck down in Amato v. Commonwealth after the government conceded the algorithmic mechanism was fundamentally unlawful and reversed the burden of proof onto impoverished citizens21. Regulatory interventions are also expanding liability for machine decision-making. The European Court of Justice recently ruled in the SCHUFA Holding case that the generation of a credit probability score constitutes "automated individual decision-making" under GDPR Article 22 if third-party lenders rely decisively on that score, severely limiting the ability of rating agencies to operate opaque algorithms1. Domestically, the EEOC secured a landmark consent decree in EEOC v. iTutorGroup, enforcing age discrimination penalties against an employer whose automated recruitment software systematically and silently rejected applicants over specific ages9. Furthermore, the Federal Trade Commission has begun deploying "algorithmic destruction"—as seen in In re Everalbum—requiring companies to completely delete machine learning models trained on improperly obtained data, targeting the "algorithmic shadow" left by initial misdeeds36.

Twenty-Claim Legal Authority Ledger

5. Quantitative recalculation laboratory

Evaluating the risks of algorithmic governance requires rigorous mathematical recalculation of the dynamics surrounding sensitivity, specificity, and base-rate neglect. Public perception—and frequently, judicial reasoning—routinely equates a "90% accurate" system with a "10% false positive rate," fundamentally misunderstanding the mathematical reality of Bayes' theorem. When predictive models are applied to rare events, the Positive Predictive Value (PPV) plummets, resulting in massive populations of false positives despite high technical sensitivity57.

Quantitative Assumption Table

Recalculation 1: Rare-Event Prediction and Base-Rate Neglect

When algorithms flag rare events (e.g., welfare fraud, terrorism, specific pre-crimes like the alerts generated by ShotSpotter networks28), the Positive Predictive Value (the probability that a flagged individual is actually guilty) crashes, regardless of how impressive the system's sensitivity appears in a laboratory. Formula: [Figure omitted from source export] (Where False Discovery Rate (FDR) \= [Figure omitted from source export]) Sensitivity Cases:

• Case A (High noise, highly rare event): [Figure omitted from source export], [Figure omitted from source export], [Figure omitted from source export].
• [Figure omitted from source export] (1.77%)
• Result: 98.23% of flagged individuals are entirely innocent (FDR). The algorithmic suspicion is overwhelmingly misdirected.
• Case B (Low noise, highly rare event): [Figure omitted from source export], [Figure omitted from source export], [Figure omitted from source export].
• [Figure omitted from source export] (8.26%)
• Result: Even with an exceptional 99% specificity, 91.74% of flagged individuals are innocent.
• Case C (Moderate noise, common event): [Figure omitted from source export], [Figure omitted from source export], [Figure omitted from source export].
• [Figure omitted from source export] (48.65%)
• Result: When the prevalence of the event rises to 5%, the system begins to approach a coin-flip reliability (51.35% FDR).

Assumptions: This calculation assumes binary classification and complete independence of the test from the base rate. It is critical to state that [Figure omitted from source export] is not an observed sociological fact, but merely an illustrative heuristic for modeling rare-event policing.

Recalculation 2: Cumulative Flag Probability under Repeated Testing

Algorithmic surveillance is rarely a single, isolated test. Citizens are scanned repeatedly across time and context (e.g., facial recognition networks across public transit, continuous welfare monitoring, automated resume scanning). The probability of an innocent person generating at least one false flag grows exponentially with repeated exposure. Formula: [Figure omitted from source export] (Where [Figure omitted from source export] is the number of independent algorithmic evaluations). Sensitivity Cases:

• Case A ([Figure omitted from source export]):
• [Figure omitted from source export] (9.56%)
• Case B ([Figure omitted from source export]):
• [Figure omitted from source export] (64.15%)
• Case C ([Figure omitted from source export]):
• [Figure omitted from source export] (99.48%)

Assumptions / Correlation Limitations: The formula strictly assumes independent trials. In real-world systems, false positives are often deeply correlated; an individual with a specific demographic trait, unusual name structure, or geographic zip code may trigger consistent false positives across different systems (as seen in the disparate impacts alleged against recruiting platforms in Mobley v. Workday5). If errors are perfectly correlated, the cumulative probability collapses back to the base FPR, though the harm becomes persistent and inescapable for that specific individual.

