COLUMN: When a Robot Judges You in a Job Interview — Welcome to the Future

Efficiency as a Key Selling Point

Advocates of AI in recruitment put forward arguments that, on paper, seem solid. The first is efficiency. In large companies that receive thousands of applications for each open position, it is humanly impossible to give each application the attention it deserves. Human recruiters spend an average of six to seven seconds on a resume before deciding whether to reject or keep it. Six seconds. In this context, claiming that the traditional process is inherently fairer than automation is somewhat naive. AI, according to its advocates, would standardize the process, evaluating each candidate according to the same criteria, at the same time, under the same conditions.

The second argument concerns the reduction of unconscious biases. Research in social psychology has extensively documented the fact that human recruiters unconsciously favor candidates who resemble them—same background, same accent, same school, same neighborhood. AI, according to its proponents, is blind to these confounding factors and judges candidates solely on their actual skills. It’s a compelling argument. And like all compelling arguments, it deserves to be subjected to rigorous critical scrutiny—something its proponents carefully avoid doing in their sales pitches.

The Promise of Algorithmic Meritocracy

There is something utopian about the vision these companies are selling. A world of recruitment where your ZIP code no longer matters, where your foreign-sounding last name no longer penalizes you, where your gender, physical appearance, and age fade into the background behind a purely objective assessment of your abilities. A world where pure meritocracy would finally be possible because an impartial machine would have made it so. It’s beautiful. It’s inspiring. And it is, to a large extent, a myth carefully perpetuated by an industry that generates billions of dollars by selling this promise to overwhelmed HR departments eager to believe that there is a technological solution to a fundamentally human problem.

Algorithmic meritocracy is version 2.0 of an old fantasy: that we might one day measure a human being’s worth with enough precision to automate judgment. This fantasy has never been harmless.

The Toxic Legacy of Training Data

The central problem with AI-powered recruitment systems is what experts call training data bias. To learn how to distinguish a good candidate from a bad one, an algorithm needs examples. These examples come from past hiring decisions—that is, decisions made by biased humans, in historically homogeneous companies, reflecting structurally unequal societies. Amazon sadly illustrated this problem in 2018, when the company had to abandon a recruitment AI tool it had been developing for several years after discovering that it systematically penalized women. The system had learned that the best engineers were men—because the vast majority of engineers hired by Amazon over the previous decades were, in fact, men. It was reproducing and amplifying the very discrimination it was supposed to eliminate.

This Amazon case is not an anomaly. It is a particularly well-documented illustration of a structural problem. Researchers at the University of Washington have shown that certain facial recognition systems are significantly less accurate for dark-skinned faces than for light-skinned faces—in part because the training data for these systems overrepresented white people. In a hiring context, this means that AI could interpret the facial expressions of a Black candidate and a White candidate differently, even when they are expressing exactly the same emotion. This is not intentional discrimination. It’s worse: it’s invisible structural discrimination, buried in thousands of lines of code that no one can audit.

Facial Analysis Under the Experts’ Microscope

The analysis of facial expressions in recruitment is one of the most controversial applications of behavioral AI. It is based on the pioneering work of psychologist Paul Ekman, who proposed in the 1960s that certain basic emotions—joy, sadness, anger, fear, disgust, and surprise—were universal and recognizable through facial expressions common to all cultures. However, decades of subsequent research have seriously qualified—or even challenged—these claims. A meta-analysis published in 2019 in the journal Psychological Science in the Public Interest, authored by Lisa Feldman Barrett and her colleagues, concluded that the idea that facial expressions accurately reflect internal emotional states was scientifically unsupported. In other words: one of the theoretical foundations upon which facial analysis in recruitment is based is, according to a large portion of the scientific community, fragile—or even erroneous.

We are therefore entrusting life-changing decisions to technologies built on contested scientific foundations. And we call this progress.

How Those Who Have Been Through It Feel

To understand what it actually means to go through an interview with a machine, you have to listen to those who’ve been through it. The accounts gathered by numerous researchers and journalists all point to a common experience: a deep sense of unease, absurdity, and sometimes humiliation. Candidates describe the feeling of talking into thin air, of smiling at a screen that doesn’t smile back, of having to convince someone—or something—that can neither hear them nor truly understand them. Some speak of a new kind of anxiety: not the fear of failing in front of a human being who might at least understand that they’re nervous, but the fear of triggering a negative signal in an algorithm for reasons that are incomprehensible and beyond their control. How do you adjust your behavior to please a judge whose criteria you don’t know? How do you remain authentic in front of a camera that analyzes your microexpressions?

