AI Hiring Weights and Bias: Why the Algorithm Doesn’t Clear the Employer
An algorithm does not need a “race” field to discriminate. It uses weights.
Hiring software turns a résumé into numbers. Those numbers get multiplied by weights—how much the model cares about school name, years without a gap, job titles that match last year’s hires, zip code, graduation year, or words that look like past “successful” employees. The output is a rank or a 0-to-5 score. If that score never reaches a human, the weight is the hiring decision. When those weights systematically bury qualified people in protected groups, federal law still treats it as a selection procedure. Buying the software from a vendor does not hand the employer a get-out-of-liability card.
How the math encodes bias. Most screening tools are supervised models or language-model rankers trained on historical applications and who got interviews. If the company’s last decade of hires were mostly young men from a short list of schools, the model learns that pattern as “fit.” Amazon’s internal tool, trained on ten years of résumés in a male-heavy tech pipeline, downgraded applications that contained the word “women’s.” The system was not broken. It was doing what the weights asked. Researchers call the same mechanism proxy encoding: strip race, gender, and age from the file, and the model still uses zip code, graduation year, college name, employment gaps, and even name embeddings that correlate with those traits. University of Washington audits of language-model résumé rankers found white-associated names preferred in about 85% of paired comparisons and Black-associated names in about 9%, with the steepest penalty at the intersection of race and gender for Black men. That is not a recruiter’s hunch. It is a documented ranking artifact.
Weights also punish people who do not look like incumbents: career changers, caregivers with résumé gaps, older workers whose timelines do not match a 25-year-old’s first job, applicants from less-famous schools. Illinois now bars zip codes as proxies in employment AI. California regulates automated-decision systems even when a human clicks “approve” at the end. Those rules exist because “the model decided” is not a legal category.
Liability stays with the employer—and can reach the vendor. Title VII, the ADEA, and the ADA apply to automated screens the same way they apply to a paper test. Disparate-impact theory looks at outcomes, not motive. If a tool selects a protected group at a much lower rate than the highest-selected group, the employer must show the practice is job-related and consistent with business necessity. A vendor badge that says “bias-free” is marketing. Courts have not treated it as a shield.
In Mobley v. Workday, a Black applicant over 40 with a disability alleged he was rejected from more than 100 jobs routed through Workday’s screening tools. A federal judge allowed claims to proceed on the theory that Workday can be an agent of the employers that use it—because the software scores, sorts, and gates applicants the way an HR department would. The EEOC filed an amicus brief supporting that agent theory. Workday denies discrimination. The practical message for every company using the platform is joint exposure: the vendor may be in the lawsuit, and so may the employer that turned the feature on. Similar theories appear in Harper v. SiriusXM (alleged racial proxies in a vendor tool) and in FCRA-style claims that scored “talent profiles” are undisclosed consumer reports. Federal enforcement of disparate-impact cases has been deprioritized since 2025; the statutes and private lawsuits remain.
What that means if you were passed over. A five-minute rejection after you applied through a corporate portal is often a weighted model, not a person. You generally cannot see the score. You can still document the pattern, ask whether automated tools were used, request an ADA alternative to a video or game assessment, and remember that “our vendor handles screening” is not a complete answer if you later pursue a charge. For employers and staffing partners, the audit question is simple: which features are weighted, which are proxies, and can you defend each weight as a real job requirement—not a portrait of last year’s staff.
Sources: Reuters on Mobley v. Workday, HR Executive on vendor certifications, Employment Law Letter on employer liability, Brookings / UW résumé-screening study, Feature selection and proxies (arXiv review), EEOC agent theory discussion.
AI hiring algorithm feature weights and proxy discrimination
employer liability for vendor AI resume screening bias
Mobley v Workday algorithmic hiring disparate impact
zip code graduation year proxy variables in hiring software
does using Workday or Eightfold absolve company of Title VII
AI Hiring Weights and Bias: Why the Algorithm Doesn’t Clear the Employer
Screening models weight school, gaps, and proxies that can bury qualified applicants. Vendor software does not cancel Title VII, ADEA, or ADA liability. Here’s how the weights work.
Comments