Workday Lawsuit and Studies Pin AI Bias on Human Choices

A federal judge in California has allowed key claims to advance against Workday over AI-powered job screening tools that allegedly filtered applicants by age, race and disability. The company denies the charges. The case sits inside a wider pattern: 2025 studies show large language models and real-world AI systems repeatedly produce the same skewed outcomes because people upstream chose the data, the labels, the risk tolerance and the teams.

Technical patches alone cannot fix what began as human decisions.

Mobley Suit Moves Past Early Dismissal

Derek Mobley, an African American man over 40 who also has anxiety and depression, applied to more than 100 jobs through systems that used Workday tools. He was rejected every time. He sued in 2023, arguing the AI scoring, sorting and ranking features created disparate impact under the Age Discrimination in Employment Act, Title VII and the Americans with Disabilities Act.

In July 2024 the Northern District of California denied Workday’s second motion to dismiss and allowed the theory that the vendor could face liability as an “agent” of employers. On 16 May 2025 Judge Rita F. Lin granted conditional certification of the ADEA collective, covering applicants aged 40 and older rejected since September 2020. In June 2026 the same court denied Workday’s bid to dismiss California Fair Employment and Housing Act claims and let an ADA claim proceed while dropping certain race and employer-status theories.

What We Know

• Workday’s tools are used by thousands of large employers worldwide.
• Conditional ADEA certification is active; discovery continues into mid-2026.
• Proxy indicators such as employment gaps are alleged to screen for disability.

What’s Unconfirmed

• Final class size or exact number of affected applicants.
• Whether intentional discrimination will be proved at trial.
• Ultimate liability split between vendor and client employers.

Workday has denied wrongdoing. The docket remains in pre-trial discovery as of August 2026. Legal observers note the agent theory could expose both the software maker and the more than 10,000 organisations that enable its AI features.

Two 2025 Studies Map the Same Pattern

Researchers at CSIRO’s Data61 and Monash University ran controlled tests that echo the lawsuit’s claims. In one exercise they fed OpenAI’s GPT-4 and Microsoft Copilot 300 synthetic software-engineer profiles, then asked each model to recommend candidates for junior, mid, senior and lead roles and to generate images of its preferred hires.

Both models preferred male and Caucasian profiles, especially for senior positions. The generated images consistently showed engineers who looked younger, slimmer and lighter-skinned. The paper, “What Does a Software Engineer Look Like?”, documents how the systems reproduced associations already present in language, imagery and employment records.

A second study manually reviewed entries in two public AI incident databases. It found that almost half of the analyzed AI incidents involved a diversity or inclusion issue. Racial, gender and age discrimination dominated. The harms traced to non-diverse training data and the neglect of diversity principles at design, development and deployment stages.

Muneera Bano, CSIRO Data61 research lead on D&I, co-authored both papers. The team has also released a public repository and decision tree to help others classify future incidents.

Where the Bias Enters

Data never arrives as a neutral mirror of reality. People decide what to collect, how to label it and whose experience counts. Those choices carry historical associations: leadership with men, technical skill with lighter skin, innovation with youth. Once encoded, models automate and scale them.

• Non-diverse training corpora that under-represent women, older workers, people with disabilities and non-Western populations.
• Design stages that skip intersectional testing (gender plus race plus age plus disability).
• Deployment without ongoing monitoring of who is rejected or ranked lower.
• Governance gaps that leave no single person accountable for reviewing live risk or incident response.

A system that passes single-axis fairness checks can still disadvantage people at the overlap of several identities. The studies and the Workday pleadings both surface that pattern.

Technical Patches Leave the Root Untouched

Re-balancing datasets, adjusting loss functions and adding post-hoc filters matter. They treat bias as a property of the model. Much of the bias originates outside it. When AI recommendations move into hiring platforms, education tools and public services, the same associations travel with them.

Western values and English-language assumptions travel too. Wealthy countries capture most of the economic upside while the harms fall more heavily on groups already underrepresented in AI development and leadership. The OECD AI incidents tracker continues to log cases that fit this description across sectors.

Intersectional Harm and Missing Voices

Bias rarely arrives as a single protected class. Age combines with race. Disability combines with employment gaps that algorithms treat as red flags. Gender combines with leadership stereotypes. The CSIRO team notes that testing each identity in isolation can hide the compound disadvantage.

Women and certain demographic groups remain underrepresented among the people who build and govern these systems. Affected communities are seldom at the table when data is selected or risk thresholds are set. That absence is itself a decision with measurable consequences.

Governance That Reaches the Decision Makers

Inclusive AI requires more than better code. Engineers need grounding in the social origins of bias. Development teams must measure whether benefits, errors and harms distribute fairly across groups, not only whether accuracy scores look good. Organisations that buy the tools need a named person accountable for live monitoring and incident response.

Participation from the people most exposed to the systems is not optional window dressing. It is how the missing information gets into the design loop. Several Australian and US public-sector frameworks already point in this direction; few private vendors have matched the practice at scale.

On X, reaction to the Workday rulings has focused on the practical stakes. Users note that resume tailoring becomes futile when the first gate is an automated filter used by thousands of employers. Others flag that discovery could force production of the exact scoring rules and client lists. One widely shared post observed that when AI leaders who raised bias concerns were sidelined, regulation stayed thin and the lawsuits arrived later.

The regulation of AI within hiring has historically been non-existent & when leaders in AI spoke out about it, they were ousted. Workday is struggling to answer these data calls because it’ll likely prove that AI bias exists and every company that used it should be investigated

Teneika Askew, Analytics & Automation, X post seen tens of thousands of times

Crowd commentary treats the case as a test of whether vendors can keep liability at arm’s length or whether the human chain of product, sales and client configuration will finally be examined in open court.

Liability Now Runs Both Ways

The agent theory surviving early motions means plaintiffs can reach the software provider even when the final hiring decision sits with a client company. Employers who enabled the features face their own exposure once discovery lists which customers turned on the AI scoring. Parallel suits against other hiring-tech vendors are already advancing on related theories, including Fair Credit Reporting Act claims.

None of this freezes AI use in recruitment. It does raise the cost of treating fairness as an afterthought bolted on after launch. The studies supply the empirical pattern; the lawsuit supplies the enforcement mechanism. Both point to the same place: the people who decided which data mattered, which risks were acceptable, and who was not in the room when those calls were made.

Algorithms do not set those parameters. People do. The current wave of litigation and research is the bill for earlier choices arriving due.