Research: Algorithmic HR — Hiring, Scoring, and Firing by Machine

July 2, 2026

AI-mediated employment decisions where a real person was demonstrably harmed and the harm is documented in a filed legal record, regulator action, primary news investigation, or a named principal’s own statements. Every claim about a real person or organization carries a source. What a vendor claims about its own tool is kept distinct from what the evidence supports; filed complaints are stated as allegations, not findings; settlements and consent decrees are not adjudications of merits; and each named principal’s denial is preserved in their own words.

The load-bearing court fact here is the Mobley v. Workday “agent” holding — the moment U.S. law stopped letting the algorithm be the alibi.

AI Resume Screening at Scale

Mobley v. Workday, Inc. — N.D. Cal. No. 3:23-cv-00770

Filed February 21, AD 2023 by Derek Mobley, a Black man over forty with anxiety and depression. Mobley alleges he applied to more than 100 jobs at companies using Workday’s applicant-screening system since AD 2017 and was rejected from every one (case docket, Civil Rights Litigation Clearinghouse).

The Workday tools at issue are Candidate Skills Match (which extracts skills from a posting and applicant materials and reports the match as strong/good/fair/low/pending/unable-to-score) and the Workday Assessment Connector. Plaintiffs allege the Connector uses machine learning to identify which protected-class members an employer disfavors and to reduce their recommendation rate. Workday denies that it scores protected characteristics (case docket).

On July 12, AD 2024, Judge Rita Lin ruled that Workday could be held directly liable under Title VII, the ADEA, and the ADA as an “agent” of the employers deploying its software — a doctrine-shifting holding that puts vendors, not just deployers, on the hook (court order, GovInfo; EEOC amicus brief, April 2024).

On May 16, AD 2025, Judge Lin granted preliminary collective certification under the ADEA. In its own representation, Workday’s software rejected 1.1 billion applications in the relevant window; the collective is potentially “hundreds of millions” of members, characterized in legal-press coverage as one of the largest ever certified in U.S. employment law (Proskauer, Law and the Workplace, June 2025).

On March 6, AD 2026, Judge Lin dismissed one ADA claim from a cancer-survivor/asthma plaintiff with leave to amend; plaintiffs filed an amended complaint March 30, AD 2026 re-upping the disability and California state-law claims (case docket).

Every allegation of discriminatory scoring is the plaintiffs’ filed claim, unadjudicated on the merits as of this writing; the “agent” liability holding and the collective certification are Judge Lin’s rulings; and the 1.1-billion figure is Workday’s own representation, not an accusation. Workday denies scoring protected characteristics.

Amazon’s scrapped resume screener (AD 2014–2018)

Per Reuters reporting, Amazon built an internal experimental screening tool trained on ten years of resumes. Because the training set was male-dominated, the model penalized resumes containing the word “women’s” (as in “women’s chess club captain”) and downranked graduates of two unnamed all-women’s colleges, while rewarding male-coded verbs such as “executed” and “captured.” Amazon disbanded the team in early AD 2018 after losing confidence the system could be made gender-neutral. Amazon’s position at the time was that the tool was never used to evaluate real candidates in production (Reuters, Jeffrey Dastin, October 10, 2018, via ACLU explainer; MIT Technology Review, October 10, 2018).

This is on-record investigative reporting sourced to internal Amazon accounts — Reuters’ finding, not an established finding of fact — and it concerns an experimental tool distinct from Amazon’s operational HR systems described below.

Eightfold AI

A class action has been reported filed against Eightfold AI for algorithmic screening discrimination, tracked by employer-defense firms as a follow-on to Mobley (National Law Review). Only secondary law-firm commentary is on hand; the primary docketed complaint has not been retrieved, and the discrimination claim is an unadjudicated allegation. [SOURCE NEEDED — primary complaint]

AI Video Interview and Facial Analysis

HireVue — EPIC FTC complaint (AD 2019–2021)

On November 6, AD 2019, the Electronic Privacy Information Center (EPIC) filed an FTC complaint alleging that HireVue’s video-interview AI committed unfair and deceptive practices. EPIC charged that HireVue’s claims to measure “cognitive ability,” “psychological traits,” “emotional intelligence,” and “social aptitudes” were unproven and discriminatory; that eye-tracking would disparately impact visually impaired applicants; and that HireVue’s denial of using “facial recognition” was misleading, since the company used “facial analysis” (EPIC docket; Washington Post, November 6, 2019).

