If you hesitate before letting an AI score your applicants, you are not being paranoid. You are being well-read. The horror stories are real, the research on bias is peer-reviewed, and the most common reassurance in the industry, "don't worry, a human reviews everything," turns out to be much weaker protection than it sounds.
This post walks through the actual evidence, then makes a specific argument: the difference between AI screening you can trust and AI screening you cannot is not whether a human is in the loop. It is whether the human gets reasons they can check, or just a number they can only accept.
The horror stories, told accurately
Three stories come up in every conversation about AI hiring, and they deserve to be told with their caveats attached, because the details matter.
Amazon's scrapped recruiting engine. In 2018, Reuters reported that Amazon had abandoned an internal, experimental recruiting tool after discovering it penalized resumes containing the word "women's" (as in "women's chess club captain") and downgraded graduates of all-women's colleges. The tool had learned from a decade of past hiring data, and the past was biased. Worth noting: it was experimental and never used in production. Amazon caught it. The lesson is not "Amazon did a bad thing"; it is that even a company with world-class ML talent could not train the bias out of a model that learns from historical hiring decisions.
The iTutorGroup settlement. In 2023, tutoring company iTutorGroup paid $365,000 to settle an EEOC lawsuit after its application software automatically rejected more than 200 qualified applicants: women 55 and older, and men 60 and older. Precision matters here too: this was rule-based auto-rejection, not machine learning, and it was settled by consent decree. But that is exactly why it is scary. The simplest possible "screening automation" produced age discrimination at scale, and nobody inside the company stopped it before the EEOC did.
Jared and lacrosse. An employment attorney told Quartz about auditing a resume-screening algorithm whose two strongest predictors of job performance turned out to be being named Jared and having played high school lacrosse. One anecdote, from one audit, but it captures the failure mode perfectly: a model will happily optimize on proxies that correlate with your past hires and have nothing to do with the job.
The peer-reviewed evidence is worse than the anecdotes
Anecdotes can be dismissed. The University of Washington's research program on AI hiring cannot.
In a 2024 study presented at AIES, researchers audited three LLM-based resume-screening models across 554 resumes and nine job categories, varying only the names. The models preferred resumes with white-associated names in 85.1 percent of statistical tests, and preferred names associated with men in 51.9 percent of tests versus 11.1 percent for women. Identical qualifications; different names; different scores.
Then in 2025, the same group ran the experiment that should end the "human-in-the-loop solves it" argument. In a 528-participant study, reviewers picking between equally qualified candidates chose evenly when they had no AI input, or neutral AI input. When the AI's recommendations were severely biased, reviewers followed them roughly 90 percent of the time. As lead author Kyra Wilson put it: "Unless bias is obvious, people were perfectly willing to accept the AI's biases."
Two honest hedges before you quote that number at a dinner party. The study used online participants in a simulated task, not professional recruiters reviewing real pipelines. And the widely cited companion statistic, that around 80 percent of organizations using AI hiring tools say they never reject an applicant without human review, comes from a ResumeBuilder survey rather than an audit. But the direction of the finding is hard to argue with, and it matches what anyone who has stared at a ranked list already suspects: a bare score does not invite scrutiny. It invites agreement.
Why "a human reviews everything" is not enough
Put the two UW findings together and the standard industry reassurance falls apart. The models can be biased, and the humans reviewing them tend to mirror whatever the model says. A rubber stamp is still a rubber stamp when a person is holding it.
The problem is structural. When a screening tool outputs "87% match" and nothing else, the reviewer has nothing to engage with. There is no claim to verify, no reasoning to challenge, no place where their expertise gets traction. The path of least resistance is to nod. And under a pile of 400 applicants, everyone takes the path of least resistance.
What trustworthy AI screening actually looks like
The fix is not removing the AI, and it is not adding a second sign-off. It is changing what the AI hands to the human. Three properties matter:
1. Criteria you wrote, not taste the model learned. Amazon's model went wrong because it learned what "good" meant from historical hires. A trustworthy scorer never gets to define "good." Your team writes the must-haves and nice-to-haves for the role; the model's only job is to check resumes against that list. If the criteria are biased, they are at least visible, written in plain language, and fixable by editing a document instead of retraining a model.
2. Reasons, not verdicts. A score should come with its work shown: which criteria matched, which are missing, what the evidence was. The breakdown converts the reviewer's job from "do I trust this number?" into "is this specific claim about this specific resume true?" That is a question a human can actually answer, and disagreeing with one line of a breakdown is a much smaller act than overruling a confident-looking percentage.
3. A place where human judgment gets recorded. If a reviewer disagrees with the machine, that disagreement should become part of the candidate's record, visible to the whole team, not a private mental note that evaporates. Judgment that is written down compounds; judgment that is not gets re-derived by the next reviewer, or lost.
How Reordinal implements this (and where the line is)
Reordinal scores every applicant against the job's criteria, the ones you wrote into the job description. Every score ships with its breakdown: matched criteria, missing criteria, strengths, and concerns, per candidate. Reviewers sort by score to decide where to start reading, open the breakdown and the parsed resume to check the machine's claims, and record their verdicts as team comments on the candidate.
The line we hold: the score orders the pile, people make the call. There is no auto-reject in Reordinal. A low-scored candidate is still in the list, still parsed, still one click from a human's eyes, which also means the long tail of applicants gets a floor of attention that a tired human skimming page 14 was never going to provide.
Held to the standard above, that is the honest pitch: not "our AI is unbiased," which nobody can promise, but "our AI shows its work, on your criteria, and never decides." That is the version of AI screening the research says you can actually supervise.
The takeaway
Distrust of black-box resume screening is not technophobia; it is the correct reading of the evidence. But the answer is not going back to gut-feel skimming, which has its own well-documented biases and no audit trail at all. The answer is AI that argues its case instead of announcing a verdict, in front of humans who keep the decision. If you are evaluating any screening tool, including ours, ask one question first: when it scores a candidate, can I see why, and can I disagree on the record?