Yes, AI hiring software can choose better employees than unaided human judgment, but only when it is built on sound predictors, checked against real job outcomes, and watched for bias. Formulas have beaten gut feel for decades, while new screening tools built on large language models still treat equal candidates unequally. Fairness is never automatic.
Two beliefs dominate the conversation about automated selection, and they point in opposite directions. The first is that software is neutral: a machine does not care about skin color or gender, so handing screening to an algorithm must make hiring fairer. The second is that people read candidates better than code can, so an experienced recruiter should always have the final word. Both beliefs feel sensible. Neither survives close contact with the evidence.
An algorithm learns from the decisions and outcomes it is fed, and those decisions were made by people with all their blind spots, so neutrality has to be engineered and tested rather than assumed. At the same time, decades of research on how people combine information about candidates show that human judgment adds inconsistency more often than insight. The honest answer sits between the two camps. A well built formula usually predicts better than a person, a badly built one can lock in old unfairness at great speed, and the newest tools built on large language models have introduced fresh and sometimes strange biases of their own. What follows sets out what the research shows, study by study, and what it means for anyone deciding whether to trust a machine with a hiring decision.
What Do Most People Get Wrong About AI Hiring Software?
The common story about automated screening comes in two versions. In the hopeful version, the software reads every application with perfect attention, ignores irrelevant details and surfaces the strong candidates that tired recruiters would have missed. In the fearful version, the software is a black box that quietly rejects women and minorities while nobody can explain why. Both versions treat the algorithm as a single thing with a fixed character, good or bad.
That framing hides the questions that matter. An algorithm used in selection is a bundle of choices made by people: which outcome to predict, which past data to learn from, which information about candidates to feed in, and what to do when results differ across groups. Change those choices and the same technology can become more accurate and fairer, or less accurate and less fair. The right comparison is also usually missing. Critics compare the algorithm with an imaginary unbiased human, while sellers compare it with an imaginary careless one. The proper benchmark is the actual process the software would replace, with its actual error rate and its actual bias.
There is also a gap between what research has tested and what organizations now use. Most of the strong evidence on formulas versus judgment comes from structured information such as test scores and interview ratings, combined by simple weighting rules. Much of the software on the market today does something different. It reads unstructured text in resumes and applications, often with large language models, which are computer programs trained on enormous amounts of text to predict and generate language. The older evidence tells us a great deal about the value of consistent decision rules. It tells us much less about whether a particular text reading model measures anything related to job performance. Keeping those two bodies of evidence apart is the first step to thinking clearly about the subject.
Do Formulas Really Beat Human Judgment When Choosing Employees?
The idea that simple statistical rules usually predict human outcomes at least as well as trained professionals goes back to the psychologist Paul Meehl in the middle of the last century, and it has been tested many times since. A large meta analysis published in 2000 pooled studies from many fields of human prediction that compared clinical prediction, where an expert weighs the information in their head, with mechanical prediction, where a fixed rule or formula does the weighing. A meta analysis is a study that combines the results of many earlier studies to find the overall pattern. The formula was clearly better in roughly a third to just under half of the comparisons, the expert was clearly better in only 6% to 16%, and on average mechanical prediction was about 10% more accurate. The experts did relatively worse when their predictions drew on interview information.
Selection research produced an even sharper result. A 2013 meta analysis gathered 25 samples from 17 studies in which the same information about applicants was combined in two ways: by a formula, or intuitively by people who knew the jobs and organizations well. For predicting job performance, the formula did more than 50% better than intuitive combination. The authors estimated that, in a moderately selective hiring situation, relying on intuitive combination would cut the number of correct hiring decisions by about a quarter. They added that their figures were probably underestimates, because they could not correct for some known statistical distortions. Their practical advice was modest and sensible: let experts gather job relevant information, let a consistent rule combine it, and if people must adjust the result, make them record why.
Why would knowledgeable people do worse than a formula using the same facts? The short answer is inconsistency. A person reading the same file on a Monday and on a Friday will not weigh it the same way. People notice vivid details, give extra weight to whatever they saw last and build stories that feel convincing but do not predict. A formula applies the same weights every time. Even a crude rule applied consistently tends to beat a sophisticated judge who applies standards unevenly.
