AI Hiring Tools Under Scrutiny: The Job Seekers Who Pay for Bias
Archi4 min read
Contents
The Test: Six Resume Screeners and a Few Dozen Fake Applicants#
We built 36 synthetic resumes. Same skill level, same years of experience, but different names, graduation years, and career paths. Then we ran them through six AI hiring tools that companies actually use to filter job applications. The tools ranged from simple keyword matchers to large language model based scorers.
What we found is not a story about rogue algorithms. It is a story about who gets filtered out when a machine decides what a good candidate looks like. The people hit hardest were not the ones with the weakest resumes. They were the ones whose careers did not fit the template.
The Setup#
Each resume had a matching job description. We used three roles: a mid level software engineer, a customer support manager, and a data analyst. For each role, we created six versions of a candidate with identical qualifications but different surface features.
One version had a graduation year of 2005 (age 43). Another had a graduation year of 2022 (age 25). One had a career gap of two years for caregiving. One had a name that sounds foreign to US hiring managers. One had a degree from a for profit online university. One was a career changer coming from a different industry.
We submitted each version three times to account for randomness. That gave us 648 total screening decisions.
The Results#
The career changer was filtered out 41 percent of the time, compared with 18 percent for the standard candidate. The older candidate was filtered out 34 percent of the time. The candidate with a caregiving gap was filtered out 29 percent of the time. The foreign sounding name made no measurable difference in our sample, which surprised us.
The tools rarely explained their decisions. When they did, the reasons were vague: "profile does not match our current needs" or "experience not aligned with team requirements." None of the six tools offered a way for a rejected candidate to request a human review.
Who Feels It Most#
The people most affected by these systems are not the ones with the most obvious disadvantages. They are the ones whose lives do not follow a straight line. A woman returning to work after raising children. A man who spent ten years in manufacturing and retrained as a data analyst. A veteran whose military job titles do not map cleanly to civilian roles.
These are not edge cases. According to the Bureau of Labor Statistics, the median tenure of American workers is about four years. People change jobs, take breaks, and switch fields constantly. The resume screener treats those normal life events as risk signals.
The Feedback Loop#
Here is the part that worries me most. The tools learn from historical hiring data. If a company has historically hired mostly 25 year olds with uninterrupted careers, the model learns that pattern and reinforces it. Each year, the pool of accepted candidates looks a little more like the last one.
One tool we tested had a feature called "culture fit score." It compared a candidate's resume language to the language used by current employees in their internal profiles. That sounds neutral until you realize that most of those employees were hired through the same biased process. The score was measuring similarity to the past, not potential for the future.
What Candidates Can Do#
If you are applying through one of these systems, there are a few practical moves. Use the same job title on your resume that appears in the posting. Spell out acronyms. Put your most relevant experience in the first third of the page. Avoid tables and graphics that some parsers cannot read.
But those are workarounds, not solutions. The real fix has to come from employers. Ask your hiring team whether they use an automated screener. Ask if there is a human review step. Ask how the tool was validated. If they cannot answer, that tells you something about how much they care about fairness.
The Regulatory Gap#
New York City's Local Law 144 requires annual bias audits for automated employment decision tools. We tested three tools that claim to comply. None of them would tell us how they measure bias across age or career gaps. The law focuses on race and gender, which are important. But a tool can be perfectly fair on race and gender and still reject every candidate who took a year off to care for a parent.
That is the gap. The rules measure the wrong things, and the people who fall through it are the ones whose lives are messy in ordinary ways.
A Note on Method#
We did not name the vendors because our goal is to describe the pattern, not to single out one product. The pattern held across all six tools. Some were worse than others. The best one still filtered the career changer 22 percent of the time. The worst filtered that candidate 58 percent of the time.
We also ran a control test with human reviewers. Three recruiters looked at the same 36 resumes without knowing about the synthetic setup. They filtered the career changer 20 percent of the time and the standard candidate 15 percent of the time. Humans have bias too. But humans can explain their reasoning, and candidates can appeal to a person. The machines offer neither.
That is the tradeoff companies are making when they adopt these tools. They save time on screening. They lose the ability to see the candidate in front of them. And the people who pay the price are the ones who already face the most friction in the job market.
More articles