The Legal Reckoning
in AI Hiring

The Bayesian Lens

Every modern AI system — every large language model, every recommender, every diagnostic algorithm — works the same way. It starts with an initial estimate. It collects evidence. It revises the estimate in proportion to how strong the evidence is. The mathematics has a name (Bayes' Theorem). The intuition does not need one. It is what every reasonable person already does without thinking about it: change your mind in proportion to how good the evidence is.

Hiring is the same problem. Before any screening begins, every candidate has some baseline likelihood of being a top performer in a given role. Each tool applied to that candidate — a resume screen, an interview, a reference check — is a piece of evidence that should raise or lower your confidence. The metric that quantifies how much each tool moves your confidence we're calling the Diagnostic Ratio (DR) for the purposes of this report. Think of it as the signal-to-noise ratio of a selection tool — a measure of how cleanly the tool separates strong candidates from weak ones. Mathematically, it is the positive Likelihood Ratio used in clinical diagnostics, forensic science, and intelligence analysis to weigh evidence under uncertainty. The DR measures how much trust the tool's signal deserves.

This is the lens through which the rest of this document should be read. The lawsuits, the regulatory landscape, the science, the legal validation — all of it is evidence that should raise or lower your confidence in one specific question: is the hiring methodology you are using right now defensible, on either scientific or legal grounds? The answer is determined by the DR of the tools the methodology uses. There is no separate legal analysis or scientific analysis. There is one analysis, looked at from two angles.

Companion documents that develop this framework in depth: The Diagnostic Ratio of Common Hiring Tools — full DR analysis of 20+ hiring tools, sourcing-channel effects on the talent pool, and the 2-Step screening process · The Science Says Your Hiring Process Doesn't Work — the parallel scientific briefing on the Sackett (2022) meta-analysis and the r² reality of selection methods.

Executive Summary

This analysis examines ten public actions involving algorithmic, AI-assisted, and credential-based hiring tools — class actions, EEOC actions, FTC investigations, settled matters, and one preemptive corporate withdrawal — and identifies the structural patterns that recur across the sample. The conclusion is reached not by litigating any individual matter but by observing what the matters share. Most actions in the sample trace back to the same upstream feature: a hiring tool used as a consequential gate without an underlying job-validated performance specification.

The science and the law converge on the same conclusion through different vocabularies. The categories of hiring tools that recur across the sample all fall at or below the noise floor of the Diagnostic Ratio framework — they look selective, but the underlying validity research shows they aren't. The full scientific case is set out in The Science Says Your Hiring Process Doesn't Work, drawing on Sackett, Zhang, Berry & Lievens (2022). The DR values below are best-estimate midpoints from that meta-analytic literature:

The same tools that the science describes as ineffective are the tools the law is now testing for defensibility. The convergence is structural, not coincidental — both legal exposure and Type II error loss trace back to the same upstream measurement failure.

This is not primarily a legal story. It is a measurement story with a legal consequence. Tools with no signal value have nothing to be related to job performance — which means they cannot satisfy the Uniform Guidelines on Employee Selection Procedures' job-relatedness requirement. Tools that randomly filter candidates produce the disparate-impact patterns the courts are now treating as actionable. Both failures follow from the same root: a tool deployed as if it had signal value it does not have.

The seven themes documented in Section 3 describe the recurring structural features across the sample. Section 4 explains why each theme is problematic in the same way — through the DR framework. Sections 5 through 8 lay out the legal and regulatory architecture that is testing these themes in court and in agency enforcement, both in the United States and across the EU and UK. Section 9 summarizes the DR profile of the tool categories implicated. Section 10 connects the analysis to the 2021 legal principles independently set out by David J. Goldstein of Littler Mendelson, which describe what a defensible methodology looks like — principles articulated before the post-2023 cases were filed.

An employer's AI hiring tool can face exposure on both tracks simultaneously. Winning on one provides no protection on the other.

The Diagnostic Reality Underneath Both Gaps

Both accountability tracks share an upstream cause that selection science has measured precisely. Every tool used in hiring has a measurable Diagnostic Ratio. The bands defined in The Bayesian Lens apply directly. The numbers below come from meta-analytic validity research, primarily Sackett, Zhang, Berry & Lievens (2022).

