Abstract
The US Equal Employment Opportunity Commission’s technical assistance document on AI in employment decisions, issued in 2023, placed the EEOC squarely in the position of applying existing disparate impact doctrine to algorithmic hiring tools - without new statutory authority and without a formal rulemaking that carries enforcement weight. Two years on, the compliance landscape is fractured: some large employers have begun bias audits of their hiring software stacks; many have not; and the audit methodologies in use vary enough that comparability across vendors is essentially impossible. The practical consequence is that Title VII protections apply in theory to algorithmic hiring discrimination, but the evidentiary infrastructure to prove harm does not yet exist.
What Disparate Impact Doctrine Requires and Where It Breaks Down
Title VII’s disparate impact theory requires showing that a facially neutral practice causes statistically significant adverse impact on a protected class, and then shifting the burden to the employer to show business necessity. Applied to algorithmic tools, the first step is straightforward in concept: compare selection rates across demographic groups. In practice, this requires demographic data on applicants, which many employers do not collect for top-of-funnel screening tools, and outcome labels, which are often unavailable because rejected applicants are never hired and their counterfactual performance cannot be observed. Resume screening tools from HireVue, Workday, and their competitors are often deployed as black boxes; the vendor audit may not cover the employer’s specific configuration or data pipeline.
The Vendor Audit Problem
New York City’s Local Law 144, which took effect in 2023, required bias audits of automated employment decision tools before use in New York City hiring, with results published annually. The law’s implementation has been more instructive as a case study in regulatory limits than as an effective control: audit firms authorized under the law varied widely in methodology, some audits were conducted on proprietary test datasets rather than production data, and the disclosure requirement created a published compliance artifact that did not translate into meaningful external accountability. Illinois and California have passed disclosure and notice requirements; none currently mandate third-party access to production data for audit purposes.
What Meaningful Accountability Would Require
A legally defensible and practically useful audit regime for algorithmic hiring tools would require access to applicant flow data at demographic granularity, the ability to test the production system with controlled inputs, and an adverse impact threshold with agreed statistical methodology. The Society for Industrial and Organizational Psychology’s guidelines on worker selection and the EEOC’s Uniform Guidelines on Employee Selection Procedures provide a starting statistical framework; neither was designed for high-dimensional embedding-based models. Closing the gap requires either vendor cooperation that goes beyond marketing-friendly audit programs, or regulatory authority that compels it - and neither is currently available at federal scale.