The Myth of the Objective Recruiter: Why AI Hires More Fairly Than People – If It's Built Right

What unconscious bias really does in hiring, what candidates think about it, what AI can structurally do better – and why the EU AI Act is not an obstacle but the decisive quality filter. With figures from our study of 1,044 job seekers.

We all believe we assess applications rationally. We read the CV, check the qualifications, look for fit – and then decide whether someone is invited to an interview or not.

The research says otherwise.

In a large-scale field experiment in Germany, published by the IZA Institute of Labor Economics, an applicant with a Turkish-sounding name and a headscarf had to send out 4.5 times as many applications on average as an identically qualified applicant with a German name in order to receive the same number of interview invitations. Identical profile. Same education. Same work experience. Different treatment.

Findings like this have been replicated for decades across Europe and beyond. Despite equal treatment legislation, despite awareness training, despite a declared commitment to diversity and equal opportunity. The conclusion is not "recruiters are biased". It is: people are biased, and recruiters are people.

And candidates already know it

Research shows that bias exists. Our own study shows that candidates feel it, too. For "Applying as a Slot Machine", the market research institute Bilendi surveyed 1,044 job seekers on our behalf in July 2026 – all of whom had applied for a job within the previous twelve months. Instead of agreement scales, we asked them to choose between sharply contrasting statements.

The result: 57% assume that hiring decisions are shaped by prejudice – regarding age, gender or origin, for example. 61% say a recruiter's personal sympathy has a very strong influence on who gets through. Among the hiring managers involved, 64% believe this. And almost one in two (46%) experience their own application as a game of chance: often arbitrary and hard to follow.

This contradicts the self-image of many HR departments, which describe their selection as structured, objective and criteria-based. Whether that perception is accurate in any individual case is secondary. Those who believe gut feeling decides in the end apply half-heartedly – or not at all.

This is exactly where a question comes in that is gaining importance in the discussion around AI in recruiting and the EU AI Act: what does a "non-discriminatory hiring process" actually mean – and what can AI contribute to it?

What bias in recruiting actually means

"Bias" is one of the overused words of our time. In recruiting, it means something specific: systematic distortions in how candidates are evaluated that have nothing to do with their professional suitability – and that may constitute discrimination under equal treatment law. The main forms have been documented for decades:

• ‍Halo effect: One striking positive trait – attractiveness, eloquence, a well-known company on the CV – outshines the perception of all other traits. Studies from the 1970s and 1990s (Schuler & Berger; Marlowe, Schneider & Nelson) consistently show that attractive candidates are invited to interviews more often despite identical qualifications. Recruiters are practically never aware of this.‍
• Affinity bias (or "similar-to-me"): Someone from the same city, the same university, with similar hobbies – comes across as more likeable and therefore, unconsciously, more competent. This is the psychological basis behind the much-invoked "culture fit", which sometimes becomes a polite cover for affinity bias.‍
• Name bias: The IZA field experiment is just one example of many. The name on an application systematically changes the probability of being invited – even when the recruiter is convinced they only look at qualifications.‍
• Confirmation bias: Once an initial hypothesis is formed, the brain interprets everything that follows as confirmation. Contradicting evidence is filtered out or reinterpreted.

These mechanisms are not character flaws of individual recruiters. They are properties of the human brain – heuristics we use to process complex information quickly. They activate stereotypes without us noticing.

Notably, those who suspect these mechanisms most are the ones with the most experience: among job seekers over 60, 64% assume prejudice in selection, compared with 55% among 18- to 29-year-olds. University graduates are more sceptical (60%) than people without a degree (53%). The more hiring processes someone has been through, the less they believe in objective assessment.

Why training and anonymised applications alone are not enough

The obvious reflex is: "Then we need to train our recruiters better." Well-intentioned, but empirically a weak strategy.