The actions documented in this whitepaper can be read as the legal system testing whether tools with these characteristics — applied as consequential screening gates without documented job-specific validation — can be defended under existing statutes. Whether any specific tool ultimately survives that test is a question for the courts. What can be observed independent of any individual ruling is that the alleged screening practices in these matters share a common feature: tools used as gates against criteria that were not validated against the role's actual performance requirements. The diagnostic exposure described above operates regardless of how the cases are decided, because the candidates rejected by low-DR tools never become employees, never become litigants, and never appear in any metric.

For the full Diagnostic Ratio analysis across 20+ hiring tools, the Bayesian sourcing-channel math, and the 2-Step screening process that closes the gap, see The Diagnostic Ratio of Common Hiring Tools.

The Sample: Ten Public Actions Examined

This analysis examines ten public actions involving algorithmic, AI-assisted, or credential-based hiring tools — a sample large enough to identify recurring patterns, varied enough to span multiple legal theories, and current enough to reflect the present state of enforcement and litigation.

The sample is drawn from federal class actions, EEOC enforcement actions, FTC investigations, ACLU complaints, state regulatory matters, and one preemptive corporate action. The legal theories invoked across the sample include disparate impact under Title VII, age discrimination under the ADEA, disability discrimination and accommodation under the ADA, consumer-report disclosure under the FCRA, state privacy and consumer-protection statutes, and Fair Housing Act vendor-as-agent theory applied by analogy. Status as of mid-2026 ranges from settled, to certified for class action, to in active litigation, to under investigation, to preemptively withdrawn from market.

The point of examining the sample is not to litigate any individual matter or to characterize the conduct of any named party beyond the public record. The point is to ask whether common structural patterns appear across the sample — patterns that, if present, would suggest the legal exposure being tested in these matters has a single underlying cause that any organization can examine in its own hiring methodology regardless of whether litigation has reached it.

The Sample at a Glance

Cluster 1 · Discriminatory Outcomes

Cluster 2 · Invisible Processes & FCRA Violations

Cluster 3 · Adjacent Cases Extending Vendor Liability Theory

Complete Case Summary

Each action is described in this analysis solely on the basis of public filings, court rulings, agency announcements, and contemporaneous reporting. Allegations recited in pleadings are allegations only and have not been adjudicated unless the table notes a settlement, withdrawal, or final ruling. Nothing in this analysis is intended to assert as fact any wrongdoing by any named party beyond what the cited public record supports. See full disclaimer.

Common Themes Across the Sample

Examining the ten actions together — without litigating any individual matter — reveals a small number of structural patterns that recur across the sample. None of these themes is novel. Each has been documented in the selection-science literature for decades. What is new is that all ten actions can be read as the legal system testing whether tools exhibiting these patterns can survive scrutiny under existing law.

Theme 1 · Screening against criteria not validated against the role's performance

Across most of the sample, the tools at issue applied gates — credentials, years of experience, degree requirements, keyword matches, personality scores, facial expressions, employment-history patterns — that the public filings indicate were not validated against documented, role-specific performance criteria. The criteria may be facially neutral. Whether they predict performance in the specific role is a separate question that, in the public record, generally was not answered before the tool was deployed as a screening gate.

Theme 2 · Black-box scoring without disclosure or recourse

Several actions in the sample involve algorithmic scoring that the candidate cannot see, contest, or correct. The candidate is filtered out, ranked low, or denied an opportunity on the basis of a score generated from inputs the candidate is not told about, against criteria the candidate cannot review, by a system the employer often cannot audit either. The FCRA theory in one matter and the ADA accommodation theory in another both turn on this opacity.

Theme 3 · Vendor-as-agent without vendor accountability

A subset of actions in the sample involves third-party tools performing functions courts have characterized — or plaintiffs have alleged — as employment-agency functions, while the vendor's contractual exposure is typically capped at subscription fees. Where this asymmetry holds, the employer carries the legal risk of a system the employer often cannot inspect, validate, or correct, while the vendor's risk is bounded by the contract.