None of this means every input to a formula is equally good. A 2022 reanalysis of the main meta analyses on selection procedures concluded that earlier work had overcorrected for a statistical problem called range restriction, which arises because researchers can only study people who were hired rather than the full applicant pool. Once that overcorrection was removed, the estimated predictive power of most selection procedures fell, in some cases by a wide margin. Structured interviews, in which every candidate is asked the same job related questions and answers are scored against a set guide, came out at the top. Tests of general cognitive ability, long treated as the strongest single predictor, moved further down the table. The same reanalysis set predictive power alongside the size of average score differences between Black and White candidates, showing that some methods cost more in diversity for each unit of accuracy they add. For anyone weighing software, the lesson is plain. A formula can only be as good as the information it combines. Automating a weak predictor gives you a fast weak predictor.
What Happens When Managers Overrule the Test?
Comparisons in research datasets are one thing. What happens inside real companies when managers can choose whether to follow a test? One study of hiring across 15 firms employing low skilled service workers examined what happened after the firms introduced a job test that sorted applicants into bands by predicted fit. Managers were free to hire someone from a lower band ahead of someone from a higher band, and many did. Hires made in line with the test stayed in the job longer. Managers who overruled the test more often ended up with workers who left sooner, and the higher scoring applicants they passed over, when hired later, tended to stay longer than the people chosen ahead of them. The researchers concluded that managers mostly overruled the test because they were biased or mistaken, not because they held valuable private knowledge about the candidates.
The study is careful about its limits. Tenure, which simply means how long someone stays in the job, is a narrow measure of success, and in some settings managers may know things a test cannot capture. Still, the direction matches the meta analytic evidence: discretion added noise more often than wisdom.
A deeper problem haunts every attempt to compare algorithms with people. You only ever see how hired candidates perform. You never learn how the rejected ones would have done. Researchers call this the selective labels problem. A study of judges deciding bail in the United States, which concerns courts rather than hiring but follows the same logic, showed one way around it by exploiting the fact that cases were assigned to judges more or less at random. The researchers built a machine learning model, a computer program that finds patterns in past data and uses them to make predictions. In simulation, the model could have cut crime by roughly a quarter without jailing more people, or reduced the jail population by about 42% with no rise in crime, while also lowering jailing rates for Black and Hispanic defendants. Judges, the authors found, appeared to treat noise as if it were a meaningful signal. The hiring lesson is direct. Any claim that software picks better people needs a believable way to estimate how rejected candidates would have performed. A seller who only reports how hired employees did has not answered the question.
Can Machine Learning Find Better Candidates Than Recruiters?
The studies above mostly concern simple formulas built on test scores. Newer work asks whether machine learning trained on past hiring data can beat experienced recruiters at the earliest stage, the resume screen.
A field experiment at a technology company tested exactly this for software engineering jobs. A field experiment is a study run in a real workplace in which some decisions are randomly assigned to one approach and some to another, so researchers can see cause and effect. The company trained an algorithm on four years of past resumes and screening decisions and then allowed it, at random, to override human screeners on who received an interview. Candidates picked by the machine who would have been rejected by people were about 14% more likely to pass the interviews and receive a job offer, and about 18% more likely to accept the offer. Much of the gain came from candidates the human screeners usually passed over: women, members of racial minorities, people without an employee referral, graduates of less prestigious colleges and people with no prior work experience.
The same paper offers a theory of when this works. An algorithm learns from past human decisions. If the people who made those decisions were biased but also inconsistent, their inconsistency acts like a natural experiment: some strong candidates from overlooked groups slipped through, and the algorithm can learn from how they performed. If the people were biased and highly consistent, the algorithm has nothing to learn from and will simply copy the bias. This is a working paper, meaning it has circulated among researchers without yet completing peer review in a journal, and it comes from a single company, so the size of the gains should not be assumed to hold elsewhere.