The reason: bias is largely unconscious. The Implicit Association Test (IAT), used in thousands of variants since 1998, reliably shows that people unconsciously link stereotypes to certain groups – even when they sincerely believe they discriminate against no one. Training that explains the halo effect changes remarkably little. Knowing about a heuristic does not switch it off.

Anonymised applications – leaving out photo, name and date of birth in the first screening round – are a good first step, but not a complete one. As soon as the personal interview begins, every relevant characteristic is back in the room and the heuristics kick in. Anonymised applications reduce bias in pre-selection but do not replace consistent evaluation across the entire process.

Then there is the volume problem. A recruiter reviewing 50 applications a week has less than two minutes per CV on average. At 200 applications, it's seconds. Under that kind of time pressure, the brain falls back almost automatically on heuristics – exactly the shortcuts that produce discrimination in the first place. In other words: the larger the application volume, the stronger the tendency towards unconscious bias. Precisely in the settings where objective assessment matters most, it is hardest to achieve.

Candidates feel that time pressure directly: 59% find it hard to reach employers during the process when questions come up. Where there is no time for questions, there is no time for methodical evaluation either.

What AI can structurally do better

Here is the point that is often overlooked – including in the public debate on AI in recruiting, which likes to focus on risks: a properly built AI has structural properties that can systematically reduce human bias. Not because it is "more neutral" than people, but because it is consistent.

• ‍Consistency: An AI evaluates every application against the same criteria, in the same order, with the same attention. It does not tire, is not influenced by the previous application, does not have a bad day. Application number 1 and application number 200 get the same treatment – something that demonstrably does not hold in human-led processes.‍
• Structured evaluation: Recruiting research has long shown that structured interviews and standardised scoring grids almost double the predictive power for actual job performance compared with unstructured conversations – and massively reduce bias effects. An AI runs structured processes by nature: that is its default mode, not a training recommendation someone might remember.‍
• Auditability: A human decision is hard to review after the fact. Why wasn't this candidate invited? "Gut feeling" is an honest answer in practice, but a risky one legally. An AI, by contrast, can log every step: criteria applied, score calculated, recommendation given. That is what makes patterns of discrimination visible in the first place – and therefore correctable.‍
• Scalability of fairness: Reducing bias is not only a question of will but of capacity. With ten applications, a recruiter might force themselves to work methodically. With two hundred, that gets hard. A well-built AI maintains its methodology at any volume.‍
• Acceptance on the candidate side: A common counter-argument is that candidates don't want AI in the process. The data says otherwise. 54% of job seekers would leave pre-selection to an AI if it decides on verifiable criteria whether they get an interview with their future manager – the classic review of documents by recruiters comes in behind. And 32% name a process free of human prejudice as one of the three most important features of an ideal hiring process. Those who suspect gut feeling in the process trust transparent criteria more than the person at the desk.

The important caveat: AI is not automatically fair

At this point, honesty is required – this is not a statement about all recruiting AIs, but about what AI can fundamentally do when it is deliberately built for it.

The counter-story is well known: from 2014, Amazon developed an internal AI tool for sorting CVs and had to shut it down because it systematically rated female applicants for tech positions lower. The reason was banal: the model was trained on historical hiring data dominated by men – and learned that "male" correlates with "successfully hired". Exactly what the AI was supposed to prevent, it replicated.

The lesson is not that AI is unsuitable. The lesson is that design is decisive. An AI trained unfiltered on historical data replicates the distortions in that data. An AI trained on curated, representative data, whose evaluation criteria are job-relevant and verifiable, and whose live results are continuously monitored for patterns of discrimination, can do the opposite: it makes evaluations more objective, not more subjective.

Candidates draw a clear line here as well. Given the choice, 56% opt for a classic process with human decisions throughout, even if it takes four to ten weeks. 44% would prefer a largely AI-driven process that reaches a decision in 24 to 48 hours – provided a human decides at the end. So AI is welcome at the pre-selection stage. For now, people should decide. Exactly this division of labour is the architecture the following section is about.