HireVue’s assertion that its system validly measures those psychological constructs is the company’s marketing claim, not established science; EPIC’s complaint is precisely that the construct-validity claim is unproven.

In January AD 2021, HireVue announced it would stop relying on facial analysis to assess candidates, while continuing to analyze speech, intonation, and behavior — the same biometric pipeline with the visible face-scoring layer pulled. Per HireVue at the time, 10–30% of a candidate’s score had come from facial expressions, the remainder from language (EPIC announcement).

The FTC never publicly disposed of the complaint. HireVue was accused of deception and discrimination; it was not found to have committed either.

ACLU / Public Justice v. Intuit and HireVue (Colorado CRD / EEOC, AD 2025)

On March 19, AD 2025, the ACLU, the ACLU of Colorado, Public Justice, and Eisenberg & Baum filed charges with the Colorado Civil Rights Division and the EEOC on behalf of “D.K.,” an Indigenous and Deaf woman denied a promotion at Intuit after a HireVue automated video interview (filed complaint, ACLU; Public Justice).

Per the filed complaint, D.K. requested human-generated captioning; Intuit told her HireVue’s software included subtitling; when she began the interview no subtitling was available; she fell back on Chrome auto-captions, which were unreliable; she was rejected; and the system gave her auto-generated feedback telling her to “practice active listening” and provide “more concise and direct answers.” The charge invokes the Colorado Anti-Discrimination Act, the ADA, and Title VII.

HireVue CEO Jeremy Friedman publicly responded that the complaint is “entirely without merit” and that “Intuit did not use a HireVue AI-based assessment.” The account of D.K.’s experience is the filed administrative charge, unadjudicated; which of the two companies bears liability is disputed; and Friedman’s denial stands in his own words.

MIT Technology Review — MyInterview / Curious Thing test (AD 2021)

MIT Technology Review investigators submitted a candidate reading a Wikipedia entry aloud in German to MyInterview’s video-interview engine, which is built on Big Five personality scoring. The system returned a 73% job-match score for a fake English-language job and rated her highly on English language skills — despite a transcript that was nonsense, German-to-English phoneme misrecognition treated as content. The same outlet replicated the result on Curious Thing (MIT Technology Review, July 7, 2021).

This is a documented journalistic test; the vendors were given the standard right of reply in the piece. It is the clean demonstration that the construct-validity marketing is not supported by the tools’ actual behavior.

Aon Consulting — ACLU FTC and EEOC complaints (AD 2024)

On May 29, AD 2024, the ACLU filed an FTC complaint against Aon Consulting. The complaint alleges that Aon deceptively markets three hiring tools — ADEPT-15 (personality), gridChallenge (cognitive), and vidAssess (video interview) — as “bias-free” when they likely discriminate by race and disability. The ACLU’s specific claim is that ADEPT-15 asks questions that track closely with clinical assessments for autism and other mental-health disabilities (for example, “I have difficulty determining how someone feels by looking at their face”), screening out autistic candidates (filed complaint, ACLU; ACLU model-card analysis; ACLU press release).

In late AD 2023 the ACLU filed parallel EEOC class charges against Aon and an employer using its assessments, on behalf of a biracial autistic applicant. These are unadjudicated advocacy-filed complaints: “bias-free” is Aon’s own marketing per the complaint, “likely discriminate” is the ACLU’s expert reading, and no FTC or EEOC disposition is on hand.

Algorithmic Firing Patterns

Amazon Flex — Bloomberg Businessweek investigation (June 28, AD 2021)

Bloomberg’s “Fired by Bot” investigation profiled Stephen Normandin, a 63-year-old Army veteran who delivered for Amazon Flex in Phoenix for almost four years. The algorithm deactivated him via automated email after marking him down for failed deliveries — including predawn drops at apartment complexes whose gates and offices were locked, where the app itself instructed drivers to leave packages at offices that were not yet open. The Flex algorithm rates drivers in four buckets: Fantastic, Great, Fair, At Risk. Bloomberg’s central finding was that Amazon “knew delegating work to machines would lead to mistakes and damaging headlines, but decided it was cheaper to trust the algorithms than pay people to investigate mistaken firings so long as the drivers could be replaced easily” (Bloomberg Businessweek, June 28, 2021). Amazon disputed the framing to Bloomberg.