Studies conducted in several European countries show that candidates from minority groups—people of color, people with disabilities, neurodivergent individuals, and people for whom French or English is not a native language—report significantly higher levels of anxiety when facing automated interviews. People on the autism spectrum, in particular, point out that the behavioral criteria on which these systems rely—sustained eye contact, fluency in nonverbal communication, and emotionally calibrated responses—correspond exactly to the areas in which their neurodiversity sets them apart from the norm. A system that measures employability against the standard of neurotypical conformity is a system that, de facto, discriminates against neuroatypical people—who are often brilliant, often highly qualified, and often unable to “perform” according to the behavioral codes that the algorithm has been trained to value.

The Silence of Automated Rejections

There is a particular cruelty in automated rejections that advocates of these technologies seem reluctant to acknowledge. When a human recruiter rejects you, there is at least the possibility—even if it isn’t always seized—of feedback, an explanation, or a learning opportunity. “You lack experience in team management” is feedback you can process, challenge, or incorporate. When an algorithm rejects you, there is nothing. A generic email. Sometimes not even that. Just silence. And an unanswered question: What did the machine see in me that I don’t see? What did I say, do, or express that triggered the red flag? This lack of transparency isn’t a minor detail. It’s a silent form of violence inflicted on people who are often already vulnerable due to the job search—an experience that carries enough stress and uncertainty on its own without adding the unfathomable mystery of an algorithmic judgment.

We have collectively decided that efficiency is worth this cost. I’m not sure we’ve truly weighed what that cost means for those who pay it.

The Opacity of Black Boxes

One of the most concerning issues with AI in recruitment is algorithmic opacity. Companies that sell these systems jealously guard their proprietary models. The specific criteria the algorithm uses to evaluate candidates are generally not disclosed—neither to the candidates nor even to the client companies. These companies purchase a tool whose internal workings they cannot audit. They receive scores, recommendations, and rankings—but not the mechanisms that generate them. This is what computer scientists call a “black box”: a system whose outputs are observed without access to its internal processes. In this context, how can a company ensure that the tool it uses complies with anti-discrimination laws? How can it defend its hiring decisions in court if it itself is unaware of the criteria that produced those decisions?

The answer, in most cases, is that it cannot. And that is precisely why several regulators are beginning to take this issue seriously. In 2021, New York City passed a law—which took effect in 2023—requiring employers using AI recruitment tools to undergo annual equity audits conducted by independent third parties. The European Union, through its now-adopted AI Act, classifies AI systems used in recruitment as high-risk systems, subject to enhanced requirements for transparency, traceability, and auditability. These regulatory advances are welcome. They are also, for now, lagging far behind the pace at which these technologies are being deployed.

Conflicts of Interest in the Ecosystem

It is instructive to examine who is funding the research that validates these tools. Studies demonstrating the effectiveness and fairness of AI recruitment systems are, to a troubling extent, funded or conducted by the very companies that sell these systems. HireVue publishes studies on the effectiveness of HireVue. Pymetrics publishes studies on the fairness of Pymetrics. This isn’t illegal. But it is a conflict of interest that HR departments at client companies would do well to keep in mind when reviewing the available literature. Independent research on these tools is significantly more nuanced—and often much harsher—than what sales brochures suggest.

The HR AI industry sells certainty in a field that should demand humility. And it does so with enough conviction that smart, well-intentioned business leaders trust it without double-checking.

Europe Is Ahead, but Not Fast Enough

Globally, the European Union is the jurisdiction that has made the most progress on the regulatory framework for AI in recruitment. The European AI Act, which officially took effect in 2024, represents a historic step forward. It classifies AI systems according to their risk level and imposes substantial obligations on high-risk systems—including recruitment tools—such as detailed technical documentation, decision logging, mandatory human oversight, robustness and accuracy testing, and the right of affected individuals to clear explanations. On paper, it is an ambitious and coherent framework. In practice, its implementation remains a work in progress—and tech companies, with their teams of lobbyists and specialized lawyers, generally know how to navigate regulatory gray areas.

