EEOC: AI Guidance in Assessing Adverse Impact in Algorithmic Decision-Making Tools

The EEOC issued guidance titled 'Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII' to clarify how existing civil rights law applies to the use of algorithmic tools in hiring, promotion, and other employment decisions. The core principle is that Title VII's adverse impact (disparate impact) prohibition applies to algorithmic selection procedures just as it does to traditional selection methods like written tests or structured interviews. An AI tool that disproportionately screens out applicants of a particular race, gender, or other protected class—even without discriminatory intent—may violate Title VII if the tool cannot be shown to be job-related and consistent with business necessity. The guidance establishes that employers cannot escape liability by using a third-party vendor's algorithm: the employer remains responsible for the discriminatory outcomes of selection procedures it uses, regardless of whether it developed the tool internally or purchased it. This has significant implications for the growing use of AI in resume screening, candidate scoring, video interview analysis, and other employment selection contexts, requiring HR professionals to actively assess and monitor algorithmic tools for disparate impact rather than assuming vendor-provided technology is compliant.

Adverse impact (also called disparate impact) occurs when a selection procedure produces substantially different selection rates between demographic groups, disadvantaging members of a protected class under Title VII even without discriminatory intent. The four-fifths (or 80%) rule is the primary statistical benchmark used to identify potential adverse impact: if the selection rate for any protected group is less than four-fifths (80%) of the selection rate for the highest-selecting group, adverse impact is indicated and triggers further analysis. For example, if 50% of white applicants pass a screening tool but only 30% of Black applicants pass (30/50 = 60%, below the 80% threshold), the tool has produced adverse impact against Black applicants. Applying this analysis to algorithmic tools requires access to selection outcome data disaggregated by race, gender, and other protected characteristics—data that employers must be proactively collecting for any selection procedure they use. Algorithmic tools present unique challenges for this analysis: selection criteria may be complex and non-transparent, making it difficult to identify which specific features of the algorithm are driving the disparate outcomes. Employers using AI selection tools should require vendors to demonstrate adverse impact analyses of their tools and should conduct their own internal analyses using their actual applicant population data rather than relying solely on vendor-provided testing results.

The EEOC's guidance makes clear that employer responsibility for Title VII compliance does not transfer to third-party software vendors when those vendors' tools produce discriminatory outcomes. Employers are liable for the discriminatory effects of selection procedures they use, regardless of whether they developed the tool internally or purchased it from a specialized vendor. This creates concrete due diligence obligations for HR professionals selecting and managing AI-based hiring tools. Before deployment, employers should require vendors to provide documented evidence of adverse impact testing conducted on the tool—ideally using demographically representative testing populations—and should understand what protected-class data, if any, was used in training the algorithm. Contracts with AI vendors should include representations and warranties about the tool's compliance testing, provisions for ongoing monitoring data sharing, and contractual responsibilities for updates when adverse impact is identified. During use, employers should independently monitor their own applicant flow and selection outcome data for adverse impact patterns rather than relying solely on vendor assurances. Key questions to ask vendors include: What protected-class data was used in the training dataset? How was the tool validated for job-relatedness? Has adverse impact analysis been conducted, and what were the results? How are model updates tested for disparate impact before deployment?

Self-analysis for adverse impact in algorithmic hiring tools requires HR professionals to collect, analyze, and act on selection outcome data disaggregated by protected class characteristics. The first step is data infrastructure: ensuring that applicant tracking systems capture both demographic self-identification data and selection outcomes at each stage of the hiring process where an algorithmic tool is applied—initial screening, assessment scoring, interview scheduling, or final selection. With this data in hand, HR can apply the four-fifths rule and statistical significance tests to identify whether meaningful selection rate differences exist across protected groups at each stage. When adverse impact is indicated, the analysis must go deeper: is the differential explainable by legitimate, job-related factors, or does it reflect algorithmic bias? Job-relatedness analysis examines whether the tool's selection criteria are predictive of actual job performance—if they are not, the business necessity defense is unavailable. Where adverse impact is identified and cannot be justified by job-relatedness and business necessity, employers should explore whether alternative procedures exist that achieve the same selection purpose with less adverse impact. Proactive self-analysis, documented and acted upon, demonstrates the good-faith compliance effort that reduces legal exposure and enables organizations to course-correct before facing regulatory scrutiny or litigation.

Responsible implementation of AI hiring tools requires a governance framework that treats algorithmic fairness as an ongoing operational commitment rather than a one-time vendor due diligence exercise. Before deployment, conduct a thorough vendor evaluation that includes adverse impact testing documentation, validation evidence demonstrating job-relatedness, and clear contractual obligations for ongoing compliance support. Involve HR, legal counsel, and DEI leadership in the evaluation process—not just IT or procurement. Establish a baseline of current adverse impact data before deploying new tools so that the tool's impact can be isolated from other selection procedure effects. During deployment, maintain robust applicant flow and selection outcome data collection disaggregated by protected class, and establish regular review cadences—at minimum quarterly for high-volume tools—to monitor for emerging adverse impact patterns. Train HR staff and hiring managers on the tool's appropriate use and limitations: AI tools should support, not replace, human judgment in selection decisions. Document decisions to use, modify, or discontinue tools based on compliance analysis. Engage external audit support periodically to validate internal analyses with independent expertise. As AI regulation in employment continues to evolve—with state and local laws in New York City, Illinois, and elsewhere now imposing specific AI hiring tool requirements—staying current on applicable legal obligations is an ongoing HR compliance responsibility.