AI – Melissa Brandt & Associates (MB&A)

Efficiency, bias, and the hidden false-negative problem in automated hiring

AI can support hiring without being entitled to decide who receives human consideration.

What happens when the first person to evaluate a candidate is not a person at all?

Employers are increasingly using automated assessments to screen applicants before a recruiter or hiring manager becomes involved. These systems may evaluate recorded interviews, written responses, games, simulations, technical exercises, and behavioral assessments and then help determine who advances.

The appeal is easy to understand. Automation promises speed, consistency, lower costs, and relief for recruiting teams overwhelmed by application volume. Every candidate can receive the same assessment, and thousands of results can be processed quickly. The process also produces something organizations find particularly persuasive: a score that appears objective.

But a standardized score is not the same thing as an objective judgment. Every automated assessment begins with human decisions about which qualities matter, how those qualities should be measured, and what successful performance looks like. The system then introduces another layer of risk through the data, correlations, proxies, and probabilistic inferences it uses to evaluate candidates. It can inherit human assumptions and biases, generate errors of its own, compound the two, and apply the result consistently at scale.

At each stage, uncertainty is converted into something that looks increasingly like fact. The greatest risk is not simply that the system might make a mistake. It is that qualified or exceptional candidates may be eliminated before anyone encounters them—and the organization may never acquire the evidence needed to discover what it missed.

Why the first wall matters

AI can reduce administrative burden through scheduling, workflow management, and information organization. Once candidate evidence exists, it can summarize structured interviews, map examples against a rubric, identify inconsistencies for human review, evaluate direct demonstrations of technical skill, and help recruiters compare large amounts of information. In those uses, a responsible person can examine both the evidence and the system’s interpretation of it. The ethical balance changes when the system controls who receives that human review, because an error at that stage can remove a candidate before anyone has the opportunity to recognize it.

Asynchronous video assessment introduces an additional limitation by removing the interaction that makes a skilled interview informative. A recruiter can follow an unexpected answer, distinguish nervousness from lack of competence, clarify an apparent contradiction, or discover that the original job specification is wrong. An automated assessment evaluates the candidate against assumptions established in advance. It cannot question those assumptions or negotiate meaning with the candidate.

These limitations matter most in executive, confidential, scarce-talent, succession, turnaround, and transformation searches. In those settings, evaluation and recruitment occur simultaneously. The strongest prospect may not be an applicant, may refuse an impersonal audition, or may understand the business problem differently from the organization that designed the assessment. That difference may be precisely the value the company needs.

The validation loop has a missing half

Once the system is in use, the employer can compare assessment scores with the performance of candidates who advanced and were ultimately hired. It cannot make the same comparison for rejected candidates because they never entered the organization and generated no performance data.

That distinction matters. Evidence that high-scoring candidates sometimes become successful employees does not prove that low-scoring candidates would have failed or that the system identified the best available talent. Because the assessment determines who becomes available for later measurement, it also shapes the evidence used to validate itself.

The system may therefore prevent the organization from collecting the very information that would reveal its false negatives.

If the system favors candidates who perform well in solitary recorded auditions, the company hires more of them and acquires more evidence about them. Candidates who think brilliantly in dialogue but speak awkwardly to a camera disappear from the dataset. The loop can become self-confirming.

Even the observed performance data are filtered through offer acceptance, manager quality, assignment distribution, early turnover, subjective ratings, and promotion politics. A model can end up optimizing for the organization’s existing definition of success while leadership believes it is identifying talent objectively.

Consistency may not equal fairness

Standardized questions and job-related scoring rubrics can improve a hiring process vulnerable to affinity bias, fatigue, intuition, and irrelevant impressions. But repeatability proves only that the same method was applied. It does not establish that the method is fair or that it measures the right things.

Bias can enter through the definition of success itself. If a model is developed around employees the organization already considers successful, it may learn patterns influenced by who received the best assignments, strongest sponsorship, most valuable territories, or greatest access to promotion. Historical performance data do not arrive morally neutral. They contain the history of the organization that produced them.

Automating those patterns does not neutralize them. It can make them more difficult to see and easier to apply consistently at scale.

The transcription problem comes first

Anyone who regularly uses voice-to-text knows that clear speech does not guarantee an accurate transcript. Proper names disappear, familiar words become unrelated ones, and regional pronunciations become statements the speaker never made. In a text message, that is inconvenient. In a scored employment assessment, it can alter the evidence used to evaluate a candidate.