The GDPR (General Data Protection Regulation), in effect since 2018, also offers certain relevant protections for candidates subject to algorithmic evaluations. Article 22 of the GDPR stipulates that individuals have the right not to be subject to decisions based solely on automated processing when such decisions produce significant legal effects or similarly affect them. In theory, this should protect candidates from purely algorithmic rejections. In practice, companies generally circumvent this provision by maintaining the appearance of human oversight—a human who, nominally, “validates” the algorithm’s decisions—without this validation being truly substantive.

Toward a Right to an Explanation in Hiring

Many legal experts and digital rights advocates are calling for the creation of a right to an explanation specific to algorithmic recruitment. Such a right would allow any candidate rejected by an automated system to obtain a concrete and intelligible explanation of the reasons for their rejection—not a generic email, but a detailed explanation that enables them to understand, challenge if necessary, and learn. This right exists in part under the GDPR, but its application in the context of recruitment remains unclear and is rarely enforced. Countries such as Belgium and France have begun to incorporate considerations on this subject into their labor laws, but there is still a long way to go.

A candidate rejected by a human can ask why. A candidate rejected by an algorithm—who, exactly, can they ask? The question is rhetorical. And the silence that greets it speaks volumes about the state of our labor democracy.

Estonia: A Pioneer by Nature and by Necessity

To understand why the Baltic countries are particularly receptive to AI technologies in recruitment, one must understand their unique relationship with digitization. Estonia, in particular, has made digital transformation a national priority since the 1990s. After regaining its independence in 1991, the country deliberately chose to build its public administration around digital technology—partly because it lacked the human and financial resources to maintain a cumbersome traditional bureaucracy, and partly because its leaders at the time understood that digital technology represented a competitive advantage for a small nation of 1.3 million people seeking its place in the global economy. Today, 99% of Estonian public services are accessible online. The electronic signature has legal standing. The e-residency program allows entrepreneurs from around the world to establish an Estonian company without ever setting foot in Tallinn.

Within this cultural and institutional context, adopting AI tools in recruitment is seen not as a radical break with tradition, but as a natural step in an already well-advanced process of digitization. Estonian startups—and the country has a remarkable number per capita, with unicorns such as Transferwise (Wise), Bolt, and Pipedrive—have been among the first in Europe to integrate automated recruitment tools into their processes. This isn’t out of recklessness. It stems from familiarity with technology and pragmatism in the face of a tight labor market, where unemployment is at a historic low and the war for talent is particularly fierce.

Specific Risks in Small Economies

But this favorable context hides specific risks. In small economies like those of the Baltic states, the labor market is tight. There are few companies in certain sectors. Professional networks are tight-knit, and circles are often closed. In this context, a biased recruitment algorithm can have disproportionate effects. If a system systematically favors certain profiles—graduates of specific universities, native speakers of a certain language, people with a specific socioeconomic background—in a small economy, the exclusionary effect is amplified. There aren’t ten similar companies to apply to—there are only two or three. If the algorithm bars you from these few opportunities, the consequences for your career path can be severe and long-lasting.

In small economies, algorithmic errors are not abstract statistics. They are professional paths that veer off course, careers that come to a halt, and potential that is extinguished—invisibly, silently, and effectively.

When Recruiters Deliberately Choose the Human Touch

Faced with the wave of recruitment automation, some companies are making a deliberate and public choice to resist it. Not out of nostalgia, not out of a rejection of progress, but out of a strategic conviction that the quality of hiring depends on authentic human interaction. Recruitment firms specializing in certain sectors—notably the creative industries, innovation sectors, and, paradoxically, some tech companies themselves—maintain recruitment processes where the initial interactions are human by design. Basecamp, the American software company, has publicly taken a stand against algorithmic assessment tools. Several design and communications agencies in Paris, London, and Amsterdam highlight their human-centered recruitment processes as a key aspect of their employer brand.

These companies teach us something important: the automation of recruitment is not inevitable. It is a choice. A choice that may make sense in certain contexts—massive volumes of applications, highly standardized roles, the need for rapid scalability—but one that isn’t universally necessary. And above all, a choice that should be made with full awareness of its implications, not simply because the technology is available and the sales rep from the HR startup gave a great presentation to the executive committee.