Amazon warehouse productivity terminations

Per a signed letter from Amazon’s outside counsel obtained by The Verge (April AD 2019), Amazon fired “hundreds” of associates at a single Baltimore fulfillment center between August AD 2017 and September AD 2018 — roughly 300 full-time associates — for failing productivity quotas. The system tracks “time off task” (TOT) and auto-generates warnings and terminations without supervisor input, per the counsel’s own description (The Verge, April 25, 2019; MIT Technology Review, April 26, 2019). The Baltimore figure is a self-disclosure — Amazon’s own counsel’s account.

In AD 2025 the EEOC issued a determination that Amazon engaged in systemic, nationwide violations of the ADA over more than six years, forcing disabled workers onto unpaid leave and then penalizing or firing them for exhausting it (A Better Balance summary). An EEOC determination is an agency finding that an employer can contest before it becomes a court judgment.

Amazon delivery vans — Netradyne / Driveri / Mentor (AD 2021–)

Netradyne’s Driveri four-lens camera array surveils Amazon delivery drivers continuously, paired with the Mentor scoring app; each driver gets a weekly report card ranging Fantastic to Poor based on infractions per 100 trips. A privacy researcher quoted in primary coverage called it “the largest expansion of corporate surveillance in human history.” Senators Markey, Booker, and others questioned Amazon about the system in March AD 2021 (Vice / Motherboard; CNBC, February 3, 2021). The “largest expansion” line is the researcher’s characterization; Netradyne is the vendor, Amazon the deployer.

Xsolla — algorithmic mass firing (August 3, AD 2021)

CEO Aleksandr Agapitov fired roughly 150 employees at Xsolla’s Perm, Russia office in a single email, citing a big-data team’s analysis of activity in Jira, Confluence, Gmail, chats, documents, and dashboards. Workers were “tagged as unengaged and unproductive” by a system measuring 30 characteristics on a 100-point scale, including reading and writing wiki articles and attending internal meetings. Agapitov’s stated goal was to reduce the salary budget by 10% after growth slipped from a 40% pace. The leaked dismissal email opened, “Many of you might be shocked, but I truly believe that Xsolla is not for you” (Game Developer, August 2021; AIAAIC incident repository). Agapitov did not dispute authorship of the email and commented publicly on the episode; “unengaged” is the company’s tag, not a judgment of the workers.

Uber and Lyft driver deactivations (California, AD 2017–2026)

The AD 2017 Lyft settlement ($27M, federal court) prohibited Lyft from deactivating California drivers without providing a reason and limited reasons to a pre-defined list — the first written constraint on rideshare algorithmic dismissal. A separate AD 2022 Uber California class settlement was $8.4M.

In April AD 2026, drivers sued Uber in San Francisco Superior Court alleging that Uber deactivates drivers on grounds not specified in its Platform Access Agreement and provides no real appeal mechanism; the suit is structured as a constitutional challenge to gaps left by Proposition 22 (CalMatters, April 2026). The settlements are settled court records and are not admissions of liability; the April AD 2026 allegations are the plaintiff drivers’ unadjudicated claims.

DoorDash — NY Attorney General settlement (February AD 2025)

Attorney General Letitia James secured a $16.75M settlement with DoorDash after the office found the company used customer tips between May AD 2017 and September AD 2019 to subsidize its “guaranteed pay” floor rather than passing tips through. The settlement requires DoorDash to give dashers — including deactivated ones — access to four years of delivery history (NY Attorney General press release, 2025). This is a wage/tip case rather than a pure deactivation case, but it is the largest gig-economy enforcement of the period, and the four-year-history clause directly affects deactivation contestability. A settlement resolves without a court adjudication of liability.

Workplace Surveillance AI

Microsoft Productivity Score (November AD 2020)

Microsoft 365 shipped a per-user “Productivity Score” dashboard tracking 73 metrics — emails sent, Teams meetings attended, files read or edited, camera-on time, @mentions, recipients per email. Wolfie Christl (Cracked Labs, Vienna) flagged it on November 24, AD 2020, calling it a “full-fledged workplace surveillance tool” enabled by default. The named-user feature was removed on December 1, AD 2020 after the backlash (The Register, December 1, 2020). The 73-metric dashboard and its de-naming are Microsoft’s own product behavior; “surveillance tool” is Christl’s characterization, against Microsoft’s own “productivity insights” framing.

