In one of the most frequently cited studies in labor economics, researchers sent nearly identical resumes to real job postings and varied only the name printed at the top. The resumes were constructed to be equivalent in every measurable way: same experience, same education, same formatting. The only variable was a name chosen to read as White or Black, and, separately, as male or female.
The results were stark. Resumes with White-associated names were preferred in roughly 85% of comparisons where a preference was expressed, against roughly 8.6% for Black-associated names. Male names were preferred over female names in roughly 51.9% of comparisons, against 11.1% the other way. No resume in the study was more qualified than any other. The name did the work.
A resume with a White-associated name drew a callback preference in roughly 85% of comparisons. A resume with a Black-associated name drew one in roughly 8.6%. The only difference between the two was the name.
What changes when the reader is a model instead of a person
That study predates any of the AI hiring tools now screening resumes at scale, and the timing matters. The bias it documented was not introduced by automation. It was already there, in the judgment of human recruiters and hiring managers making fast, informal decisions about who to call back. The problem automation introduces is not the bias itself. It is the scale and the speed at which that bias gets applied.
An AI hiring tool trained on historical hiring data learns from a record of who got hired, promoted, and retained in the past. If that historical record reflects the pattern described above, even partially, a model trained on it doesn't need to be told to discriminate. It only needs to be accurate to the data, which means it can reproduce the pattern faithfully and apply it to every resume that crosses its desk, in seconds, at whatever volume the company is processing. A biased recruiter reviews dozens of resumes a day. A biased model can review thousands. The underlying decision is the same. The reach is not.
The legal exposure is not hypothetical
Employers do not get to treat a vendor's algorithm as a liability shield. In the United States, a hiring process that produces a disparate outcome along race or sex lines can expose an employer to claims under Title VII of the Civil Rights Act, regardless of whether a human or a model made the individual decision. Where a tool disadvantages older applicants, exposure runs through the Age Discrimination in Employment Act. Where it disadvantages applicants with disabilities, including through video analysis, resume parsing, or gamified assessments that assume a particular baseline of ability, exposure runs through the Americans with Disabilities Act.
Regulators and courts have been explicit that automating a decision does not automate away responsibility for its outcome. An employer that cannot explain why its screening tool rejects a disproportionate share of a protected group is not in a stronger legal position than an employer whose human recruiters did the same thing. In some respects it is in a weaker one, because the pattern is now documented in the tool's own output, at scale, in a form that is straightforward to audit and difficult to explain away as an isolated judgment call.
Automation is not an excuse
This is the sense in which automation doesn't fix the problem described by the callback studies. It industrializes it. A pattern that once depended on one person's unexamined assumptions can now run unattended, at the speed of the applicant pool, with no single human decision point where anyone might notice or intervene. That absence of a decision point is not a defense available to the employer that deployed the tool. It is the reason regulators are paying closer attention to hiring automation, not less.
Dealt has an obvious stake in this problem, since the entire premise of verification only holds up if the process built on top of it treats candidates fairly. But the legal and ethical case here doesn't depend on any particular company's product. It depends on a body of evidence, now decades old, that hiring bias is real, measurable, and, left unexamined, exactly the kind of pattern a model will learn to reproduce with total consistency.