Theme 4 · Training-data inheritance

Several of the actions involve AI systems trained on historical hiring or workforce data. Where that data reflects prior selection patterns — including patterns that produced demographic skew — models trained on it will reproduce the skew in their outputs, regardless of whether protected characteristics are explicitly used as inputs. The Amazon disclosure (the company's preemptive withdrawal of its internal AI recruiting tool in 2018) is the canonical example documented in the public record.

Theme 5 · Facially neutral inputs functioning as protected-class proxies

One action in the sample alleges that facially neutral variables — educational background, employment history, zip code — operate as proxies for race in their cumulative effect. The legal theory is not new; it descends from longstanding disparate-impact doctrine. What is new is its application to algorithmic systems that combine many neutral variables in ways the employer often cannot describe in human-readable terms. The June 2026 order in Mobley v. Workday illustrates the same proxy logic in the disability context, allowing a claim to proceed on the theory that employment gaps functioned as a proxy for disability status.

Theme 6 · Automated assessment of psychological or affective traits without job-specific validation

Several actions involve tools that infer personality traits, emotional states, "employability," or related psychological characteristics from voice, video, or response patterns. These instruments raise distinct legal issues under state psychological-testing statutes, the ADA's restrictions on medical inquiries, and disparate-impact doctrine — separate from but compounding the validity questions in the underlying I/O literature.

Theme 7 · Deployment without a human review path

Across multiple actions, the tools at issue operated as filters before any human review of the candidate's application. Candidates were rejected — sometimes within minutes of applying, sometimes silently — by automated systems whose outputs were treated as dispositive rather than advisory. This deployment posture has become a central legal issue: courts and plaintiffs have increasingly characterized such automated gates as decision-making rather than recommendation, with corresponding liability consequences for both vendor and employer.

Why These Themes Are Problematic — Through the DR Lens

Each theme above is problematic in roughly the same way. Each describes a tool, or a pattern of tool deployment, with a Diagnostic Ratio at or near 1.0 used as a consequential gate without underlying performance validation. The DR framework introduced earlier explains why all seven themes converge on a single failure mode and why the science and the law penalize that failure mode through the same logic.

A tool with a DR at or below 1.0 cannot reliably distinguish a top performer from an average one. It generates Type II errors — top performers filtered out — at scale, while producing the visible appearance of selection. From a science perspective, the tool is not measuring what its users believe it is measuring. From a legal perspective, the tool cannot satisfy job-relatedness because there is no documented relation to the role's performance to begin with. Both failures are the same failure: the tool was deployed as if it had signal value it does not have.

How the Themes Map to the DR Framework

Each of the seven themes corresponds to a specific failure mode the DR framework predicts:

The Two Accountability Gaps

Employers face simultaneous legal exposure on two independent tracks. Winning on one provides no protection on the other.

Gap 1: Unfair Outcomes

The framework most employers are aware of: civil rights and anti-discrimination law applied to algorithmic screening. The core legal theory is disparate impact — a facially neutral practice that produces statistically unequal results for protected groups violates federal law regardless of discriminatory intent.

The Workday and iTutorGroup cases anchor this track. The legal standard is the same one governing traditional hiring under the Uniform Guidelines on Employee Selection Procedures: if a selection method produces adverse impact, the employer bears the burden of demonstrating the method is job-validated.

Most AI hiring tools cannot satisfy this burden because they were not built against job-specific performance criteria. They were built against historical hiring patterns — which are themselves the record of prior discriminatory decisions.

Gap 2: Invisible Processes

The emerging framework introduced by Kistler v. Eightfold. The Fair Credit Reporting Act requires any entity compiling third-party information about consumers for consequential decisions — including employment — to: notify the subject; obtain employer certification; allow access; and provide dispute rights.

The Eightfold case argues that AI platforms aggregating third-party data to generate candidate scores function as consumer reporting agencies under the FCRA — and have been operating without any required disclosures.

The critical implication: this theory requires no proof of discrimination. A fully fair AI tool can still create FCRA liability if the process was undisclosed. The universe of potentially exposed employers is dramatically larger than under disparate impact theory alone.

The Vendor Liability Trap

What vendor contracts typically provide versus what employers may retain legally — the contractual asymmetries that, where present, can leave HR exposed.