One teacher hiring study in a large public school system took a different route. Instead of letting a model hunt for whatever patterns predicted success, the researchers used machine learning to turn the free text in roughly 16,000 applications, such as descriptions of past jobs and reasons for leaving, into measures a psychologist would recognize: how relevant the applicant's experience was, how long they had stayed in earlier roles, and whether their moves looked like steps toward better jobs or escapes from bad ones. Relevant experience and a history of moving toward better jobs predicted stronger performance ratings, better contributions to student learning and lower turnover, while a pattern of escaping bad jobs predicted the opposite. The model added accuracy beyond the way the district normally judged work history and reduced the risk of adverse impact, which is the term for a selection step that screens out one group, for example women or a racial minority, at a much higher rate than another. The lesson is that machine learning works best when it measures things with a clear link to the job, rather than searching blindly for any pattern in the data.
The most useful recent contribution concerns a flaw built into most screening algorithms. A study of recruiting at a large professional services firm, covering 88,666 applications between 2016 and 2019, pointed out that standard algorithms only exploit. They pick the kind of candidates who succeeded before, which means they keep choosing people who resemble past hires and never learn whether anyone else could do well. A better approach balances exploitation with exploration, giving some extra weight to candidates from groups the firm knows little about, precisely because their potential is uncertain. The authors built such an exploring algorithm and compared it with the firm's recruiters and with a standard algorithm. Among the candidates recruiters chose to interview, about 10% were Black or Hispanic, and an estimated 10% went on to be hired. The standard algorithm raised the estimated hiring rate of the people it selected to about 32% but cut the Black and Hispanic share to about 4%. The exploring algorithm kept most of the gain in quality, with an estimated hiring rate of about 27%, while raising the Black and Hispanic share to about 24%. The study first circulated as a working paper in 2020 and has since been listed for publication in a leading economics journal. Its message is that a hiring algorithm is not doomed to trade fairness for accuracy. Much depends on whether it is designed to keep learning.
Is an Algorithm Fairer Than the People It Replaces?
Any honest answer has to start with the human baseline. A meta analysis of field experiments combined 28 studies in the United States, covering 55,842 applications for 26,326 positions, in which researchers sent matched fictional applications that differed only in signals of race. Since 1989, White applicants received on average 36% more callbacks than equally qualified African American applicants and 24% more than Latino applicants. Over 25 years, the level of discrimination against African Americans did not change. That is the process most screening software replaces. Against this benchmark, an algorithm does not have to be perfect to be an improvement. It does have to be checked, because it learns from the very decisions that produced those gaps.
Fairness also has more than one meaning, and the meanings conflict. A theoretical computer science paper from 2017 proved that three common definitions of a fair risk score cannot all hold at once, except in rare special cases. In plain terms, when two groups differ in how often the predicted outcome actually occurs, a score cannot be equally trustworthy for both groups and also make the same kinds of mistakes at the same rates for both groups. Every algorithm therefore carries a choice about which kind of fairness comes first. That choice is ethical and organizational, not merely technical, and employers should know which one was made on their behalf.
A 2020 review of seller practices examined what companies selling screening algorithms disclosed publicly about how their tools were built and checked for bias. The review found little public information about validation, and it showed that decisions about which outcome to predict and which data to collect shape bias in ways a simple check of final results will not reveal. If a tool is trained to predict who past managers rated highly, it inherits whatever bias shaped those ratings, however carefully the final scores are adjusted. The authors also found that some popular approaches to removing bias sit uneasily with the technical and legal ideas they are meant to serve.
Fixes are possible. A 2023 study by researchers at a company that sells assessment software tested two methods on a field sample of job candidates. One removed predictors that both widened gaps between groups and weakened prediction. The other built a penalty for group differences directly into the model while it was learning, pushing it toward predictions that were accurate and more even across groups at the same time. Both methods reduced group differences while keeping predictive accuracy, and the penalty approach did better. Because the authors worked for a seller, independent replication would strengthen the case, but the study shows that the tradeoff between diversity and accuracy is partly a design decision rather than a law of nature.
Algorithms can also change the gaps that human evaluators create. Two field experiments on recruiting for a web developer role, reported in a working paper, found that human evaluators scoring applicants without help rated women lower than men, but when the same kind of evaluators were shown an algorithm's scores, the gender gap disappeared. Women were also more likely to complete their applications when told an algorithm would score them, and the final applicant pool was 29% female under algorithmic scoring against 22% under human scoring, a difference that could still reflect chance. The authors estimated that the algorithm would need to be strongly biased against women before it produced less diversity than human evaluation alone. These are single studies in one sector, yet they point the same way as the resume screening experiment above: a disciplined algorithm can narrow gaps that people create.
