Recruiting AI: The Bias Measurement Gap That Systems Struggle to Address

The evidence: adoption rising, measurement stalled

Three independent signals converge on a distance every talent strategy must address. AI adoption for resume screening has grown from approximately one quarter to over 40% of HR teams between 2024 and 2025, according to SHRM data.

Over the same period, evidence on the quality of these systems has deteriorated. A 2024 University of Washington study tested three production selection models across over three million resume-to-position comparisons. The sample size and scope make the finding difficult to dismiss as an isolated case.

Names associated with white candidates were preferred in 85% of cases, those associated with Black candidates in just 9%, according to analysis covering the study^[1]. In direct comparisons between a Black man and a white man with identical resumes, the white name won in nearly every case.

The models functioned correctly. They reproduced the pattern already present in the historical hiring data on which they were trained. The consequence for organizations purchasing these systems is clear: the technical accuracy of the model says nothing about the fairness of its outcomes.

What the model actually sees

Most selection bias emerges from what the model is authorized to see, rather than from unpredictable behavior.

A system trained on video interviews implicitly evaluates tone of voice, accent, background and appearance, regardless of intent. A system trained solely on resumes evaluates pedigree proxies.

These proxies include school prestige, employer brand recognition and keyword density. They track more closely who wrote the resume than what the person actually knows how to do.

Here lies the structural problem. The algorithm inherits the social geography of the input data and converts it into an apparently neutral score. The score obscures the origin of the judgment, making it harder to contest than a decision made by a person.

Bias persists, changes form

The promise of AI in recruiting was to remove human judgment distorted by bias from the process. The evidence shows a different outcome.

When an organization replaces manual screening with an algorithm trained on its own past hiring, it encodes those same choices into a replicable rule. Discrimination becomes faster, more consistent and harder to see.

The critical point concerns measurement. Most systems in use lack a methodology to establish whether bias has been eliminated or simply shifted further down the pipeline.

Screening that discards obvious racial proxies can still penalize postal codes, educational pathways or employment gaps that correlate. The score appears clean, the outcome remains imbalanced. Without an outcomes test, the organization has no way to distinguish the two cases.

The regulatory window shifted, the problem remains

The EU AI Act classifies AI applied to hiring as high-risk, under Annex III. That classification carries with it a precise compliance regime.

The regime requires risk assessments, technical documentation, human oversight and transparency to candidates. The original deadline fell in August of this year.

The Digital Omnibus on AI, published in the Official Journal in July, moved that deadline to December 2027. The bias problem remained where it was. The postponement shifts the formal obligation, not the substantial exposure of the system.

Regulators will ask the same question the University of Washington study was designed to answer: what is your system actually evaluating, and how do you know? Those waiting for the new deadline to respond will have less time than the postponement makes it appear.

Why compliance misses the mark

Regulatory compliance verifies the presence of documents, processes and statements. It rarely verifies the statistical outcome of the system on protected groups.

An organization can produce an impeccable risk assessment and maintain an algorithm that systematically favors a particular demographic profile. The two levels operate separately.

Here lies the legal and reputational risk. Systematic discrimination in invisible form passes formal controls and remains exposed to individual litigation, audits and journalistic investigation.

The Guardian documented in August 2026 how AI screening tools continue to generate discriminatory outcomes in English-speaking labor markets, in dedicated coverage^[2]. Media attention arrives well before the regulatory deadline. The reputational cost therefore matures on a different calendar than compliance.

What high-performing organizations do

Organizations closing this gap treat bias measurement as a design condition, rather than as an individual choice of the recruiting team.

They start with an architectural question: what the system is authorized to see. They deliberately restrict input to verifiable skills and remove pedigree proxies from the pipeline.

Then they set up continuous outcome testing. They compare advancement rates between demographic groups at every stage of the funnel, rather than trusting the vendor's declared neutrality.

Braintrust, for example, built its AIR recruiting product around a tighter constraint on input. The approach recognizes that controlling the input data precedes any policy.

The design question for CHROs and boards

For the CHRO the priority becomes one: demand outcome evidence from vendors, rather than compliance reassurances. The operational question is which fairness metric is monitored at every stage.

For the CEO the conversation to bring to the board concerns exposure. An unaudited selection algorithm is a human capital and legal risk that sits outside the radar of formal compliance.

For the Talent & Compensation Committee the metric to monitor is the advancement differential between groups across the selection funnel. It is board-level data, measurable every quarter.

The distance between 40% adoption and the share of systems that actually measure their own bias describes the governance challenge of this evidence. Those who close it now build a defensible advantage.