The procedural rulings in Mobley v. Workday suggest that contractual structures alone may not protect employers from liability where a court classifies a vendor's tool as performing an employment-agency function. Where that classification holds, employers may not be able to redirect liability to the vendor on the basis of contractual indemnification language alone.

The practical implication: organizations using automated candidate screening platforms — particularly those with limited audit, validation, or correction rights under their vendor contracts — may carry legal exposure that the contract itself does not transfer to the vendor. Whether a specific platform fits this description depends on the contract, the implementation, and the jurisdiction; this is a question for employment counsel rather than a conclusion this document can draw.

The Regulatory Landscape

Litigation is only one dimension of exposure. A parallel regulatory regime is constructing a multi-state compliance minefield.

Federal Framework

No comprehensive federal AI-in-hiring law exists as of mid-2026. Federal exposure flows from existing statutes applied to new technology:

• Title VII (1964) — race, color, religion, sex, national origin discrimination
• ADEA (1967) — protection for workers and applicants 40+
• ADA (1990) — disability discrimination and accommodation requirements
• FCRA (1970) — consumer report compilation and disclosure requirements
• UGESP (1978) — job validation requirements for all selection tools

State & Local Regulation

A company operating across New York City, California, Colorado, and Illinois must simultaneously satisfy four distinct regulatory regimes — each with different audit, disclosure, and record-keeping requirements — while managing federal exposure under five separate statutes. For employers whose hiring footprint extends to the EU or UK, the picture is more complex still: an entirely separate regulatory architecture — built on different statutory foundations but addressing the same underlying measurement failures — is examined in the next section.

The EU & UK Regulatory Landscape

The United States is testing AI hiring tools case by case, primarily under anti-discrimination statutes written before AI existed. The EU and UK are testing the same tools through purpose-built statutory frameworks. The vocabulary differs; the underlying measurement failure being addressed is identical. An employer whose hiring footprint touches either jurisdiction faces obligations that cannot be satisfied by US-only compliance work — and that bear directly on the same Diagnostic Ratio and Type II error realities documented throughout this report.

Three Independent Legal Layers

Multinational employers operating across the US, the UK, and the EU face three distinct legal assessments that must each be satisfied independently. A tool that satisfies UK automated decision-making rules does not automatically satisfy EU requirements. A tool with EU AI Act conformity does not satisfy UK GDPR. And neither addresses US Title VII or ADEA exposure. The three layers:

• UK GDPR (as amended by the Data (Use and Access) Act 2025) — automated decision-making rules with the ICO as enforcement authority
• EU GDPR Article 22 — the right not to be subject to solely automated decisions producing legal or similarly significant effects, enforced by national data protection authorities
• EU AI Act (Regulation 2024/1689) — purpose-built framework treating hiring AI as a high-risk system requiring conformity assessments, technical documentation, and human oversight

The UK Equality Act 2010 runs alongside all three regimes independently. As one published analysis put it bluntly: in a UK Employment Tribunal, "the algorithm produced the shortlist" is not a defense. Bias testing is the employer's obligation — not a question that can be outsourced to the vendor.

EU AI Act — Hiring Explicitly Classified as High-Risk

The EU AI Act entered into force on August 1, 2024. Annex III of the Act explicitly classifies AI systems used for recruitment, candidate selection, CV screening, ranking, video interview assessment, and employment decision-making as high-risk. The classification triggers obligations on both providers (the vendors building the tool) and deployers (the employers using it). The latter point matters: deployer obligations cannot be contracted away to the vendor.

Deployer obligations under Article 26, as applicable to high-risk hiring AI, include:

• Use the system only within the provider's documented intended purpose — deviation voids the compliance framework
• Implement and maintain meaningful human oversight, with reviewers who have the authority and competence to override the system's output
• Monitor system performance during operation, including logging and bias-related performance tracking
• Maintain documentation of the deployer's own risk assessment and use-case analysis
• Cooperate with national market surveillance authorities and report serious incidents

The August 2026 Deadline — and the May 2026 Delay

The high-risk system obligations were originally scheduled to apply from August 2, 2026. On May 7, 2026, the European Parliament, Council, and Commission reached a provisional political agreement (the "AI Omnibus") to delay the high-risk obligations for employment-related AI systems to December 2, 2027 — a sixteen-month postponement driven primarily by the failure of harmonized technical standards to be finalized on the original timeline.