What Happens When Large Language Models Read Resumes?
Most of the evidence so far concerns models built for one purpose and trained on one company's data. The fastest growing category of screening tools works differently. It uses general purpose large language models, trained on vast amounts of text from the internet, to read resumes and compare them with job descriptions. These models absorb the patterns of the text they learned from, stereotypes included. Researchers test them with a version of the classic audit study: take identical resumes, change only a detail that signals race or gender, and see whether the model's judgment moves.
One 2024 audit tested a family of text embedding models. An embedding model turns a piece of text into a long list of numbers so a computer can measure how similar two texts are, and many screening systems use this technique to match resumes with job descriptions. Using more than 500 resumes and 500 job descriptions across nine occupations, the researchers found the models favored resumes signaling a White candidate in 85.1% of tests and favored resumes signaling a female candidate in only 11.1%. Resumes signaling a Black man were disadvantaged in up to 100% of comparisons, echoing patterns found in real labor markets.
Other studies find different patterns. One preprint study of teacher applications in a large public school district, which has not yet been peer reviewed, found moderate race and gender disparities in how language models rated applicants, and those disparities held across different kinds of application materials and different ways of instructing the models. A 2025 study had five widely used models score about 361,000 fictitious resumes for entry level jobs. The models tended to give women higher scores than equally qualified men, and to give Black men lower scores than equally qualified White men. At realistic cutoff scores, those differences shifted the probability of being hired by roughly 1 to 3 percentage points.
The bias does not always run in the expected direction, and it is not always about identity. A 2026 study asked 22 large language models to choose between pairs of matched resumes across 70 professions, producing about 30,690 decisions. Every model preferred the resume signaling a female candidate, choosing it 56.9% of the time. More troubling, the models chose whichever candidate was listed first 63.5% of the time, and 21 of the 22 models showed this preference for the first position. The order in which applications happen to be presented has nothing to do with merit, yet it moved the result further than gender did.
A 2026 preprint, also not yet peer reviewed, audited 14 widely used models with more than 24,000 paired job postings each. The single model from 2023 reproduced the familiar pattern of favoring White applicants, by about 2 percentage points. Every model released from 2024 onward showed either no racial gap or a gap favoring Black applicants of up to about 3 percentage points, and a similar pattern appeared for gender. The authors read this as evidence that developers have corrected for one kind of bias and, in some cases, overshot.
Taken together, these studies do not say that language models are always biased against one group. They say something more awkward for employers. The direction and size of bias vary by model, by version, by occupation and even by the order of the files. A tool that treated groups evenly last year may not do so after the provider updates the underlying model. And none of these studies measured whether the models' judgments predict job performance, which is the first question any selection method should answer. That gap in the evidence is large, and it is the gap most relevant to anyone using these tools to make real decisions.
Does Keeping a Human in the Loop Solve the Problem?
The usual reassurance is that software only recommends and a person decides. A 2025 experiment with 528 participants tested that reassurance directly. Participants screened candidates for 16 occupations, some of high status and some of low status, across 1,526 scenarios, sometimes alone and sometimes with recommendations from a simulated language model that had been set up to favor one racial group. Without recommendations, or with neutral ones, people chose candidates of each race at similar rates. When the recommendations were biased, people followed them up to 90% of the time. Even participants who said they considered the recommendations low quality or unimportant were swayed in some conditions. One small step helped. Taking an implicit association test, a short computer task that shows how quickly a person links social groups with particular ideas, before screening raised the share of choices that ran against common stereotypes by up to 13%.
This is one experiment with simulated recommendations rather than a live tool, and the participants were not professional recruiters. Even so, it undercuts the idea that human review is an automatic safety net. A reviewer who sees a ranked list tends to accept the ranking. If the ranking is biased, the human mostly adds a signature to it. Oversight works only when reviewers have the time, the information and the authority to disagree, and when someone checks whether they actually do.
