No Hiring Tool Removes Bias. Including Ours.

Where does bias in hiring begin?

Most conversations about AI bias in recruitment focus on the model. But hiring outcomes are shaped long before any model runs, and long after.

Think about everything that decides who gets hired. How the job is defined, and which requirements are written down. Where it is advertised, and who ever sees it. Which candidates apply in the first place. Which questions are asked, and what counts as a good answer. How recruiters and hiring managers interpret the evidence, what happens when a candidate does not fit the template, and the organisation's own habits and history.

Technology can sit anywhere in that chain, but it is only one part of it. A badly defined requirement produces unfair outcomes whether a person or a machine applies it. And a process that relies heavily on people is not automatically fair either.

What can hiring technology change about bias?

Technology is not irrelevant to bias. It changes how a process operates, and some of those changes are useful.

It can ask every candidate for a role the same job-related questions. It forces requirements to be written down, because criteria have to exist before they can be configured. It records answers against those requirements instead of leaving them in one recruiter's memory. It makes it possible to see later what a decision was based on, and who made it. And it can take repetitive processing off recruiters' desks, which may leave more time for careful review.

These are process mechanisms. They make some kinds of inconsistency less likely and others easier to see. But consistency is not the same as fairness.

A process can be highly consistent and still produce problematic outcomes. If a requirement is not really needed for the job, applying it the same way to everyone applies the same unnecessary barrier to everyone. If a question is harder for some groups to answer for reasons unrelated to the job, asking it consistently does not make it fair.

The difference is simple. Consistency is a mechanism: you can design it in. Fairness is an outcome: you have to measure it.

What the research does, and does not, tell us

The best evidence on discrimination in Dutch hiring comes from field experiments, where researchers send matched fictitious applications to real vacancies.

The study in the opening is one of them. Thijssen, Coenders and Lancee (2021) ran it between October 2016 and April 2018, with applicants aged 23 to 25 from 35 ethnic minority groups. The difference was larger for non-Western than for Western minority groups, and adding grades, performance details, social skills or a photo to the applications did not reduce it.

That finding matters, and it is also narrow. It measured callbacks at the application stage, with fictitious applications, and it left out the lowest segments of the labour market. So it does not directly describe frontline logistics hiring.

The authors call for more attention to interventions such as formalising hiring processes, anonymous applications and stronger anti-discrimination policy, while noting that research on the effects of such interventions is scarce.

Research from the United States does find smaller average score differences between groups for structured interviews than for cognitive ability tests (Sackett and colleagues, 2022). That is not evidence that structure reduces discrimination in hiring, and no study reviewed for The Radius Papers 01: A New Standard for Frontline Hiring shows that structured or AI-assisted screening does.

What should you ask a hiring tool about bias?

"Does your tool remove bias?" invites a yes, and the yes means very little.

Better questions are harder to fake. What exactly is the system evaluating, and against which criteria? Who wrote those criteria, and could you defend each one as job-related? What evidence does it produce for each candidate, and can a recruiter inspect and challenge it? Does it rank or reject candidates, or only support evaluation? And are outcomes, including differences between groups where that is lawful, actually monitored over time? The full list is in our guide to choosing AI recruitment software.

Some things should simply be ruled out. Since 2 February 2025, the EU AI Act has banned AI that infers emotions at work. The ban covers inferring emotions from biometric data such as a face or voice, with exceptions for medical or safety reasons. The European Commission's guidelines, which are not binding, say it also covers job candidates (C(2025) 5052 final).

What Radius Hire does, and where it can fail

Radius Hire asks every applicant the same structured questions, built on the requirements the employer sets, and shows the recruiter the evidence behind each requirement, labelled Proven, Claimed or Missing. It does not rank candidates or reject anyone on its own. A recruiter decides.

Candidates answer by voice, in their own language, and Radius Hire does not assess accent, appearance or background. These choices remove some known sources of inconsistency and make the evidence visible. They do not prove that Radius Hire is unbiased, and we do not claim that it is. We have no published evidence that it produces fairer outcomes than any other method, including a careful human process.

Our product also carries risks of its own. Speech recognition makes more errors for some accents and dialects than for others (Koenecke and colleagues, 2020), so a transcript can miss part of a real answer. A question asked in several languages has to mean the same thing in each, and a translation can drift. Either can make an answer look weaker than it was. That is why the recruiter sees the answer behind every label, not only the label, and why we treat a transcript as evidence to check, not a verdict.

What does responsible AI hiring look like?

Whatever tools you use, the practice is less glamorous than the marketing. Start from job-related requirements, written in the words of the people who supervise the work, and ask about them consistently. Make the evaluation criteria visible to everyone who applies them. Collect evidence rather than relying only on claims. Keep a named person responsible for every decision, with the power to challenge and overrule any output.

Then do the part most teams skip: watch the outcomes. Look at who progresses, who is hired and who stays, and review the process when the evidence shows a problem. Comparing outcomes between groups can involve special categories of personal data, such as ethnic origin, which the GDPR protects strictly. The AI Act allows such data, under strict conditions, to detect and correct bias in AI systems. Whether that covers your own outcome monitoring is a question for counsel.

Be open with candidates, too: the GDPR requires you to tell them how their data is used. What the AI Act already requires, and from when, is set out in our guide to the EU AI Act and recruitment. And treat all of this as ongoing work, not a checkbox at purchase.

Closing thought

The honest promise of AI hiring is not "we removed bias".

In our view, it is closer to this: we designed the process so that important decisions rest on clearer evidence, with people responsible for the decision, and outcomes monitored over time.

That is a smaller promise. It is also one you can check.