Two important caveats apply to the delay:

• The delay is not yet law. The provisional agreement must still be formally adopted by the European Parliament and Council and published in the Official Journal before August 2, 2026. If formal adoption fails, the original August 2, 2026 deadline applies as written. Legal analysts have advised organizations to continue preparing against the August 2026 date until the Omnibus is published.
• Other obligations remain on the August 2026 timeline. Transparency obligations under Article 50 (including chatbot disclosure and AI-generated content labeling) come into effect August 2, 2026 regardless. The prohibition on emotion recognition in workplace and educational settings (Article 5) has been enforceable since February 2, 2025. General-purpose AI model obligations have applied since August 2, 2025.

Penalties under the AI Act reach €15 million or 3% of global annual turnover for high-risk system violations, and up to €35 million or 7% of global turnover for prohibited practices — whichever figure is higher in each case.

GDPR Article 22 — Already In Force, Already Being Enforced

Regardless of the AI Act's deferred timeline, GDPR Article 22 has been enforceable across the EU since 2018 and applies to recruitment decisions today. Article 22 gives individuals the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects on them — language that EU data protection authorities have consistently interpreted to cover recruitment screening, ranking, and rejection decisions.

UK Parallel Track — The ICO's "Recruitment Rewired"

The UK is not an EU member state, but its data protection framework (UK GDPR) mirrors the EU GDPR closely, and the Information Commissioner's Office (ICO) has been the most operationally active regulator on AI hiring globally. The UK regulatory architecture has three components employers cannot ignore:

• UK GDPR + Data (Use and Access) Act 2025 (DUAA) — effective February 5, 2026, the DUAA updated UK GDPR's automated decision-making provisions and created a clearer framework for lawful use of automation.
• UK Equality Act 2010 — applies to every set of AI recruitment tools regardless of vendor origin. Employer liability for AI-driven indirect discrimination is absolute. Bias testing is the employer's obligation.
• ICO regulatory guidance and enforcement — operationalized through audits, direct letters to named organizations, and published guidance.

On March 31, 2026, the ICO published Recruitment Rewired: an update on the ICO's work on the fair and responsible use of automation in recruitment. Based on evidence gathered from over 30 UK employers between March 2025 and January 2026, the report's central finding cuts to the heart of the matter:

The ICO sent direct letters to 16 named organizations identified as likely using automated decision-making in recruitment. All 16 committed to acting on the ICO's recommendations. The 2024 audit of AI recruitment tool providers produced almost 300 recommendations for improving compliance, with the regulator finding that some tools allowed recruiters to filter out candidates with protected characteristics, while others inferred gender and ethnicity from candidate names rather than asking directly.

Why This Section Belongs in This Report

The thesis of this whitepaper — that the same upstream measurement failure (low-DR tools deployed as consequential gates without job-specific validation) produces both legal exposure and Type II errors — does not stop at the US border. The EU and UK regulatory architectures are addressing the same problem through different statutory vocabulary. The EU AI Act's bias testing and conformity assessment requirements, the GDPR's meaningful-human-involvement standard, the UK ICO's rubber-stamping critique, and the UK Equality Act's indirect-discrimination test all converge on the same operational question: does the tool actually separate strong candidates from weak ones, or does it merely automate an unvalidated screening process at greater speed?

The convergence matters because it changes the strategic calculation. A US employer can plausibly tell itself that the litigation risk documented in Section 2 is statistically small — one lawsuit per X thousand hires. The same employer hiring across the EU and UK faces a regulatory architecture where the question is not whether the tools will be assessed but against what standard they will be assessed. The standards being constructed (the AI Act's conformity assessment, the ICO's meaningful-involvement test, the AP's right-to-explanation guidance) are all asking the same underlying question that the Diagnostic Ratio framework answers directly. An employer whose tools sit at DR ≤ 1.5 will fail every one of these tests — not because the tools are exotic edge cases, but because they cannot demonstrate the separation between strong and weak candidates that every framework, regardless of jurisdiction, ultimately requires.