Recalculation 3: Downstream Error Multiplication and Appeal Latency

When a false positive triggers an automated denial of service—such as the arbitrary budget cuts seen in K.W. v. Armstrong6 or the aggressive debt collections in Amato v. Commonwealth21—the burden of proof shifts heavily to the citizen. If the human review process is infected by automation bias (precautionary deference), the final error rate multiplies, overriding intended procedural safeguards. Formula (Effective System Error): [Figure omitted from source export] (Where [Figure omitted from source export] is the deference rate: the probability a human reviewer blindly confirms the machine flag without substantive investigation). Sensitivity Cases (Assume initial [Figure omitted from source export]):

• Case A (High Independence, [Figure omitted from source export]): [Figure omitted from source export] (0.5%). The human acts as an effective filter.
• Case B (Moderate Automation Bias, [Figure omitted from source export]): [Figure omitted from source export] (2.5%). Half of the false positives pass through to enforcement.
• Case C (High Deference/Robodebt scenario, [Figure omitted from source export]): [Figure omitted from source export] (4.75%). The human-in-the-loop is merely a rubber stamp, allowing the vast majority of machine errors to inflict harm.

Appeal Latency Burden: The theoretical existence of an appeal mechanism does not guarantee due process if the scale of machine errors overwhelms human adjudicative capacity. [Figure omitted from source export] If an automated system instantly generates 10,000 algorithmic flags (as seen in aggressive fraud-detection algorithms), and the human appellate body processes 100 cases per month, [Figure omitted from source export] for the 10,000th citizen is 100 months. In such scenarios, due process is theoretically extant but practically void, functionally operating as a denial of service.

6. Counterexamples and disanalogies

While the evaluation of concepts such as the "Judgment Burden Index" and "Cascading Denial Velocity" effectively highlights the systemic risks of machine governance, objective analysis requires documenting where these conceptual frameworks break down, misrepresent reality, or fail to account for countervailing systemic forces. A rigorous verification methodology requires establishing a registry of counterexamples.

Counterexample Registry

7. Falsifiability and uncertainty

A central failing in the contemporary public discourse surrounding algorithmic governance is the reliance on non-falsifiable claims. To maintain epistemological rigor, any conceptual framework advanced by research institutions or governance bodies must explicitly articulate the precise empirical conditions under which it would be considered invalid. The claim that "machine algorithms inherently violate due process" is non-falsifiable if "due process" is emotionally or subjectively defined as "the right to have a human look you in the eye." However, if due process is defined procedurally—as the right to timely notice, the right to review the underlying evidence, and the right to a neutral arbiter—then the claim becomes highly falsifiable. For instance, the ruling in State v. Loomis7 falsifies the absolutist claim that proprietary algorithms are per se unconstitutional at sentencing, even while Houston Federation of Teachers47 simultaneously supports the claim that entirely uninterpretable, opaque algorithms used for termination violate due process. The conflicting jurisprudence indicates that claims regarding machine judgment must be heavily caveated by jurisdiction and specific administrative context26. Furthermore, extreme caution must be exercised regarding data unavailability. The true False Positive Rate of commercial surveillance systems is frequently guarded as a highly lucrative trade secret59. Asserting that a proprietary system "has an unacceptably high error rate" without having access to the proprietary validation data constitutes an unsupported extrapolation. The proper, falsifiable claim is that "the system lacks publicly accessible validation data, permanently preventing the independent, adversarial verification of its error rate."

8. Public wording corrections

To prevent the dissemination of legally inaccurate, mathematically unsound, or jurisdictionally overstated claims, public-facing terminology must be ruthlessly precise. Broad extrapolations derived from single-jurisdiction trial court rulings or unverified sociological theories severely damage institutional credibility.

Public Wording Recommendation Table

'Do Not Overclaim' List

9. Claims requiring further research

The verification process has identified several critical gaps in both empirical data and legal doctrine that require further, extensive research before definitive claims can be established by policymakers or litigators:

• The Scope and Limits of Algorithmic Destruction: While the Federal Trade Commission utilized algorithmic destruction in In re Everalbum36 to force the disgorgement of a model built on improperly obtained photos, it remains unclear whether this remedy will be widely adopted by state attorneys general, or if it can be retroactively applied to predictive policing models built on historically contaminated arrest data.
• The Confrontation Clause in the AI Era: The New York Court of Appeals in People v. Wakefield22 determined that a proprietary algorithm's source code is not a "witness" subject to cross-examination. Further research is required to determine if other state supreme courts or federal circuits will adopt this interpretation, or if the Sixth Amendment will eventually be construed to encompass probabilistic genotyping software and other machine witnesses.
• Automation Complacency Metrics: While the EU AI Act (Article 14\) mandates "human oversight"51, there is insufficient peer-reviewed psychological data establishing the precise user-interface design requirements necessary to effectively override automation bias in high-volume, low-margin bureaucratic settings.
• Applicant Standing in AI Discrimination: As demonstrated in Mobley v. Workday40, the application of disparate impact protections to job applicants (as opposed to current employees) under the ADEA remains an actively litigated frontier. The final resolution of these claims will dictate the liability scope of major HR software vendors.