Surveillance-tool landscape — Time Doctor, Hubstaff, ActivTrak, Teramind

All four are commercial-grade. Per the vendors’ own materials and press coverage, Teramind is the most invasive (live screen recording, full keystroke logging, email-content monitoring, behavioral analytics); ActivTrak markets itself as “privacy-first” with focus-time analytics; Time Doctor and Hubstaff sit between, with optional screenshots and activity tracking (Brookings; Washington Post, September 24, 2021). These capability descriptions come from the products’ own marketing; a capability is not evidence that any named employer misused a tool. “Most invasive” is comparative; “privacy-first” is ActivTrak’s own positioning. Specific employer adoption is difficult to source, since vendors do not publish customer lists. [SOURCE NEEDED — named-employer deployment]

Performance-Evaluation Algorithms

Bridgewater Associates — Dot Collector

Every Bridgewater employee carries a proprietary iPad running Dot Collector, the system Ray Dalio designed for “radical transparency.” Employees rate each other on roughly 75 attributes during meetings on a 1–10 scale, and every dot given is public. Behind the app runs a proprietary algorithm that constructs “believability-weighted averages” — the opinions of employees with stronger track records count more — with employee-to-role matching driven by the aggregated “pointillist painting” of each person (Quartz, on Dalio’s TED Talk). In AD 2022, Coinbase reportedly piloted a Dot Collector-style real-time-feedback app (TechCrunch, May 28, 2022). This cluster is descriptive and self-disclosed; no firing or harm is documented here.

IBM AskHR / Watson Career Coach

Per CEO Arvind Krishna to the Wall Street Journal (AD 2023), IBM’s AskHR AI agent automated 94% of routine HR tasks — vacation requests, pay statements — and Krishna said IBM used the freed budget to hire more programmers and salespeople, with total headcount going up. IBM also separately markets Watson Career Coach for skills-gap analysis and career-path mapping inside large organizations (Entrepreneur, citing the WSJ). The 94%-automation and net-headcount-up figures are Krishna’s own claims to the WSJ, not independently verified; the original headline glosses this as “replaced hundreds of HR staff,” whereas Krishna’s framing was reallocation with net headcount up.

Google GRAD

Google’s performance system, introduced in May AD 2022, is Googler Reviews and Development (GRAD) — a 5-point scale, 360-degree feedback, with separated pay and review conversations (Google Careers). This is descriptive, from Google’s own documentation; no wrongdoing is alleged.

Remote and Hybrid Worker Monitoring

The vendor landscape above overlaps this section; the distinguishing factor is that during and after the COVID-19 work-from-home wave, employer adoption jumped sharply. Per Fortune (September AD 2021), only 10% of remote companies monitored employees before the pandemic, and deployment “jumped over 50%” in the first 18 months (Fortune, September 1, 2021). The Washington Post’s September AD 2021 deep-dive documenting keystroke counts, screenshots, and webcam-on requirements for in-home work is the strongest primary news source for the home-monitoring lifestyle (Washington Post, September 24, 2021).

No major U.S. class action has resolved purely on remote-monitoring grounds. The closest analogue is the body of BIPA cases in Illinois (biometric keystroke and facial scanning) — see the case law below and Deyerler v. HireVue (February AD 2024), which held that the Illinois Biometric Information Privacy Act and the Illinois AI Video Interview Act impose concurrent obligations.

Biometric Employment Verification

E-Verify, I-9, and the AI background-check stack

E-Verify cross-checks Form I-9 against DHS and SSA records; it can be run only after hire and only on completed I-9s, and pre-hire use violates program terms and may itself be discriminatory per program terms and EEOC guidance. In AD 2024–2026, HR-tech vendors have layered automated background screening on top of I-9/E-Verify — resume parsing, social-media scraping, criminal-record matching — with DISA, Sterling, and Checkr among the vendors in this market. [SOURCE NEEDED — market-share claim]

The ban-the-box versus algorithm tension is that ban-the-box laws delay disclosure of criminal history to give a human resume reviewer an unbiased first read, while AI screeners often surface the same information from public records the moment they parse a name — defeating the purpose of the law before a human sees the file. AD 2024 California Civil Rights Council rules provide that AI tools producing disparate impact on protected groups trigger Title VII / FEHA liability for the deploying employer, and that employers must retain automated-decision data for four years.

DISA, Sterling, and Checkr are named here only as market participants providing automated screening; no wrongdoing is alleged against any of them.

REAL ID and employment

REAL ID enforcement at airports began May 7, AD 2025. REAL ID does not yet replace the I-9, but it creates a federally normed baseline biometric and biographic credential that downstream identity-verification vendors are wiring into onboarding workflows. TSA’s ConfirmID alternate-verification system was announced December 1, AD 2025 and set to launch February 1, AD 2026, normalizing paid biometric verification by federal contractors — a precedent the private employment-verification market is following (Biometric Update, May 2024). The enforcement dates and program existence are government-program facts; the “wiring into onboarding” trajectory is analysis, not a claim that any named vendor is doing anything improper.

