AI’s quiet shift in selection
Something is shifting in the way we select. That change around AI in selection is just not happening in a way that is immediately apparent. For a long time, selection was primarily a human process, based on conversations, experience and intuition. We relied on our ability to assess someone well, to recognize potential and interpret behavior. Sometimes that worked. But just as often judgment turned out to depend on context, on the person sitting across from you, or simply on the moment. That discomfort never completely disappeared, but we accepted it for a long time.
Why AI in selection felt like progress
When technology made it possible to examine behavior more systematically, and especially when AI began recognizing patterns in data, something that felt like progress emerged. Suddenly, we could make decisions that were more consistent, came about faster, and, in many cases, also better predicted who would be successful. For many organizations, this made AI in selection not only interesting but also a logical step. If something demonstrably works better than what you were used to, there is little reason not to use it.
Predicting is not yet understanding
But right there is also an uncomfortable question, one that has long been on the back burner. For what does it actually mean that something works? When a model finds patterns associated with performance, we can conclude that it has predictive value. That is relevant and, in many cases, practically useful. But that does not make it clear exactly what is being measured, or why that correlation exists. The difference seems subtle, but in reality, it is fundamental. After all, a system can predict just fine without understanding what it is doing. And as long as the outcome is better than the alternative, that may not feel like a problem.
The AI Act changes the rules of the game for selection
Until you are asked to explain why you made a certain decision. And that moment is getting closer and closer. Not only because organizations are becoming more critical, but also because the context is changing. With the introduction of the European AI Act, which came into force in 2024 and whose main obligations take effect in 2026, the standard for selection is shifting. For AI in selection, this means a clear shift: systems used in recruitment and selection often fall into the high-risk application category. That does not mean they can no longer be used, but it does mean that there will be higher requirements for how they are deployed and accounted for.
It is not only the outcome that counts
This is not just about whether a system predicts well, but how it arrives at that prediction and how carefully that process is designed. Organizations must be able to demonstrate that they understand their systems, manage risk, and maintain human control. That sounds logical, but it becomes complex when you work with models optimized primarily for outcomes rather than explainability.
The tension between smart and explainable
That’s where a tension arises that you don’t solve directly by simply using better technology. Many of the tools developed in recent years are strong precisely because they detect patterns that are difficult for humans to see. But that very fact makes it more difficult to reconstruct exactly what is happening and why a particular score came about. Especially with AI in the selection process, it becomes clear how dependent you are on your ability to make sense of the process. As long as everything goes well, this does not seem to be a problem. But as soon as questions are asked – by a candidate, a manager or a supervisor – that changes immediately.
Measuring itself also changes
This tension is amplified by the fact that the measurement context itself is also changing. Candidates are beginning, tentatively but noticeably, to use AI when completing assessments. This does not mean that all outcomes are suddenly unreliable, but it does mean that answers are less obviously a direct reflection of behavior. They also partly reflect how well someone understands and knows how to use the system. At the same time, the work itself is changing rapidly. It is becoming more international, more digital and increasingly dependent on collaboration and interaction with technology. As a result, behavior is less easily separated from the context in which it arises.
A higher bar for AI in selection
When you add these developments together, it becomes clear that the bar is shifting not so much because technology is failing, but because our expectations are changing. It is no longer enough to be able to say that a system predicts better than what you used before. With AI in selection, the question increasingly becomes whether you understand what you are measuring, whether you can explain how you arrived at a decision, and whether you are prepared to justify that decision.
The core of selection is exposed
For HR and selection, this represents a shift that goes deeper than choosing a new tool. It goes to the heart of what selection is: making decisions about people. Decisions that must not only be effective, but also explainable, consistent and defensible. And perhaps that is the most important change that the AI Act makes visible. Not that we are going to measure differently, but that we can no longer suffice with measurement alone.
The question that remains
In the end, you are left with a question that sounds simpler than it is. Not whether you can predict who will be successful, but whether you understand what you are measuring and whether you want to take responsibility for the resulting decision. In the end, that is the real test for using AI in selection.