Successful hybrids: AI + humans = the winning formula

The most successful experiences integrating AI into recruitment aren’t those that replace humans, but those that augment them. Hybrid models are beginning to emerge, where AI handles low-value-added tasks—initial resume screening, verifying minimum criteria, scheduling interviews—while leaving high-value-added interactions to human recruiters: in-depth interviews, cultural fit assessments, and conversations about aspirations and motivations. In these models, AI is a decision-support tool, not an autonomous decision-maker. Human oversight is real, substantial, and not merely formal. It’s a slower, more costly, and more demanding approach. It’s also a fairer approach—and, in the long run, likely more effective at identifying the candidates who will truly make a difference.

The real question isn’t “AI or humans.” It’s “how to ensure that technology serves humans rather than replacing them”—and this question requires leaders to make courageous decisions rather than comfortable ones.

The Scenario of Drift: When Machines Decide Everything

If current trends continue unchecked, the most likely scenario by 2030 is one of increasing and widespread automation in recruitment. AI tools will become more sophisticated, more accurate, and less expensive—and thus accessible to an ever-wider range of companies, including smaller ones. Human recruiters will become a rare sight in certain sectors, confined to the highest levels of the hierarchy or the most strategic positions. The majority of initial interactions between candidates and employers will be handled by automated systems. In this scenario, inequalities in access to employment risk becoming further entrenched around the ability to perform well for algorithms—a new, abstract skill that is fundamentally different from the ability to do a job. The person who knows how to speak to a camera to maximize their facial trust score will get the job. The person who is simply the best for the job but doesn’t understand the ins and outs of automated recruitment will fall through the cracks.

This scenario is not inevitable. But it is the path of least resistance—the one we drift toward when no one bothers to actively question the direction we’re heading. And the warning signs are already visible: coaches offering training to “optimize your behavior in front of AI cameras,” tools that let you analyze your own video interview before submitting it to identify potentially negative cues, and entire industries springing up to help candidates navigate a system designed to filter them out.

The Disruption Scenario: When Regulation Regains the Upper Hand

An alternative scenario exists, and it is just as plausible. It is one in which regulation catches up with innovation, where the rights of workers and candidates are strengthened by clear legislation and robust enforcement mechanisms, where mandatory fairness audits expose and compel the correction of algorithmic biases, and where a culture of transparency regarding AI in recruitment becomes the industry standard. In this scenario, AI becomes a legitimate and reliable tool because it has been held to the same standards as any other tool used to make decisions that affect people’s rights. This scenario requires political will, civic vigilance, and companies that are responsible enough to anticipate regulatory requirements rather than merely comply with them.

Both scenarios are realistic. Which one comes to pass depends largely on the choices we make collectively over the next five years—and on our ability to ask the right questions while there is still time to change course.

Understanding the Rules of the New Game

While we await more robust regulations, candidates facing automated recruitment processes aren’t entirely at a loss—as long as they understand the rules of this new game. The first step is to actively research the tools used by the company you’re applying to. More and more companies are mentioning in their job postings that they use automated assessment systems. If they don’t, it’s perfectly legitimate—and increasingly recommended by HR experts—to ask the recruiter directly: “Does your process include AI-based assessments? If so, based on what criteria?” The question itself signals your level of professionalism. And the answer—or lack thereof—will tell you a lot about the organization’s culture of transparency.

The second step is to take the time to prepare specifically for asynchronous video interviews. This format is challenging in its own way: there’s limited time to think, no interaction to adjust your response based on the interviewer’s cues, and a constant awareness of the camera that can unsettle even the most qualified candidates. Practicing with simulation tools, recording your own responses to review later, and working on your nonverbal communication in a video interview setting—all of this can make a real difference. Not to manipulate the system, but to avoid being penalized by a technical misstep in a format that’s still unfamiliar.