The Bias Case Law

Federal — agency action

EEOC v. iTutorGroup, Inc. (E.D.N.Y. No. 1:22-cv-02565) settled August 9 with a consent decree entered September 8, AD 2023: $365,000 to a group of roughly 200 applicants whose applications were auto-rejected because they were female and 55+ or male and 60+. Per the EEOC’s account, the discovery happened only because one applicant submitted two identical applications with different birth dates and got an interview only on the younger one; the EEOC monitors compliance for five years (EEOC press release; case docket, Civil Rights Litigation Clearinghouse). A consent decree resolves the suit without an admission of liability on the merits.

EEOC Technical Assistance (May 18, AD 2023), “Select Issues: Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII of the Civil Rights Act of 1964,” holds employers liable for disparate impact even when the tool is built by a vendor and even when the vendor certifies it is bias-free. [SOURCE NEEDED — EEOC guidance permalink]

Federal — landmark private litigation

Mobley v. Workday (above) is the structural landmark. Judge Lin’s July AD 2024 ruling that a software vendor is an “agent” of the employer under Title VII, the ADEA, and the ADA — combined with the May AD 2025 collective certification and Workday’s own “1.1 billion applications rejected” representation — puts vendor liability on the same footing as employer liability.

State and municipal

NYC Local Law 144 (effective July AD 2023) requires annual independent bias audits of any Automated Employment Decision Tool used on NYC-based hires, with results published publicly; employers must notify candidates at least 10 business days before an AEDT is used; penalties run $500–$1,500 per day, enforced by the NYC Department of Consumer and Worker Protection (NYC DCWP).

Illinois HB 3773 (signed August 9, AD 2024; effective January 1, AD 2026) amends the Illinois Human Rights Act to prohibit AI use in recruitment, hiring, promotion, discipline, or discharge where the AI causes discrimination on a protected-class basis, and bans ZIP code as a proxy for protected class. It layers on top of the Illinois AI Video Interview Act (2020) and BIPA; Deyerler v. HireVue (February AD 2024) held that compliance with one does not excuse another (Illinois Public Act 103-0804 (HB 3773)).

Colorado SB 24-205 (signed May 17, AD 2024) establishes a “consequential decisions” framework covering employment, housing, credit, education, and healthcare. Originally effective February 1, AD 2026, it was delayed to June 30, AD 2026 by a special session. It requires impact assessments, consumer disclosures, and opt-out notice, and defines “algorithmic discrimination” as any AI-driven differential treatment based on protected characteristics (Colorado General Assembly, SB 24-205).

Gig-Economy Classification

California AB5 (AD 2019) codified the ABC test for independent-contractor classification: a worker is presumed an employee unless (A) free from control, (B) outside the usual course of the hirer’s business, and (C) independently established in the trade. It was aimed at Uber, Lyft, DoorDash, and Postmates.

In Olson v. State of California (9th Cir.), Lydia Olson and Miguel Perez, plus Uber and Postmates, challenged AB5 in the Central District of California and then the Ninth Circuit; after Prop 22 passed, the Ninth Circuit en banc panel (June AD 2024) sided with the state and rejected the appeal, and the constitutional carve-out for app-based drivers remained in Prop 22 (Ninth Circuit opinion, March 17, 2023).

Proposition 22 (November AD 2020), the $200M+ Uber/Lyft/DoorDash ballot measure, exempted app-based drivers from AB5, keeping them independent contractors with limited benefits. The California Supreme Court unanimously upheld Prop 22 in Castellanos v. State of California on July 25, AD 2024 (CalMatters, July 2024). Prop 22 effectively constitutionalizes the position that the algorithm — not a manager — can manage the worker without crossing into “control” sufficient to trigger employee status; the April AD 2026 San Francisco Superior Court deactivation suit is the next round.

Whistleblowers and Insider Voices

No on-record HireVue, Workday, or major applicant-tracking-system vendor whistleblower surfaces in publicly indexed coverage as of AD 2026. The critics with credibility have been external: EPIC, the ACLU, Wolfie Christl of Cracked Labs, and Meredith Whittaker of the AI Now Institute, who called HireVue “pseudoscience” and “a license to discriminate.”

