← The Atlas Blog

Workday Hiring AI

Filed February 2023 · EEOC amicus brief filed 2024

Class action alleges AI screening tool discriminated by race, age, and disability. EEOC filed amicus brief.

Human Agency Data Governance

What Happened

Derek Mobley applied for more than 80 jobs through employers using Workday's AI-powered hiring platform.

He was rejected from all of them — without a single human ever reviewing his application.

Mobley, who is Black and over 40, filed a federal class action lawsuit in February 2023 alleging that Workday's AI hiring tools discriminated against him and similarly situated applicants based on race, age, and disability status. He alleged that the AI screening tools systematically filtered out protected-class candidates before any human recruiter saw their applications.

The case gained significant weight in 2024 when the U.S. Equal Employment Opportunity Commission filed an amicus brief in support of the plaintiff — one of the first times the EEOC formally weighed in on AI-driven hiring discrimination in a private lawsuit.

District Judge Rita F. Lin allowed the case to proceed, ruling that Workday could be held liable as an "agent" of the employers using its platform — not merely a neutral technology vendor.

The Atlas Analysis

A textbook illustration of what happens when Human Agency is eliminated from a high-stakes decision process affecting millions of people.

Human Agency ≈ 15/100 — Level 1

The Human Agency pillar asks: can humans understand, override, and be protected from AI decisions? In the Workday case, the answer to all three questions was effectively no for affected applicants.

Applicants could not understand the AI's criteria — screening methodologies are not disclosed. Applicants could not override the determination — rejection before a human saw the application meant there was no human to appeal to, because no human had been involved. Applicants could not verify what had happened — a rejection without review produces no feedback, no basis for appeal, and no mechanism for the applicant to understand why they were filtered.

Signal #118 — "Can a user appeal a decision the AI made about them?" Score: 0 — there was no appeal mechanism because the applicant never knew a human had not reviewed their application.
Signal #116 — "Does this AI make decisions that affect people's lives?" Score: 3 — hiring decisions affect employment, income, career trajectory, and family stability. This is among the most consequential category of decision an AI system can make.

The combination of these two signals at maximum consequence and zero recourse is a Level 1 Human Agency finding regardless of any other dimension.

Data ≈ 20/100 — Level 1

AI hiring systems trained on historical hiring data reproduce historical hiring patterns. Historical hiring patterns in most industries reflect decades of documented discrimination by race, age, gender, and disability status. A system trained to identify "successful" candidates using historical data will systematically underrepresent candidates who were historically excluded from "success" — not because of capability, but because of the discriminatory patterns baked into the training signal.

Signal #122 — "Is the output fair across gender, race, language, geography?" Score: 0 — no public bias testing results disclosed.
Signal #85 — "Was the training data independently audited?" Score: 0 — training data composition and bias evaluation not publicly disclosed.
Governance ≈ 25/100 — Level 1

Judge Lin's ruling that Workday could be held liable as an "agent" of its employer-clients — not a neutral tool — is a Governance finding of significant consequence. The ruling means that AI vendors who build systems that make consequential decisions cannot disclaim accountability by characterizing themselves as technology providers rather than decision-makers.

This is the same principle that the Air Canada tribunal established for chatbot outputs: you own what your AI system does. Workday, as the builder of a system that filtered applicants before human review, could not separate itself from the consequences of those filters.

Signal #113 — "Has the AI system been through a third-party ethics audit?" Score: 0 — no public audit results disclosed.

Signals That Would Have Caught It

01

No demographic parity validation before deployment. A pre-deployment analysis of screening outcomes by race, age, gender, and disability status would have identified disparate impact before a single real applicant was filtered. This is a standard requirement in employment law for any selection procedure — AI or otherwise.

02

No applicant-facing transparency about AI involvement. Applicants using employer hiring portals built on Workday had no way to know that their applications were being screened by an AI before any human reviewed them. Disclosure is the minimum transparency requirement for high-stakes AI decisions.

03

No human review gate before rejection. Any system that can reject an applicant without a human ever seeing their application has eliminated Human Agency from a consequential decision. A minimum requirement: human review before any rejection determination is finalized.

What It Cost

The case is ongoing. The class action, if certified, could include a significant number of applicants who were rejected by Workday-powered screening tools without human review.

The EEOC amicus brief is the broader cost signal. When a federal agency files in support of an AI discrimination plaintiff, it signals that regulatory enforcement in this area is moving from theoretical to active. Every enterprise deploying AI hiring tools now faces a documented regulatory risk that did not formally exist at this scale two years ago.

The Lesson

A technology vendor whose AI makes consequential decisions is not a neutral tool. It is an agent with liability.

AI hiring tools occupy one of the highest-stakes Human Agency contexts possible: decisions that affect people's employment, income, and livelihood. The Atlas framework applies its most stringent requirements to systems in this category precisely because the consequences of failure are borne entirely by people who had no say in whether the system was deployed.

The Workday case establishes two principles that will shape AI accountability in employment for years. First: a technology vendor whose AI makes consequential decisions is not a neutral tool — it is an agent with liability. The Air Canada ruling established this for chatbots. The Workday ruling extends it to hiring platforms.

Second: historical data produces historical discrimination. An AI system trained to find "successful" candidates in an industry with a history of discriminatory hiring will reproduce that discrimination at scale, automatically, without any individual actor making a discriminatory choice. Algorithmic discrimination is not accidental — it is the predictable output of training systems on historically discriminatory data without bias evaluation and remediation.

References

  1. Derek Mobley v. Workday, Inc., Case No. 3:23-cv-00770-RFL, U.S. District Court for the Northern District of California, filed February 2023. District Judge Rita F. Lin.
  2. U.S. Equal Employment Opportunity Commission Amicus Brief in support of plaintiff, 2024.
  3. Reuters coverage of the Workday lawsuit.
  4. HR Dive coverage of the EEOC amicus brief.

FREE · 15 MINUTES

Book a free Atlas Readiness Review

Book Your Review →