Web-Ready Content Package

How to Read Evidence States (150 words)

When evaluating claims about algorithmic governance, it is critical to understand the precise epistemological weight of the statement. Not all claims are established facts. We categorize statements using strict evidence states to maintain clarity. A "Verified Fact" is supported by binding legal authority, empirical consensus, or documented administrative records. A "Supported Interpretation" represents a highly plausible conclusion drawn from existing data, though differing expert opinions may naturally exist. A "Conceptual Framework" or "Simulation Tool" is an illustrative model—like a mathematical equation—designed to explain complex systemic behaviors, but it is not a historical event. A "Normative Proposal" is an argument for how the law or society should operate, rather than how it currently does. Finally, a "Disputed Claim" lacks sufficient empirical backing or contradicts established authority. Understanding these tags prevents the dangerous conflation of theoretical modeling with binding legal reality.

Evidence-State Glossary

• Verified Fact: A claim substantiated by Tier 1 or Tier 2 sources (binding law, government records) or overwhelming peer-reviewed consensus.
• Supported Interpretation: A logical deduction from verified facts that remains open to alternative expert analysis.
• Conceptual Framework: An analytical lens or metric created to evaluate a system (e.g., Judgment Velocity Ratio). It is valid as an exploratory tool, not as a historical event.
• Simulation Tool: A mathematical or logical model demonstrating potential outcomes based on stated assumptions.
• Normative Proposal: An advocacy position outlining proposed policy, legal reforms, or ethical standards.
• Disputed Claim: An assertion that contradicts controlling authority or lacks empirical validation, requiring heavy revision or removal.

10 FAQs About Verification and Uncertainty

1. Why can't we just call algorithmic bias illegal? Legal liability depends heavily on jurisdiction, standing, and specific statutory frameworks (e.g., Title VII, ADEA). Blanket statements of absolute illegality ignore complex judicial realities.

2. Does an AI have a high error rate if it is 99% accurate? Yes, if the event it detects is extremely rare. Due to base-rate neglect, a 99% accurate system scanning 1,000,000 people for 100 actual infractions will generate roughly 10,000 false positives.

3. Did a court rule that Robodebt was illegal? Yes. In Amato v. Commonwealth, the Federal Court of Australia issued consent orders declaring that calculating debts solely by averaging ATO income data was fundamentally unlawful13.

4. Do humans always blindly trust algorithms? No. Automation bias is real, but in high-liability environments, users sometimes exhibit "algorithm aversion," actively rejecting valid machine output due to mistrust19.

5. Is an algorithm's source code available during a criminal trial? Usually no. In cases like People v. Wakefield, courts have ruled that defendants do not have a constitutional right to examine proprietary source code under the Confrontation Clause22.

6. What is Actuarial Justice? A sociological framework popularized by Feeley and Simon (1992) describing the shift in criminal justice from individual rehabilitation to the statistical risk management of demographic populations12.

7. Can you sue a company if their AI rejects your job application? Potentially. Cases like Mobley v. Workday are currently testing whether AI software vendors can be held liable directly as "agents" under federal anti-discrimination laws5.

8. What does the GDPR say about automated decisions? Article 22 generally grants individuals the right not to be subject to a decision based solely on automated processing if it produces legal or significant effects, though exceptions exist for consent and contract necessity1.

9. How do algorithms violate the APA? If a government agency adopts a new algorithmic methodology for determining welfare benefits without undergoing proper public notice and comment procedures, it can be enjoined, as seen in Arkansas DHS v. Ledgerwood56.

10. What is Algorithmic Destruction? A powerful regulatory enforcement tool where an agency (like the FTC) orders a company to completely delete machine learning models that were trained on improperly obtained data36.

10 Examples of Safer Public Wording

1. "The integration of algorithmic decision-making frequently strains traditional administrative due process."

2. "Under specific conditions, statistical risk assessments can conflict with principles of individualized suspicion."

