Does Artificial Intelligence Decide for Us in Recruitment?

AI has infiltrated recruitment like water slipping under a door: discreetly, inexorably. From the automated screening of CVs to voice analysis during video interviews, from pre-screening chatbots to predictive matching algorithms, every stage of the candidate journey can now be mediated, accelerated, or simply decided by a machine. As early as 2018, 64% of recruiters surveyed in a large online study reported using these tools. Since then, their deployment has only accelerated. But what really happens in practice when a recruiter delegates part of their judgment to an algorithmic system? Who wins, who loses? And at what cost?

By: Giseline Rondeaux

Imagine this. You send in your CV. It is read, or rather scanned, in a matter of seconds by an algorithm. Before a human even lays eyes on it, an artificial intelligence has already decided whether you deserve to move forward. This scenario is not science fiction: it has quietly become the everyday reality of millions of candidates in Europe and elsewhere.

AI Recruits. Then What?

The promises are appealing: greater objectivity, time savings, reduced discriminatory biases, better matching between job offers and job seekers...

The signals are clear: between 2023 and 2024, the proportion of employers reporting the use of AI in recruitment tripled according to the iHire platform, rising from 5% to nearly 15% in just one year. This acceleration is fueling the enthusiasm of HR directors.

Yet reality is rougher. Researchers specializing in artificial intelligence sometimes speak of artificial incompetence to describe systems that process information on a massive scale but struggle to grasp the uniqueness of a human career path, the nuance of a career change, or the richness of an atypical profile. The data feeding these algorithms does not fall from the sky: it was produced by humans, along with their biases, blind spots, and institutionalized blind spots.

This is precisely what the ERUSIA research project seeks to document rigorously and empirically in real companies, with real recruiters, real employment agencies, and real candidates.

What We Still Don’t Know, and Why It Matters

Here is a striking paradox: at a time when recruitment AI is being presented as a certified revolution, scientific research into its actual uses remains surprisingly incomplete. Most existing studies focus on the acceptance of these technologies, in other words on what users think about AI, far more than on what they actually do with it, hour by hour, case by case. Many studies list AI’s potential without relying on empirical feedback analyzing real practices and uses, which are often less disruptive.

Yet there is often a huge gap between the technology as prescribed and the technology as actually used. What does a recruiter choose to do when they receive an algorithmic recommendation they consider absurd? Do they follow it blindly because “AI knows better”? Do they quietly work around it, at the risk of wasting time? Or do they report it to management, at the risk of appearing old-fashioned? The organizational practices we have analyzed reveal that humans remain paramount in the final decision. These daily, microscopic arbitrations are, however, the real life of AI-augmented recruitment. Recruiters therefore choose, depending on the case, not to follow the tool or the prescribed process, relying instead on their expertise or even their intuition. It is these adjustments that ultimately determine whether AI improves or degrades the quality of work and the quality of matching in the labor market.

Over two years, through three complementary approaches, the research project aims to understand: a large quantitative survey involving various organizations in Hauts-de-France and Belgium; several qualitative case studies based on in-depth interviews with HR professionals, labor market intermediaries, and candidates; and finally, an unprecedented analysis from the perspective of the designers of these systems (R&D teams and HR AI start-ups) to understand how AI systems are built and where the risks of bias arise, sometimes without anyone truly realizing it, from the very first lines of code.

The Blind Spot: The Designers

This third component of the study may be the most original and the most urgent. We often talk about AI bias as a problem of data quality or quantity. That is true, but incomplete. Biases can also result from design choices and implementation strategies that are co-constructed through dialogue with the client: which criteria should be retained? Which variables should be considered relevant? How much weight should be given to a CV without continuous professional experience, to a career path marked by back-and-forth moves, or to skills acquired outside conventional pathways?

These choices are rarely neutral. They reflect representations of the “ideal candidate” that are themselves historically and socially constructed. And once encoded into an algorithm, these representations acquire an appearance of objectivity that makes them all the more difficult to challenge.

The European AI Regulation (AI Act, which entered into force in 2024) classified AI systems used in recruitment among high-risk systems, that is, systems likely to affect individuals’ fundamental rights. A welcomed recognition. But between regulatory obligations and the actual practices of companies, which are not always aware of or informed about these risks, the gap can be considerable. Current proposals (November 2025), such as the Digital Omnibus Regulation Proposal, are nevertheless emerging and carry the risk of weakening various components of the AI Act (delayed key obligations for high-risk systems, exemptions for existing systems, lighter requirements for certain sectors, and relaxed documentation requirements) in the name of business competitiveness.

What Candidates Experience Without Anyone Really Asking Them

There is one final voice often absent from the debate: that of the candidates themselves. How do they feel when they learn that their application was filtered by an algorithm before reaching a human desk? Do they perceive this process as fair? As more or less discriminatory than human screening? What impact does it have on their relationship with job seeking and on their trust in institutions?

These questions are not anecdotal. The labor market is also a market of trust. If candidates lose faith in the fairness of selection processes, or if they develop strategies to “game” the algorithm, such as artificially formatting their CVs to pass automated filters or discreetly using ChatGPT during remote interviews, the entire system becomes more fragile. Moreover, it appears that their expectations are evolving toward a faster process.

Neither Techno-Optimism Nor Technophobia

The goal is neither to condemn AI in recruitment nor to celebrate it. The goal is to understand it in its real-world uses, in its measurable effects, and in its blind spots that are carefully ignored by marketing narratives that are often disconnected from reality.

What research can provide, and what technology solution vendors will never tell you, is nuance. AI can be a powerful lever for efficiency and discrimination. It can free up time for recruiters and erode their expertise. It can reduce certain human biases and amplify others, more subtle and imperceptible, behind the apparent neutrality of code.

The question is therefore not “Should AI be used in recruitment?” but rather: Who decides how it is used? In service of what? With what guarantees for workers, candidates, and the labor market as a whole?

These are political questions as much as technical ones. Questions that deserve to be asked loudly and clearly, and on the basis of solid evidence.

This article was written with the support of Dr. Arnaud Stiepen, expert in science communication