AI in recruitment: between efficiency and bias

AI in recruitment: between efficiency and bias

Have you ever thought that an algorithm, unlike a human, can't be biased? A recent study from Princeton and the University of Chicago, reported by MIT Technology Review , shows just the opposite — AI's bias in recruiting may be even greater than that of a human recruiter. And that's increasingly important, as more and more companies let an algorithm be the first to read their candidates' resumes.

The bad news isn't just that AI "inherits" stereotypes from the data it's trained on — something that's been known for some time. The really worrying news is that it can invent its own stereotypes on the fly, even when it has no real reason to do so.

Princeton–Chicago study on AI bias in recruitment

The researchers tested several large language models, including ChatGPT, Claude, and Gemini, in a job simulation game adapted from a classic psychology study on stereotype formation. Each model was put in the role of a consultant hired by the mayor of a fictional city, tasked with filling 20 different positions—from doctors and lawyers to babysitters and cleaners.

The candidates came from four completely invented ethnic groups (Tufa, Aima, Reku and Weki) specifically created so that the models could not "cheat" using stereotypes already existing in the training data. At each of the 40 rounds, the model chose a candidate, found out whether or not they had succeeded in the job, and moved on. Importantly: in reality, all candidates had exactly the same chances of success, regardless of the group they belonged to.

The result? The models quickly began to separate candidates by ethnicity, based on just a few early results. 📊 For example, if a single Aima candidate failed as a doctor (a position that theoretically requires a lot of competence and warmth), the model avoided hiring other Aima doctors and redirected them to cleaning jobs, which were considered “easier.” On a scale of 0 to 2, the humans in the original study scored 0.84 — the AI models scored about 65% higher on average, and OpenAI’s o3 model scored 1.83, close to the maximum possible. The most worrying aspect? The newer models, with more advanced reasoning capabilities, showed the strongest bias tendencies — not the weakest.

Why algorithms stereotype faster than humans

The researchers' explanation is, in short, that models are trained to find patterns quickly from a small number of examples — the very skill that makes them good at math, coding, or logic problems. But that same skill becomes a problem in a social context: a model that is "rushed" to generalize from a single failure ends up drawing hasty conclusions about an entire group of people, instead of waiting for more data before making a decision.

What doesn't work is interesting. Researchers have tried simply asking models to be "fair" and "non-discriminatory," but the behavior hardly changed. In contrast, when models were given an explicit incentive for diversity in hiring, bias decreased significantly. Another practical detail for any HR professional: when models were given information about candidates that was truly relevant to the job—like age or education—they became less biased, because they had something to base their decision on besides ethnicity. But when the only additional information was irrelevant to the job—like hair color—the models reverted to their old habits and chose based on ethnicity, not anything else.

The risk increases as chatbots gain more advanced memory and personalization capabilities — a model that "remembers" previous interactions with a type of candidate risks repeating and amplifying those exact patterns, not just forming them once. And this is relevant right now, because more and more companies are using AI not only for CV filtering, but also for automated interviews.

How to prevent algorithmic bias in your recruitment process

The practical conclusion is not "don't use AI at all in recruitment", but neither is it "let the AI decide completely objectively, without supervision". Both are myths. The fact that AI works with data and calculations, not impressions, does not automatically make it correct — just like a human, it forms its "opinion" based on the examples it has seen so far, and if those examples are few or inappropriate, its conclusions can be just as wrong. The difference between a useful tool and an algorithm that amplifies discrimination lies, most of the time, in how much human control remains in the process. ⚖️

I've already written about how to avoid bias in the selection process , and the principles there (structured interviews, clear evaluation criteria, numerical scoring) are even more valid now that algorithms also have their own bias to manage. Basically, the same rules that protect candidates from a tired or rushed recruiter also protect them from an AI model that generalizes too quickly.

Some concrete things you can do:

  • Keep the final decision in the hands of humans. The algorithm can recommend, but it should not decide on its own (the approach known in the industry as "human-in-the-loop");
  • Standardize evaluation criteria before introducing any automated tool into the process. An AI without clear criteria has even more freedom to form its own patterns;
  • Do a little audit from time to time : check if the results filtered by the algorithm show reasonable diversity, not just higher speed;
  • Constantly track results : not just the speed of recruitment, but also the diversity of candidates who reach the final stages.

At Busy Bee 🐝, the recruitment process remains yours, not a black box algorithm. Our platform centralizes applicants in one place so that the final decision stays where it belongs: with you and your team.

Fairer recruitment, with or without AI

No one denies that AI brings real benefits to recruiting — speed, consistency, less time wasted on repetitive tasks. But fast doesn’t mean automatic and correct. Without human oversight and clear criteria, an algorithm can end up repeating, and even amplifying, the very patterns we’ve been trying to eliminate from the selection process for years. The good news is that the solutions are largely the same as for human bias — structure, transparency, and constant monitoring. ✅

If you recognize the chaos of CVs filtered half by an algorithm and half "by eye", you know there is a better way. Busy Bee 🐝 centralizes all your applicants in one place while always maintaining control over the final decision.

Try Busy Bee now !

Are you ready to buzz peacefully?

Create a branded careers site, publish roles, and start collecting applications—effortless and quick.

Get Started for Free