Consider someone with a pronounced Boston accent, a Southern dialect, a speech disability, an unfamiliar cadence, or English as an additional language. If the transcription is inaccurate, the assessment may analyze words the candidate never said. Even if every word is captured correctly, another question remains: Is the candidate’s linguistic style genuinely related to job performance, or is it simply different from the speech most heavily represented in the system’s development and validation data?

Evidence suggests that these errors have not affected all speakers equally. In a 2020 peer-reviewed study of systems developed by Amazon, Apple, Google, IBM, and Microsoft, researchers found an average word error rate of 35 percent for Black speakers, compared with 19 percent for white speakers. Error rates also varied by location and were positively correlated with the density of African American Vernacular English features in the speech samples. Some systems have improved since the study was conducted, but progress has not been uniform, and employers may still be using older or insufficiently tested technology. When a transcript becomes the basis for a candidate score, uneven transcription error becomes an employment-selection risk.

Do not confuse probabilistic inference with knowledge of human potential. The system’s most dangerous error may be the candidate it rejects so early that the organization never learns what it lost.

This does not mean that all AI-enabled hiring tools present the same concerns. A representative technical exercise is fundamentally different from a system that attempts to infer empathy, initiative, judgment, or leadership potential from language or behavior. Platforms that exclude facial recognition or have discontinued facial analysis have removed important sources of risk, but questions involving transcription, competency design, proxies, scoring thresholds, validation, and human oversight remain.

The relevant question is not simply whether an employer uses AI. It is whether a particular tool, used for a particular role and decision, measures something sufficiently job-related to justify the authority it has been given.

The law recognizes risk, but compliance is not proof of wisdom

Federal employment law does not contain an AI exception. The Equal Employment Opportunity Commission has made clear that an employer’s use of AI can violate Title VII when a seemingly neutral practice produces unjustifiable disparate impact. The Americans with Disabilities Act creates additional concerns when a tool screens out qualified people with disabilities, fails to provide reasonable accommodation, or elicits disability-related information.

State and local regulation is developing as a patchwork:

Illinois. The Artificial Intelligence Video Interview Act requires notice, an explanation of how the technology works and what it evaluates, and applicant consent before AI analyzes a video interview. It also provides deletion rights and requires demographic reporting when an employer relies solely on AI analysis to determine who receives an in-person interview. Since January 1, 2026, the Illinois Human Rights Act has also prohibited discriminatory employment uses of AI and required notice of covered uses.

New York City. Local Law 144 requires covered automated employment decision tools to undergo an independent bias audit, publication of audit information, and advance candidate notice.

Maryland. Employers must obtain written consent before using facial-recognition technology to create a facial template during an employment interview.

California. Employment regulations effective October 1, 2025 clarify that automated decision systems can violate the Fair Employment and Housing Act, require covered automated-decision records to be retained for at least four years, and address disability-related inquiries. Separate privacy regulations establish additional automated-decision requirements on a phased schedule.

Colorado. Beginning January 1, 2027, covered automated systems used in consequential decisions, including employment, will be subject to disclosure and recordkeeping requirements, rights to correct inaccurate personal data, and an opportunity to request meaningful human review following an adverse outcome.

These protections matter, but a bias audit, disclosure, or consent form is not a certificate of truth. A system may produce similar selection rates across demographic groups while rejecting excellent candidates from every group because it measures conformity more effectively than potential. Nondiscriminatory, statistically valid, and strategically wise are three different standards.

The employer remains responsible for how the technology is deployed: which competencies are measured, how results are weighted, where cutoffs are set, whether accommodations and human review are meaningful, and whether the tool is appropriate for the role. Purchasing a validated product does not make every use of it valid.

A more defensible standard

Organizations do not need to choose between rejecting AI and surrendering judgment to it. A responsible system begins with job analysis, favors direct demonstrations of work over inferred personality, validates the full selection process for the actual role and population, tests transcription quality across relevant speech groups, provides an accessible alternative, and preserves meaningful human review before rejection.

It also examines false negatives deliberately. That may require independent human review of randomized rejected responses, parallel scoring in which the model does not control advancement, monitoring each selection stage for adverse impact, and tracking whether automation actually improves quality of hire rather than merely reducing recruiter time.

Those safeguards may reduce some of the promised efficiency. That is not evidence that they are excessive. It is evidence that consequential judgment was never costless.

The question is not whether AI belongs in hiring. It already does, and it can improve poorly structured processes. The question is what authority it should be granted.

Use AI where the evidence is direct, the method is demonstrably job-related, limitations are disclosed, candidates can request accommodation or correction, and a person remains accountable for the decision. Use it to organize information, reduce administrative burden, strengthen consistency, and surface questions worth investigating.