3. "The Dutch court in SyRI ruled that the specific fraud detection system violated ECHR Article 8 proportionality requirements."

4. "Bayesian mathematics dictate that highly sensitive algorithms will still produce massive false-positive volumes when applied to rare events."

5. "The EU AI Act classifies certain predictive policing models as prohibited practices."

6. "When human reviewers are subjected to severe time constraints, the risk of automation complacency increases."

7. "The use of proprietary risk assessment algorithms limits traditional adversarial testing."

8. "If errors multiply across interconnected administrative databases, the burden of correction falls heavily on the citizen."

9. "Algorithmic hiring tools have been subjected to regulatory scrutiny by the EEOC regarding potential disparate impacts."

10. "The K.W. v. Armstrong decision affirmed that property interests in Medicaid benefits require adequate notice before algorithmic reduction."

10 Examples of Wording That Should Be Avoided

1. "Algorithms are completely illegal to use in hiring."

2. "The Supreme Court has banned predictive policing."

3. "Machines are inherently racist."

4. "The algorithm is 99% accurate, so anyone it flags is definitely guilty."

5. "Artificial intelligence has eliminated the presumption of innocence."

6. "You have a constitutional right to see the source code of any machine that accuses you."

7. "Robodebt proved that all government algorithms are unlawful."

8. "Automation bias means human judges are entirely irrelevant."

9. "The algorithm deliberately targeted minorities."

10. "Data privacy laws make algorithmic surveillance impossible."

Compact Source-Quality Guide

• Tier 1 (Binding Law): Supreme Court rulings, finalized appellate decisions, statutes. Use to state what the law is.
• Tier 2 (Official Records): Consent decrees, Inspector General audits. Use to state what institutions did.
• Tier 3 (Academic & Technical): Peer-reviewed journals, NIST standards. Use to state empirical facts and sociological consensus.
• Tier 4 (Journalism & Commentary): Investigative reports, legal blogs. Use to show public impact, controversy, or theoretical proposals.
• Tier 5 (Vendor Claims): Proprietary marketing. Use only to describe what a system is marketed to do, never to verify its actual efficacy.

Proposed Visual Legend

• 🟢 Verified Fact: Solid Green Circle (Established by Tier 1-3 sources).
• 🔵 Supported Interpretation: Blue Square (Logical extrapolation based on verified data).
• 🟣 Simulation / Conceptual Tool: Purple Diamond (Mathematical models, heuristic frameworks).
• 🟡 Normative Proposal: Yellow Triangle (Advocacy, policy recommendations).
• 🔴 Disputed / Unverified: Red Octagon (Contradicts established law/data, requires revision).

Works cited

1. SCHUFA Case C-634/21: GDPR Implications \- Scribd, https://www.scribd.com/document/850160515/1-JusGov-Research-Paper-Series-Schufa-Case-C-634-21-on-ADM

2. SCHUFA Holding and Others (Scoring) \- EUR-Lex, https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:62021CJ0634

3. Who Is Worthy of Constitutional Protection? A Commentary on Ewert, https://research.schulichlaw.dal.ca/en/publications/who-is-worthy-of-constitutional-protection-a-commentary-on-ewert-/

4. Ewert v. Canada \- SCC Cases, https://decisions.scc-csc.ca/scc-csc/scc-csc/en/item/17133/index.do

5. DEREK MOBLEY v. WORKDAY INC (2024) \- FindLaw Caselaw, https://caselaw.findlaw.com/court/us-dis-crt-n-d-cal/116378658.html

6. K.W. v. Armstrong, No. 14-35296 (9th Cir. 2015\) \- Justia Law, https://law.justia.com/cases/federal/appellate-courts/ca9/14-35296/14-35296-2015-06-05.html

7. CRIMINAL LAW: Use of Risk Assessment Tools in Sentencing, https://www.nlrg.com/legal-content/the-lawletter/criminal-law-use-of-risk-assessment-tools-in-sentencing-upheld-.-.-.-for-now

8. SCHUFA case | OQ v Land Hessen | Automated decision-making, https://www.michalsons.com/blog/schufa-case-oq-v-land-hessen-automated-decision-making/76054

9. EEOC v. iTutorGroup, Inc. Case Study: AI Hiring Discrimination, https://www.forensisgroup.com/resources/expert-legal-witness-blog/eeoc-v-i-tutor-group-inc-case-study-ai-hiring-discrimination-automated-recruiting-bias-and-age-discrimination-enforcement

10. iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit, https://www.eeoc.gov/newsroom/itutorgroup-pay-365000-settle-eeoc-discriminatory-hiring-suit

11. Disparate Impact in AI Accountability Sources, https://civilrights.org/disparate-impact-ai-sources/

12. Actuarial Justice \- SozTheo, https://soztheo.com/glossary/actuarial-justice/

13. Code vs Code: How is Artificial Intelligence Reshaping Planning Law?, https://wallmans.com.au/news-resources/code-vs-code-how-is-artificial-intelligence-reshaping-planning-law/

14. the dangers of AI-generated evidence and its ethical implications, https://www.ibanet.org/actual-intelligence-v-artificial-intelligence-dangers-of-ai-generated-evidence

15. UNCHECKED POLICE USE OF ALGORITHMIC TECHNOLOGIES, https://digitalcommons.law.scu.edu/cgi/viewcontent.cgi?article=1697\&context=chtlj

16. Reference Guide on Forensic Feature Comparison Evidence \- NCBI, https://www.ncbi.nlm.nih.gov/books/NBK621601/

17. CJEU, 7 December 2023, Schufa Holding (Scoring), Case C-634/21, https://www.julia-project.eu/database/case-law/216

18. Critical Issues Criminal Justice | LawTeacher.net, https://www.lawteacher.net/free-law-essays/criminology/critical-issues-in-criminal-justice.php

19. The rule of reason is being rebuilt, not hollowed out \- Law News, https://lawnews.nz/administrative-public/the-rule-of-reason-is-being-rebuilt-not-hollowed-out/

20. Cothron v. White Castle System, Inc. \- Illinois Case Law, https://law.justia.com/cases/illinois/supreme-court/2023/128004.html

21. An in-depth look at our robo-debt test case | Victoria Legal Aid, https://www.legalaid.vic.gov.au/depth-look-our-robo-debt-test-case

22. Can you Depose Artificial Intelligence?, https://coverage.memberclicks.net/assets/LawSymposium-2024-UofMinnesota/ACCC\_2024InsLawSymp\_MN\_Papers\_Nov15\_1320\_MockOralArg\_CanYouDeposeAI\_20241111.pdf

23. 2023- TGG- CASE STUDY-From Courtrooms to Algorithms \- HCLA, https://www.hclanl.com/wp-content/uploads/2023/11/2023-PART-ONE-From-Courtrooms-to-Algorithms.pdf

24. Technology is never neutral: Robodebt and a human rights analysis, https://www.tandfonline.com/doi/full/10.1080/1323238X.2024.2409620

25. Scores as Decisions? Article 22 GDPR and the Judgment of the, https://academic.oup.com/ilj/article/53/4/840/7745471

26. The Automated Administrative State: A Crisis of Legitimacy, https://scholarlycommons.law.emory.edu/cgi/viewcontent.cgi?article=1418\&context=elj

27. Centrelink's Big Data 'Robo-Debt' Fiasco of 2016-20 \- Roger Clarke, http://www.rogerclarke.com/DV/CRD17.html

28. ShotSpotter Gunshot Detection Technology, Predictive Policing, and, https://repository.law.umich.edu/cgi/viewcontent.cgi?article=2561\&context=mjlr

29. “Bang\!”: ShotSpotter Gunshot Detection Technology, Predictive, https://mjlr.org/journal/bang-shotspotter-gunshot-detection-technology-predictive-policing-and-measuring-terrys-reach/

30. SENATE SENATE MAJORITY LEADER v. HOUSE SPEAKER (2025), https://caselaw.findlaw.com/court/mi-court-of-appeals/117880045.html

31. PEOPLE v. WAKEFIELD (2022) | FindLaw, https://caselaw.findlaw.com/court/ny-court-of-appeals/2170317.html

32. Should Risk-Assessment Technology Be Part Of Our Criminal, https://www.litedepalma.com/should-riskassessment-technology-be-part-of-our-criminal-justice-system

33. K W et al v. Armstrong et al, No. 1:2012cv00022 \- Justia Law, https://law.justia.com/cases/federal/district-courts/idaho/iddce/1:2012cv00022/29058/596/

34. Chapter 3 – Parliament of Australia, https://www.aph.gov.au/Parliamentary\_Business/Committees/Senate/Community\_Affairs/Centrelinkcompliance/Second\_Interim\_Report/section?id=committees%2Freportsen%2F024338%2F27760