HireVue principals have gone on record defending the technology: former CTO Loren Larsen characterized critics as having “limited understanding” of worker psychology, and current CEO Jeremy Friedman called the ACLU / Public Justice Intuit complaint “entirely without merit.” Whittaker’s “pseudoscience” and “license to discriminate” are her attributed words; Larsen’s and Friedman’s lines are the principals’ own on-record defenses, quoted so both sides speak in their own voices.

The open question is whether the Mobley collective notice — going to “hundreds of millions” of rejected applicants under Workday’s own representation — produces named affiants from inside Workday’s development team. As of AD 2026 no such affidavit appears in the public docket. The absence of a vendor-side whistleblower is itself a data point: the gag instruments that silence workers are why the docket is often the only witness.

Chapter-Anchor Incidents

Durable, named, sourced incidents:

  1. Stephen Normandin, Phoenix, AD 2017–2021. A 63-year-old Army veteran, four years of Amazon Flex deliveries, deactivated via automated email after the algorithm marked him down for predawn drops at locked apartment complexes where Amazon’s own app told him to leave packages at offices that were not open (Bloomberg “Fired by Bot,” June 28, AD 2021).

  2. Derek Mobley, AD 2017–present. A Black man over 40 with anxiety and depression; applied to 100+ jobs at Workday-using companies since AD 2017, rejected every time; plaintiff in the case that established vendor liability and produced the largest collective certification in U.S. employment-AI history.

  3. D.K., an Indigenous, Deaf woman at Intuit, AD 2024–2025. Denied accommodation for human captioning in a HireVue video interview; the system told her to “practice active listening” (ACLU / Public Justice / EEOC complaint, March 19, AD 2025).

  4. The two applications, iTutorGroup, AD 2020s. The same person, two applications, different birth dates; the younger one got an interview. Discovering the discrimination required someone to test the system on themselves (EEOC consent decree, AD 2023).

  5. The German Wikipedia entry, MIT Technology Review test, AD 2021. A candidate reads aloud in German; MyInterview returns a 73% match for an English-language job and rates her highly on English language skills. The transcript is gibberish; the system grades her anyway.

  6. Aleksandr Agapitov and 150 Xsolla workers, August 3, AD 2021. A mass dismissal email — “Many of you might be shocked, but I truly believe that Xsolla is not for you” — after an algorithm read Jira tickets and wiki edits and assigned engagement scores. The goal was a 10% budget cut.

  7. Wolfie Christl versus Microsoft Productivity Score, November 24, AD 2020. One Vienna researcher reads the Microsoft 365 documentation, names the surveillance feature publicly, and within a week Microsoft has stripped names from the dashboard.

  8. The Mobley July AD 2024 “agent” ruling. Judge Rita Lin holds that a software vendor can be directly liable under Title VII as the employer’s agent. Before this ruling, vendors hid behind “we just sell the tool.” After it, they cannot.

  9. The Baltimore fulfillment center, AD 2017–2018. Amazon’s own counsel admitted to The Verge that “hundreds” of associates — about 300 — were fired in 13 months at a single facility for productivity quotas, with no supervisor input on the terminations.

  10. The four-bucket report card. Amazon Flex driver ratings — Fantastic, Great, Fair, At Risk — and the same four-tier weekly card pushed by Netradyne’s Mentor app to delivery-van drivers. The vocabulary of the day-care preschool, applied to adults whose livelihoods turn on the bucket assignment.

The Load-Bearing Finding

Judge Rita Lin’s July 12, AD 2024 ruling in Mobley v. Workday that AI vendors are subject to Title VII, ADEA, and ADA liability as “agents” of the deploying employer. Combined with the May 16, AD 2025 ADEA collective certification — across a class Workday itself sized at 1.1 billion rejected applications — this is the moment American law stopped letting the algorithm be the alibi. The vendor is not an oracle the employer consults; the vendor is the employer’s hand. If the hand discriminates, the employer is liable and so is the hand’s maker.

How to Read This Record

The load-bearing spine of the hiring-and-firing story is court-record and regulator-action fact, not accusation. Three tiers of certainty govern how each cluster reads:

The unadjudicated complaints (Workday scoring, Aon, Eightfold, Intuit, the AD 2026 Uber suit) are the highest-exposure items; the Eightfold cluster is not yet sourced to a docketed complaint and names no parties beyond the vendor.

Get updates on the Evil Robots series

Newsletter essays on AI escape, deception, and the humans who built them.