Collective Defense

Individual advocacy has its limits. When facing structural systems, the responses must be structural. Digital rights organizations such as AlgorithmWatch in Germany and the AI Now Institute in the United States are documenting and exposing abuses by AI-powered recruitment systems. Unions are beginning to include the issue of AI in recruitment on their collective bargaining agendas. Lawyers specializing in labor and digital law are developing expertise to support candidates who believe they have been discriminated against by an algorithm. These resources exist. Using them, supporting them, and amplifying them—this, too, is a form of active resistance against a dispossession that would otherwise occur in silence and invisibility.

We are not powerless. But powerlessness is the feeling these systems seek to instill—because a candidate who does not understand what is happening to them cannot challenge it, and a candidate who does not challenge it costs nothing to correct.

What are we really looking for in the hiring process?

Behind all the technical and regulatory debates about AI in recruitment lies a deeper, more fundamental question that our societies consistently avoid: What makes for good hiring? What are we really looking for when we choose someone for a position? If the answer is simply “the profile that best matches a list of measurable criteria,” then AI is indeed the ideal tool—it measures better and faster than any human. But if the answer is more complex—if it includes potential for growth, adaptability, interpersonal skills, the ability to challenge established assumptions, to bring a different perspective, and to create something the job description didn’t anticipate—then today’s algorithmic tools are profoundly inadequate. Because they excel at identifying compliance with past standards, yet are blind to what transcends them.

The world’s most innovative companies did not become so by hiring candidates who best fit established profiles. They did so by hiring people who didn’t fit into the boxes—and who, precisely for that reason, were able to create new ones. Steve Jobs probably wouldn’t have passed a recruitment algorithm designed for a major tech company in 1985. Frida Kahlo wouldn’t have gotten a job at a design agency that uses behavioral analysis to measure emotional stability. Alan Turing—whose algorithms are, in a sense, the forerunners of the AI that evaluates candidates today—would likely not have passed facial analysis. History is full of brilliant and unconventional people whom conformity criteria would have weeded out. And recruitment algorithms are, at their core, machines of conformity.

The Irreducible Value of Human Encounter

There is something in the job interview—in its human and imperfect form—that AI cannot replicate: the encounter. The moment when two people face each other and ask one another: Can we build something together? This question cannot be reduced to skill scores, analyses of microexpressions, or measures of verbal fluency. It involves a form of mutual intuition, a bet on the future, trust granted before it is earned. It involves humanity—in the most literal sense of the word. And if we abandon this humanity in the name of algorithmic efficiency, we won’t just be changing our recruitment processes. We’ll be changing what we think work is, and what we think human beings are worth.

The real question posed by AI in recruitment is not a technical one. It is a question of civilization: do we decide that the value of a human being can be calculated? And if so, to whom do we entrust the task of making that calculation?

What This Moment Reveals About Us

The rise of automated job interviews isn’t just a technological trend to observe with curiosity or concern. It’s a barometer—a mirror held up to our societies that shows us what we truly value, what we’re willing to sacrifice in the name of efficiency, and what we believe human beings owe one another in the workplace. This mirror isn’t flattering. It shows companies outsourcing their ethical responsibility to algorithms they don’t understand. It shows job candidates reduced to data sets to be analyzed. It shows regulators scrambling to keep up with innovation that is outpacing them. And it shows, in a way, a collective abdication of our responsibility to set limits on what we allow machines to decide for us.

But this mirror also shows something else: a resistance that is taking shape. Researchers documenting biases. Regulators enacting laws. Lawyers making their cases. Job candidates asking questions. Companies choosing a different path. Voices—like that of the Baltic Times, which brought this issue to the forefront—that refuse to let this transformation take place amid indifference and invisibility. It’s not enough. But it’s a start. And in times of technological upheaval, beginnings matter.

The Urgency of a Deliberate Choice

We are at a crossroads. The decisions we make today—as companies, as regulators, as candidates, as citizens—will determine the shape of the labor market for the next generation. We can choose the path of algorithmic convenience and accept all the consequences that entails in terms of equity, dignity, and diversity. Or we can choose the more demanding path of responsible innovation—a path that harnesses the power of AI where it is truly useful, while preserving the irreplaceable value of human connection where it is indispensable. This choice is not a technical one. It is political, ethical, and deeply human. And it is up to us—entirely and fully—to make it.

A robot can analyze your microexpressions. It cannot look you in the eye and decide that it believes in you. We cannot afford to let that difference disappear.

Signed, Jacques Pj Provost