35. The European Court of Justice (CJEU) holds that credit scoring can, https://hellodpo.com/the-european-court-of-justice-cjeu-holds-that-credit-scoring-can-be-an-automated-decision-under-gdpr/

36. Algorithmic Destruction \- SMU Scholar, https://scholar.smu.edu/cgi/viewcontent.cgi?article=4930\&context=smulr

37. Liability Rules and Standards, https://www.ntia.gov/issues/artificial-intelligence/ai-accountability-policy-report/using-accountability-inputs/liability-rules-and-standards

38. Ewert v. Canada: Shining Light on Corrections and Indigenous People, https://digitalcommons.osgoode.yorku.ca/cgi/viewcontent.cgi?article=1388\&context=sclr

39. Automated facial recognition and policing: A Bridge too far?, https://research-portal.uea.ac.uk/en/publications/automated-facial-recognition-and-policing-a-bridge-too-far/

40. The Algorithm Is on Trial: What the Workday Ruling Means for Every, https://integralrecruiting.com/icims-consultant-mobley-workday-ai-hiring/

41. Mobley v. Workday, Inc., No. 3:2023cv00770 \- Document 267 (N.D., https://law.justia.com/cases/federal/district-courts/california/candce/3:2023cv00770/408645/267/

42. ECLI:NL:RBDHA:2020:1878, Rechtbank Den Haag, C ... \- Uitspraken, https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:RBDHA:2020:1878

43. Nederlands Juristen Comité voor de Mensenrechten et al. v. The, https://www.escr-net.org/caselaw/2020/nederlands-juristen-comite-voor-mensenrechten-et-al-v-netherlands-eclinlrbdha20201878/

44. PEOPLE v. GUTIERREZ (2023) \- FindLaw Caselaw, https://caselaw.findlaw.com/court/ny-supreme-court/2186451.html

45. Case v. Montana Attempts to Clarify the Emergency Aid Exception to, https://nccriminallaw.sog.unc.edu/2026/02/02/case-v-montana-attempts-to-clarify-the-emergency-aid-exception-to-the-warrant-requirement/

46. Antiracism in Action \- Gretchen, https://gretchen.law.nyu.edu/cgi/viewcontent.cgi?article=2463\&context=fac-articles

47. Validity First: The Politics, Governance, and Consequences ... \- MDPI, https://www.mdpi.com/2227-7102/16/5/766

48. Sorting Teachers Out (Chapter 10\) \- Money, Power, and AI, https://www.cambridge.org/core/books/money-power-and-ai/sorting-teachers-out/EA29DD9114CBB80B89742FA729486346

49. EL PASO COUNTY, COLORADO PERSONNEL POLICIES MANUAL, https://www.agendasuite.org/iip/elpaso/file/getfile/52975

50. Colorado SB 26-189 — Automated Decision-Making Technology in, https://docs.modulos.ai/frameworks/colorado-sb189

51. A Framework for Compliance with Regulation (EU) 2024/1689 for, https://www.mdpi.com/2624-800X/5/3/40

52. Regulation (EU) 2024/1689 of the European Parliament and of the, https://www.ndfr.nl/content/CELEX-02024R1689-20260727-ENG

53. Law / proposed law in the European Union \- AI Laws of the World, https://intelligence.dlapiper.com/artificial-intelligence/?t=01-law\&c=EU

54. Regulation (EU) 2024/1689 of the European Parliament ... \- EUR-Lex, https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=OJ:L\_202401689

55. Arkansas Department of Human Services v. Ledgerwood :: 2019, https://law.justia.com/cases/arkansas/supreme-court/2019/cv-18-639.html

56. ARKANSAS DEPARTMENT OF HUMAN SERVICES, Appellant v, https://caselaw.findlaw.com/court/ar-supreme-court/1992988.html

57. unknown\_url

58. California Federal Court Clarifies Limits On AI Bias Testing And, https://blogs.duanemorris.com/classactiondefense/2026/06/02/california-federal-court-clarifies-limits-on-ai-bias-testing-and-applicant-data-disclosure-in-mobley-v-workday/

59. Rebecca Wexler \- Senate Committee on the Judiciary, https://www.judiciary.senate.gov/download/2024-01-24-pm-testimony-wexler

60. Misguided Governments, Flawed Technologies, and Social Control, https://ir.lawnet.fordham.edu/cgi/viewcontent.cgi?article=2758\&context=ulj