The Type II error point applies identically across both legal systems. A tool that filters out half the strongest candidates before any human review is not just a Quality-of-Hire problem in San Francisco — it is also the operational reality that triggers every EU and UK obligation discussed above. One intervention addresses both. The next section examines, tool by tool, what that intervention requires.

The DR Profile of Common Hiring Tools

The themes documented in Section 3 implicate specific categories of tools. The table below summarizes the Diagnostic Ratio profile of those categories, drawn from the meta-analytic validity literature, and the corresponding legal posture each category occupies in the current landscape. The DR is the bridge — it is the property that makes both the operational failure and the legal indefensibility true at the same time.

Each row pairs a tool category's DR with the legal posture associated with it. The DR values are best-estimate midpoints from Sackett, Zhang, Berry & Lievens (2022) and related meta-analyses, translated through Taylor-Russell-style assumptions about top-quartile thresholds (see disclaimer for full caveats).

Implications for HR Leaders

Immediate Risk Assessment

Any organization using the following tools or practices should conduct an immediate legal review:

• Automated resume screening that ranks or filters candidates before human review
• AI video interview platforms that score facial expressions, speech patterns, or behavioral signals
• Personality or cognitive assessments used as screening gates without documented job validation
• Keyword-matching ATS systems that automatically disqualify on credential requirements
• Third-party platforms aggregating candidate data from sources beyond the job application

The Two Questions Every Audit Must Answer

The legal standard established by the UGESP and now enforced through AI litigation reduces to two questions:

• Question 1: Does the selection method produce adverse impact on any protected group?
• Question 2: If so, is the method validated against job-specific performance criteria?

Most AI hiring tools fail on Question 1. Almost none can satisfy Question 2 — because they were not built against defined performance criteria.

A Note on Enforcement Direction

The Trump Administration's 2025 executive order directing agencies to deprioritize disparate-impact enforcement has been read by some as reducing AI hiring risk. This reading is incorrect for three reasons:

• Private litigation — the primary driver of every case in this whitepaper — is unaffected by executive enforcement priorities
• State enforcement (NYC, California, Colorado, Illinois, and 40+ pending states) is explicitly independent of federal priorities
• The FCRA theory in Kistler v. Eightfold is a consumer protection claim — unaffected by disparate-impact deprioritization

Organizations that interpret reduced federal enforcement as reduced legal risk are misreading the direction of travel.

How the 2021 Goldstein Principles Map to the Themes

Goldstein's 2021 legal analysis identified the principles by which a hiring methodology can be defended under existing employment law. The table below — prepared by The Adler Group, not by Goldstein or Littler Mendelson — pairs each of those principles with the corresponding theme observed across the sample in Section 3. The mapping describes a structural correspondence between the 2021 analysis and the patterns the post-2023 actions are testing, not a prediction of outcomes in any specific matter.

The legal principles for mounting a job-relatedness defense were articulated in published form before the post-2023 actions were filed. If those actions continue to turn on whether the underlying methodology can be defended on validity grounds, then a methodology built around the 2021 principles would, on its face, be more responsive to the themes the actions are testing than one that was not. Whether that face-value correspondence holds in any specific matter depends on the implementation, the jurisdiction, and the validation work the employer has independently performed.

Conclusion: The Case Has Already Been Made

The argument in this document is one most readers were already partway to before they opened it. Hiring is a decision made under uncertainty. Tools used to reduce that uncertainty either give you reason to trust their output or they don't. The DR is the metric that says which. Most of the tools at the center of corporate hiring today have DRs at or below 1.0. They look like rigor. They are not.

The seven themes documented in Section 3 describe the structural patterns that recur across the ten public actions examined. The DR analysis in Sections 4 and 8 explains why each theme is problematic in roughly the same way: a tool with low or unmeasured DR used as a consequential gate without an underlying job-validated performance specification. The scientific failure and the legal failure are not two problems. They are the same problem expressed in two vocabularies.