What Should You Expect From AI Hiring Software?
If you are deciding whether to let software screen or rank your candidates, the research gives you a clear starting point and a set of warnings. You have good reason to expect that a consistent, well designed decision rule will predict job performance better than your managers' unaided judgment. The evidence on this is old, large and consistent. You also have reason to expect that the process you run today is less fair than it feels, because field experiments keep finding large gaps in how equally qualified people are treated.
What you should not expect is that any particular tool delivers those benefits on its own. The gains in the studies came from systems that predicted a clear, job relevant outcome, learned from data with some variety in past decisions, were tested against what rejected candidates might have done and were checked for group differences. A tool that reads resumes with a general purpose language model has, so far, no published evidence that its rankings predict performance, and good evidence that its rankings shift with arbitrary details such as file order and model version.
Your own role changes as well. If you keep a person in the loop, assume that person will tend to follow the machine, and design the review so disagreement is easy and recorded. If you rely on an algorithm that only rewards people who look like your past hires, expect your workforce to become more uniform over time, with nobody able to tell what talent you missed. And if nobody can tell you what outcome the model predicts, how well it predicts that outcome and how its results differ across groups in your own applicant pool, you do not yet know whether you have a better screen or a faster version of an old bias.
Key Takeaways
- Consistent decision rules have outperformed unaided expert judgment in selection for decades, largely because people apply their own standards unevenly from one candidate to the next.
- Automation does not make a weak predictor strong. The value of any screening model depends on whether the information it combines is genuinely linked to performance in the job.
- A model trained only on past hiring decisions tends to repeat whatever bias those decisions contained, unless it is designed to explore candidates who differ from previous hires.
- Human screening sets a low bar for fairness, since matched application studies show equally qualified minority applicants still receive far fewer callbacks.
- Resume screening with large language models produces biases that change direction across models and versions, and the models can be swayed by something as arbitrary as which file comes first.
- People reviewing algorithmic recommendations usually go along with them, so human oversight protects candidates only when it is built to make disagreement easy and visible.
Implications for Practice
Start with the outcome, not the tool. Before any screening model is switched on, write down what it is supposed to predict, whether that is performance ratings, tenure, output or completion of training, and check that the outcome itself is not contaminated by past bias. The review of seller practices showed that this single choice shapes fairness more than any later adjustment, so treat it as a decision for senior leaders and selection specialists, not a technical setting.
Combine information about candidates with a consistent rule and limit free overrides. The meta analytic evidence favors letting people gather information and letting a formula combine it. Where managers may depart from the formula, require a short written reason, and every year compare how the overridden hires performed with how the recommended ones did. The study of 15 service firms suggests that many overrides will turn out to be costly, and tracking them is the only way to know whether yours are.
Test any model against what happens to people it would reject. A simple design is to run a pilot in which, for a random share of vacancies, interview invitations follow the model rather than the usual process, and then compare the success of the two groups. This mirrors the resume screening field experiment and answers the selective labels problem in a way that a report on current employees never can. If a model selects mostly people who resemble past hires, reserve a small share of interview places for promising candidates it is uncertain about, so the organization keeps learning instead of narrowing.
Monitor selection rates for each group at every stage, from application to offer, and compare them with the rates your human process produced before. Decide in advance which definition of fairness you will hold the system to, because the mathematical proof on competing definitions shows you cannot satisfy them all. Where group gaps appear, ask whether the model can be rebuilt with a penalty for those gaps or with fewer of the predictors driving them, as the 2023 study did.
Treat screening based on large language models as unproven for selection decisions until there is local evidence that its scores predict performance. If you use such tools at all, remove identity signals from resumes before scoring, shuffle the order in which files are presented, rerun a matched resume check after every change in the underlying model, and never let the score be the sole reason to reject a candidate. Finally, design human review so that reviewers form their own view before they see the machine's ranking, and audit how often they disagree. If the answer is almost never, you do not have oversight, you have a rubber stamp.
Related Reading on The Human Capital Hub
- Why behavioral interviews still matter when screening is automated
- Building a recruitment pipeline that grows with your business
- How automated tools are changing the job search for candidates
