The New Technology, the Same Old Discrimination: AI Hiring and Mental Health

Employment Testing: Failing to Make the Grade — Ten Years Later, Part 1 

Ten years ago, the final post on this blog addressed the growing influence of artificial intelligence and decision-support systems on employment, education, and other important opportunities. The concern was not simply whether computers could make decisions more efficiently than people. It was whether anyone could adequately measure the consequences of those decisions. 

A decade later, that question has become considerably more urgent.

 

Employers increasingly use automated applicant-screening systems, predictive assessments, and AI-assisted recruiting tools to evaluate job candidates. These technologies promise greater efficiency, consistency, and objectivity. 

But they also raise an old question in a new form: Are these systems identifying people who can perform a job, or are they excluding qualified applicants because their characteristics do not fit an algorithm's definition of the ideal employee? 

The Problem Was Never Limited to Technology 

In August 2013, this blog examined employment assessments marketed by Kronos and its subsidiary Unicru in Kroger and Kronos: Chaos and Disorder. Among other concerns, the post questioned whether assessments using personality characteristics and historical employee data could disadvantage applicants with mental health conditions. 

The underlying problem remains relevant today.

Consider an applicant who has experienced major depression but is fully qualified for a position. Another applicant may have post-traumatic stress disorder or bipolar disorder that has been successfully treated. A third may be autistic and possess exceptional technical skills but communicate differently from other candidates. 

An employment assessment that rewards a narrow range of emotional, social, or behavioral characteristics may disadvantage these applicants without reliably measuring their ability to perform the essential functions of the job. 

An algorithm does not have to ask whether someone has a psychiatric diagnosis to create such a problem. 

The ADA Still Matters 

The Americans with Disabilities Act (ADA) restricts disability-related inquiries and medical examinations before an employer makes a conditional job offer. It also prohibits certain qualification standards and selection criteria that unlawfully screen out qualified individuals with disabilities. These are related but distinct protections. 

In Karraker v. Rent-A-Center, 411 F.3d 831 (7th Cir. 2005), the court held that the use of the Minnesota Multiphasic Personality Inventory in an employment selection process constituted a prohibited medical examination because the test was designed, at least in part, to reveal mental illness. 

That decision does not mean every personality assessment is unlawful. Nor does it mean that every algorithm that evaluates behavior is a medical examination. But it does establish that calling something a personality test does not necessarily determine its legal status. 

Similarly, an assessment that is not a medical examination may still raise ADA concerns if it unlawfully screens out qualified applicants with disabilities. 

Artificial Intelligence Does Not Eliminate Human Bias 

An algorithm may identify patterns in employees previously labeled successful. But what if the employer's definition of success reflects longstanding hiring preferences, organizational culture, or discriminatory practices? 

The system may reproduce those preferences on a much larger scale. An applicant can be rejected without anyone knowing—or asking—whether a disability affected the assessment. 

The technology may be new. The underlying concern is not. 

Employers should have to demonstrate that their selection systems measure qualifications relevant to the job, rather than characteristics that unnecessarily exclude people with disabilities. That requires more than trusting the technology or accepting the vendor's assurances. It requires examining the actual selection process, its relationship to job performance, and its effects on applicants with protected characteristics. 

Ten years after this blog last examined artificial intelligence, the essential question remains: Are we using technology to identify talent—or to automate exclusion? Next: How seemingly neutral algorithms can disadvantage applicants without ever asking whether they have a disability.