The damage from low-DR tools is far larger than the litigation reveals. Every applicant cycle, those tools generate Type II errors at scale — top performers filtered out before any human evaluation begins. The Harvard Business School and Accenture "Hidden Workers" research (2021) documented more than 27 million such filtered candidates in U.S. labor markets. None of them sue. None of them appear in employer metrics. The cost is fully invisible until the empty seat the rejected candidate would have filled becomes a missed quarter, a delayed product launch, or a vacancy held open while the recruiting funnel re-runs the same filters that rejected them the first time. The legal exposure is the smaller and louder half of the problem. The Type II errors are the larger and silent half.

The diagnostic principles described in this document are drawn from decades of meta-analytic research culminating in Sackett et al. (2022). The legal principles supporting a defensible alternative are set out in the 2021 analysis by David J. Goldstein of Littler Mendelson. The DR framework that unifies them is developed in The Diagnostic Ratio of Common Hiring Tools. If the validity research is correct and if the cases since 2023 turn on whether the underlying methodology can be defended on validity grounds, then organizations relying on low-DR screening tools without supporting validation work would face exposure on both the operational and legal sides simultaneously. Whether that exposure has crystallized for any specific organization is a question for that organization's counsel and I/O psychology advisors, not for this document.

Sources

• Mobley v. Workday, Inc., No. 3:23-cv-00770 (N.D. Cal., filed Feb. 2023; collective action certified May 2025)
• Mobley v. Workday, Inc., No. 3:23-cv-00770 (N.D. Cal., order denying motion to dismiss in part, June 22, 2026).
• EEOC v. iTutorGroup, Inc., No. 1:22-cv-02565-PKC-PK (E.D.N.Y., settled Aug. 2023)
• Kistler et al. v. Eightfold AI Inc., filed Jan. 20, 2026, Contra Costa County Superior Court, California
• Harper v. Sirius XM Radio, LLC, No. 2:25-cv-12403 (E.D. Mich., filed Aug. 4, 2025)
• ACLU of Colorado, Complaint re: Intuit, Inc. and HireVue, Inc., filed March 2025 (EEOC and Colorado Civil Rights Division)
• FTC Complaints re: Aon ADEPT-15, vidAssess-AI, gridChallenge, 2025 (investigation pending)
• SafeRent Solutions class action, settled $2M+, 2024 (Fair Housing Act)
• Sackett, P.R., Zhang, C., Berry, C.M., & Lievens, F. (2022). Journal of Applied Psychology, 107(11), 2040–2068.
• Goldstein, D. J. (2021). Performance-based Hiring and Legal Compliance. Performance-based Hiring Learning Systems White Paper. Littler Mendelson, Minneapolis. (Updated for the 4th edition of Hire With Your Head; original 2013.)
• Gallup. (2024). State of the Global Workplace Report. Q12 research base: 3.3M workers, 100,000+ teams.
• Schmidt, F.L., & Hunter, J.E. (1998). Psychological Bulletin, 124(2), 262–274.
• The Adler Group. (2026). The Evidence Base for Performance-based Hiring: A Systematic Review. Working Paper.
• The Adler Group. (2026). The Diagnostic Ratio of Common Hiring Tools. Performance-based Hiring Bayesian Talent Tools. Available at kanbanforhiring.com/bayes/dr-ratio-hiring-report.
• Castilla, E. J. (2005). Social networks and employee performance in a call center. American Journal of Sociology, 110(5), 1243–1283.
• Burks, S. V., Cowgill, B., Hoffman, M., & Housman, M. (2015). The value of hiring through employee referrals. Quarterly Journal of Economics, 130(2), 805–839.
• Burning Glass Institute & Harvard Business School. (2024). Skills-Based Hiring: The Long Road from Pronouncements to Practice.
• Fuller, J., Raman, M., et al. (2021). Hidden Workers: Untapped Talent. Harvard Business School Project on Managing the Future of Work, in collaboration with Accenture.
• Deloitte. (2025). Global Human Capital Trends Report. Survey of ~10,000 leaders, 93 countries.
• Uniform Guidelines on Employee Selection Procedures, 29 C.F.R. Part 1607 (1978).
• EEOC Strategic Enforcement Plan 2024–2028 (AI hiring discrimination designated priority area).

Disclaimer