An AI resume screener rejects every applicant with a career gap longer than six months. The cut falls hardest on women who took parental leave and on people who cared for sick family members. No one set it up that way. The tool learned the pattern from the company’s own hiring history, and HR finds out from a rejected candidate’s semi-viral LinkedIn post. Rebecca Taylor brought that scenario to Lawrence Covitt, VP and Head of HR at Kelly Spicers, for HR Voices.
Lawrence opened with a disclosure: he is “a huge proponent of AI.” He did not read the scenario as a malfunction. “The scary part is that the AI functioned exactly as it was designed to do,” he said. It found a pattern in past hiring and adopted it as a best practice, leaving HR to explain a disparate impact that nobody chose.
Why the Screener Did Exactly What It Was Built to Do
Lawrence used to teach at FIT, where many faculty wanted AI banned as cheating. He went the other way. He almost mandated it, on one condition borrowed from math class: students had to show their work, down to every prompt.
A tool built without the parameters and barriers a company wants, he argued, will do what it thinks is best. This one probably found that the company rarely hired people with six-month gaps, and took that as a lesson.
At a previous job, he and his boss began rolling out AI around 2022. What they learned is that the technology does not create bias. It “discovers it, operationalizes and then it scales the bias that probably already exists,” including bias no one intended.
Gap filters are not rare. In research on hidden workers from Harvard Business School and Accenture, almost half the companies surveyed weeded out resumes with a work gap, and a gap of more than six months could trigger automatic rejection. Such filters, the researchers noted, cannot infer what caused a gap.
Why the Final Human Decision Comes Too Late
Companies often reassure people that a human still makes the final decision. Lawrence called that appropriate but insufficient. If AI narrows 2,000 applicants to 80, the recruiter sees 80. For the other 1,920, “the consequential decision may have already occurred.”
During the search that led to his current role, LinkedIn placed him past the 4,000th applicant for one opening. At that scale, he argued, companies hire the best of the people they can see instead of the best person for the job.
Since no team can read every application, the question is what the filter measures. Lawrence framed it as a choice between a job-related criterion and “an easily measurable proxy that just happens to correlate with historical data.” Title VII draws a similar line: an employer whose practice causes a disparate impact by sex, race, color, religion, or national origin has to show it is “job related for the position in question and consistent with business necessity.”
Lawrence called rejecting gaps an old-school way of thinking. People step away for parental and family leave, to care for an ill parent or spouse, or for disability, medical treatment, military obligations, layoffs, or education.
What Actually Works
Measure Adverse Impact Before Launch and After
Lawrence laid out the governance he expects: validation before deployment, adverse impact testing, ongoing monitoring, and a path to escalate problems to a human.
New York City’s rules on automated employment decision tools already bar employers from using one unless it has had a bias audit within a year of use, the audit information is public, and candidates or employees get notice.
Give the Tool Rules, and Let It Ask Questions Back
Lawrence’s model for a well-governed tool is a “super smart infant.” It does not understand racial or religious bias until someone tells it, and it can ask questions back.
His team is using AI on a Fair Labor Standards Act review of which roles are exempt or non-exempt, comparing job descriptions with salary and supervision rules across roughly 15 states. It advises; humans decide. When a manager insisted a new role was salaried, Lawrence disagreed and put the question to the “impartial AI.” It pushed back, and the exchange settled it.
Build a Way for Candidates to Be Heard
A LinkedIn post should never be how HR learns that its recruiting technology screens out people in a protected class, Lawrence argued. The gap sits in monitoring, starting with candidate experience.
He described a candidate whose recruiter said, “I like you a lot more than I thought I was going to,” then raised concerns her resume already answered. She had nowhere to take it. Lawrence’s own floor is that every applicant hears back. “Might be a form letter, but you get a no,” he said. His team hates the rule, and it is his answer to the ghosting that erodes trust in the candidate experience.
Where Employee Relations Fits
Once the post is public, Lawrence would pause the tool first. Then the work is investigative: count how many other candidates were affected, run an adverse impact analysis, and look at how the company treated gaps before AI. Screened-out candidates may deserve another look, with employment counsel and the vendor involved.
His first question is not how to answer the post. It is whether the candidate is right, and how to make it right. Marketing, which owns PR and social media at his company, is his partner on damage control, since even a semi-viral post can hurt a company’s reputation as an employer. Saying nobody programmed the tool that way, he added, “is not going to be a defense in court.” Rebecca’s advice was a statement that shows its work: what the company found, how it is fixing it, and what it does not know yet.
The candidate here had nowhere to go but LinkedIn. Lawrence wants recruiters able to spot warning signs so HR hears first, which is the argument for anonymous reporting that feeds case management.
Frequently Asked Questions About AI Resume Screening
Yes. Nobody configured the screener in this scenario to reject career gaps. It learned the pattern from past hiring data, and Lawrence Covitt’s point is that it worked exactly as designed.
Because the screen acts first. If a tool narrows 2,000 applicants to 80, recruiters only review 80. Lawrence argues the consequential decision about the other 1,920 may already have happened.
Not as a blanket rule, in Lawrence’s view. Gaps often have ordinary causes, from parental leave and caregiving to layoffs, so he asks whether a criterion measures the job or only an easy proxy.
Pause the tool, then find the facts. Lawrence would ask whether the candidate is right and how many other applicants the screen affected before drafting any response.
Yes, at Lawrence’s company. Every applicant hears back however far they got, even if the answer is a form letter saying no.
The Bottom Line for HR Leaders
Lawrence’s framing refuses both easy readings of the scenario. The tool is not a villain, and it is not an alibi. His answer was more governance, not less technology, with guardrails set by the leaders and hiring managers who own the outcome.
Lawrence Covitt holds people to the same standard. Teach an AI a lesson and it learns it, he said, while people can sit through DEI training going through the motions. Either way, the work is checking what was learned before it reaches a candidate.
See how AllVoices helps recruiting and HR teams surface hiring concerns before they go public.
