When the Algorithm Decides: What Recruiters Really Need to Know

Artificial intelligence has become a standard part of hiring. More than 90% of employers in the United States now use algorithms to screen job applicants, and the trend is growing across Europe too. The promise is appealing: faster decisions, less human inconsistency, more objective outcomes.

But a major new study from Stanford University, Northeastern University, and Chapman University is asking us to look more carefully at what is actually happening inside these systems. You can read the full research here.

What the research found

The study analysed 4 million real job applications submitted to 156 employers across 11 sectors. All of them were screened by algorithms from a single vendor. What the researchers found should give every HR professional pause.

First, there is clear evidence of racial disparity. Over 25% of applications submitted by Black candidates and nearly 15% of those submitted by Asian candidates were directed to positions where the algorithm demonstrably disadvantaged them. This is not a theoretical risk. It is happening at scale, in real hiring decisions, right now.

Second, the research identified what the authors call "systemic rejection." When multiple employers use algorithms from the same vendor, candidates who are filtered out in one place tend to be filtered out everywhere. 

An applicant would need to submit around 25 applications just to have a near-certain chance of receiving a single recommendation from the algorithm. Under genuinely independent human decisions, that number would be closer to 10.

Third, and perhaps most striking, these disparities were largely invisible in aggregate data. It was only by looking at individual positions separately that the adverse impact became clear. Blended statistics can hide a great deal.

What does this mean for recruiters?

The research is not an argument against using technology in recruitment. It is an argument for using it with far more awareness and scrutiny than most organisations currently apply.

A few things are worth considering:

Algorithms are not neutral

Every system is trained on historical data, and historical data reflects historical decisions, including the biased ones. If past hiring favoured certain profiles, an algorithm trained on that data will tend to reproduce those same patterns. Efficiency and fairness are not the same thing.

Over-reliance on a single vendor creates fragile systems

When the majority of employers use the same tools, the consequences of a flawed model are amplified across the entire market. Candidates who do not fit a particular algorithm's preferences may find doors closing everywhere, not just in one place.

Human judgment still matters

A combination of structured evidence, trained human awareness, and genuine curiosity about a candidate is more robust than any single system operating alone.

What this means in practice

For HR teams and hiring managers, this research is a prompt to ask harder questions of the tools being used. Who built this algorithm, and on what data? Has it been audited for adverse impact at a granular level? Does the screening process leave adequate room for human review?

It is also a reminder that diversity, equity, and inclusion commitments cannot be outsourced to a piece of software. Technology can support a fair process, but it cannot guarantee one.

Really, the future of recruitment is not about choosing between human and machine. But about building processes where both are used thoughtfully, with clear accountability for outcomes.

At EvolveTalent, this is something we think about every day. If you would like to talk about how to build a recruitment process that is both efficient and genuinely human-centred, we would be glad to hear from you! Write to us at hello@evolvetalent.eu