Why an AI Hiring Test Cannot Measure the Performance of People It Never Hired
Imagine an employer that hires a technology company to develop an AI assessment predicting which applicants will become successful employees.
The company examines personality assessments, employment records, and performance evaluations. It identifies characteristics associated with stronger performance and develops a test favoring applicants with those traits.
The results appear promising. Assessment scores show a statistically significant relationship with job performance.
But there is a fundamental problem.
The company knows how its hired employees performed. It does not know how the rejected applicants would have performed.
And some of those rejected applicants might have become its best employees.
The Performance Records That Do Not Exist
Consider an employer that receives 10,000 applications for 500 positions.
Every applicant completes an assessment measuring personality characteristics, communication styles, and other traits the employer considers important.
The employer uses those scores, along with other information, to select 500 applicants.
Over the next two years, it evaluates their performance.
The technology company now has two sets of information for those employees: their original assessment scores and subsequent job performance.
It can examine whether higher scores predicted better performance among those hired.
But what about the other 9,500 applicants?
The company knows their assessment scores and that they were rejected. It does not know how they would have performed if hired.
Some might have struggled. Others might have performed equally well—or better.
The employer has performance data for the people it selected, not for those its selection process excluded.
That distinction matters when those records become the foundation for future hiring standards.
When the Sample Has Already Been Filtered
This problem involves two related statistical concerns: missing performance outcomes and range restriction.
Missing outcomes arise because rejected applicants never perform the job. Their potential performance cannot be directly observed.
Range restriction occurs when employees included in a study represent a narrower range of characteristics or assessment scores than the broader applicant population.
Suppose an employer historically favored applicants who scored highly on extraversion and assertiveness.
Its workforce may contain relatively few employees with lower scores.
A study comparing personality scores with job performance will therefore provide limited direct evidence about how less extroverted or assertive applicants might perform.
Researchers have statistical methods for addressing range restriction, but those techniques depend on assumptions about the selection process and available data.
The problem becomes more complicated when hiring decisions involve interviews, recommendations, and multiple assessments.
A relationship observed among selected employees does not automatically establish the same relationship among all qualified applicants.
That does not make an assessment invalid. It means researchers must examine whether the evidence supports applying their findings to a broader population.
The Disability Connection
These limitations are particularly important when historical hiring practices may have disadvantaged applicants with disabilities.
Consider an employer that traditionally favored candidates who appeared outgoing, enthusiastic, emotionally expressive, and comfortable with conventional workplace interactions.
Qualified applicants with autism, depression, anxiety disorders, or other disabilities may have been disadvantaged by those preferences.
Some may have been rejected because their communication styles or emotional expressions differed from what interviewers expected.
An AI assessment developed from the performance records of previously hired employees may therefore contain limited information about the potential performance of excluded applicants.
Yet the algorithm may learn that applicants resembling existing successful employees are more likely to succeed.
The absence of evidence that excluded applicants would succeed is not evidence that they would fail.
Under the Americans with Disabilities Act, an assessment that screens out or tends to screen out qualified individuals because of disability raises important legal questions.
The employer may need to establish that the challenged criterion is job-related and consistent with business necessity. Reasonable-accommodation obligations may also apply.
Statistical limitations do not automatically establish discrimination. But favorable validation results do not automatically establish ADA compliance, either.
What Does a Validation Study Actually Prove?
Employment assessments are often evaluated by examining whether test scores predict subsequent job performance.
This approach, known as criterion-related validation, is an established method of evaluating employee-selection procedures.
Properly conducted, these studies can provide valuable evidence about whether an assessment predicts relevant aspects of performance.
But their conclusions depend on the data, research design, and population studied.
Suppose an assessment accurately predicts performance among the 500 employees hired in our example.
That finding may be meaningful.
What it does not necessarily establish is that the other 9,500 applicants lacked the ability to perform the job—or that the employer selected the best 500 candidates.
Those conclusions require additional evidence.
The federal Uniform Guidelines on Employee Selection Procedures recognize the importance of appropriate validation methods, including consideration of restricted score ranges and the populations to which findings will be applied.
An assessment can predict performance among selected employees without proving that rejected applicants would have performed poorly.
The question is not whether every rejected applicant's performance must be observed. That is generally impossible.
It is whether the study provides a sound basis for drawing conclusions about applicants whose performance was never measured.
Why More Data May Not Solve the Problem
Artificial intelligence can analyze enormous amounts of employment information.
A vendor may possess records from hundreds of employers and millions of applicants.
But database size does not necessarily resolve its limitations.
Suppose a database contains assessment scores for one million applicants, but performance outcomes for only the 50,000 who were hired.
The system has extensive information about those employees.
It still lacks direct evidence of how the other 950,000 applicants would have performed.
If participating employers used similar hiring criteria or assumptions about desirable personality characteristics, the missing information may be especially important.
A larger database may provide more precise estimates of relationships among the employees studied without establishing whether those relationships apply equally to excluded applicants.
More data cannot, by itself, supply performance outcomes that were never observed.
AI can estimate those missing outcomes, but its predictions depend on assumptions and patterns learned from existing information.
Predictions are not the same as observed performance.
What Responsible Validation Requires
None of this means employment assessments are inherently unreliable or that existing-employee data have no legitimate value.
Researchers have long recognized the problems created by restricted samples and missing performance information.
They have developed methods for addressing some of these limitations, including statistical corrections, independent validation samples, and research examining different applicant populations.
But those methods require sound assumptions and careful interpretation.
Employers and vendors should examine how employees in validation studies were selected, whether important applicant groups are underrepresented, and whether findings can reasonably be applied to applicants with different characteristics.
They should also determine whether an assessment measures genuine job requirements and whether qualified applicants with disabilities may be improperly excluded.
The objective is not to guarantee that every rejected applicant would have succeeded.
It is to establish meaningful evidence of job-related ability without unlawfully screening out qualified individuals.
The People an Algorithm Cannot See
The previous articles in this series examined two related problems.
When the Workforce Becomes the Test explained how historical hiring decisions can produce a workforce that does not represent everyone capable of performing a job.
When Success Becomes the Test examined how biased performance evaluations can cause AI assessments to learn supervisors' preferences rather than genuine measures of ability.
There is a third problem.
Even when performance evaluations are accurate, they exist only for people who were given the opportunity to perform.
An AI assessment may identify genuine relationships among employees who were hired. It may even outperform traditional hiring methods.
But those findings do not establish that applicants receiving lower scores would necessarily have performed poorly.
Some may possess abilities the assessment fails to recognize. Their potential contributions remain unobserved.
An algorithm cannot learn directly from the job performance of people who were never given the opportunity to perform the job.
And that raises a question every employer relying on AI hiring assessments should consider:
How many capable applicants are missing from the data used to determine who is capable?
Sources and Further Reading
U.S. Equal Employment Opportunity Commission and other federal agencies, Uniform Guidelines on Employee Selection Procedures, 29 C.F.R. Part 1607, particularly §§ 1607.5, 1607.14, and 1607.15.
Society for Industrial and Organizational Psychology, Principles for the Validation and Use of Personnel Selection Procedures, Fifth Edition (2018).
U.S. Department of Justice, Algorithms, Artificial Intelligence, and Disability Discrimination in Hiring (May 12, 2022).
Next: When Retention Becomes the Test—how employment data can cause AI to mistake workplace barriers for employee characteristics.
No comments:
Post a Comment
Because I value your thoughtful opinions, I encourage you to add a comment to this discussion. Don't be offended if I edit your comments for clarity or to keep out questionable matters, however, and I may even delete off-topic comments.