This is exactly what the EU AI Act demands

This brings in a reading of the EU AI Act that usually gets lost in compliance discussions. The AI Act is not primarily a hurdle for AI in recruiting – it is a quality filter for the right kind of AI in recruiting.

A note on timing: with the Digital Omnibus Regulation adopted in summer 2026, the EU postponed the application of the core high-risk requirements for Annex III systems – which include recruiting software – to 2 December 2027. That changes nothing about the requirements themselves. Anyone choosing a system today that will need retrofitting in 2027 is buying a problem with advance notice.

What the AI Act requires of high-risk systems – and recruiting is one of them – are precisely the properties that separate a bias-replicating system from a bias-reducing one:

• ‍Data quality and data governance: Training and test data must be curated so that distortions are detectable and controllable. This is exactly the measure Amazon did not take.‍
• Bias monitoring in live operation: Not only the training data but also ongoing results are monitored for systematic deviations. If certain groups are treated differently, it becomes visible – before it becomes a scandal.‍
• Explainability: Every recommendation must be traceably justified. No "match: 73%", but understandable criteria. That makes bias something to discuss rather than guess at.‍
• Human oversight: AI prepares, humans decide. This separation not only protects candidates but also prevents AI bias from flowing uncorrected into a final decision.‍
• Documentation: Every decision is logged. Bias analysis becomes a continuous possibility rather than a one-off audit.

Put differently: anyone looking for a recruiting AI that hires more fairly than people is automatically looking for a recruiting AI that meets the AI Act. The two go together. The AI Act sorts out the tools that replicate discrimination – and leaves those that actively support equal opportunity.

What this means for HR leaders

Anyone who wants to hire more fairly in 2026 than a few years ago has better levers than ever before. Three points make the difference in practice.

1. The honest acknowledgement that human evaluation alone is not a solution. "We pay attention to fairness" is not enough – the data shows that good intentions barely influence the brain's heuristics. Structured processes and consistent criteria are the precondition for a genuinely non-discriminatory hiring process.
2. An AI built for recruiting – not a generalist tool with an HR module bolted on, not a homegrown solution. A specialised recruiting AI brings the decisive properties out of the box: consistent evaluation across all applications, continuous bias monitoring in live operation, traceable justification for every recommendation. Generalist tools inherit the distortions of their training data. Homegrown solutions have to build this architecture first – an effort that is rarely realistic alongside day-to-day business.
3. The clarity that people decide – and can now do so better than before. In a purely human pre-selection, there is often nothing tangible left at the end about why someone was rejected. A specialised recruiting AI turns this around: the recruiter sees every recommendation with its justification and can accept, adjust or reject it. What they gain is not less control – it is informed control, on a data basis that can later be explained and documented.

Our customer VD Mayr, a Munich-based security services company with around 600 hires a year, shows what this looks like in practice. Around half of the applicants for sports events have an international background. In the live interview, Paul checks whether German is available at B2 level – and nothing else. Origin and accent play no role. Self-assessed language skills often differ considerably from actual competence; the structured check makes them comparable, with the same criteria for every application.

One last thought

The most important insight from decades of bias research is not that recruiters are unfair. It is that fairness in recruiting is a question of architecture, not of character. Structured processes, consistent criteria, documented decisions, continuous monitoring – these are the tools with which discrimination can be systematically reduced.

That architecture is exactly what a well-built recruiting AI delivers. And it is exactly what the EU AI Act demands of high-risk systems. Both pull in the same direction – and both work towards a goal that is in every company's interest anyway: better, fairer and verifiably traceable hiring decisions that carry diversity as more than a slogan.

At Paul's job, we have put exactly this logic at the centre of our product: recruiting AI that treats the AI Act not as an afterthought but as a design principle – and candidates as the benchmark. What they want, we have compiled in "Applying as a Slot Machine". The full study is available free of charge.

And if you want to see what fair, AI-Act-compliant recruiting actually feels like: get in touch, we'll show you.

‍