Paul Sackett has a candidate for the worst test he has ever seen: the US citizenship test.
For years, it worked like this. There were 100 open-ended questions, published in advance. The examiner asked them orally and could choose which questions to ask, and how many. Some were easy. Nearly everyone knows the colours of the US flag. Others were much harder, like the year the Constitution was ratified. Some questions were answered correctly by fewer than 10% of people.
So an examiner who took a liking to someone could steer them towards easy questions. An examiner who didn't could do the opposite. That's bias and noise in one process.
The fix: a computer now randomly selects ten questions for each person. The examiner can no longer steer. But ten questions is too few. By chance, one person gets the flag and another gets the Federalist Papers. Better, still noisy.
Sackett is a professor of psychology at the University of Minnesota. He has spent 47 years developing and studying selection systems, and his research is the reference point for what predicts job performance. At Meeting of the Minds 2026 in Stockholm, he used the citizenship test to set up the one idea he came to hammer home.
"The key issue is consistency. Consistency of treatment of all candidates."
Sackett's research ranks dozens of selection methods by how well they predict performance. It would be tempting to take the top three and call it done. He warned against that, with a story from the 1980s.
The US supermarket industry was installing its first barcode scanners. The Food Marketing Institute asked Sackett's team to find people who would be fast on the scanner. The range between the fastest and slowest cashiers was enormous. The team built a test that worked.
When he presented it to the heads of HR from the supermarket chains, some were thrilled. Many said it was of no use to them.
One chain competed on speed and low cost, so fast cashiers were exactly what it needed. Another competed on being the friendly store, where the cashier asks how your week has been. A third was most worried about theft and wanted to hire trustworthy people.
Same job title. Three different definitions of a good cashier. Three different sets of selection tools.
"You have to start by saying what do you want."
That's a value choice every organisation has to make for itself. Define it clearly, then pick the tools that measure it.
Once you know what you're looking for, the research on methods is clear. Structured interviews, job knowledge tests and work samples sit near the top. Unstructured interviews come in much lower. Structured interviews predict performance about twice as well as unstructured ones.
The unstructured interview does carry some signal. It beats random selection. But structure is where the gains are, and structure means the same thing for every candidate: the same questions, the same information, evaluated against the same standard.
Sackett also took a detour into a common shortcut: years of experience. A minimum requirement can make sense. Ranking candidates by how far they exceed it does not.
"The number of years of experience above the minimum has no relationship with subsequent performance."
No single tool captures everything about a candidate, so good hiring combines several. The research shows that combining results across multiple tools gives better predictive accuracy.
The question is how you combine them. Most organisations hand the scores to the hiring manager and let them make of it what they will. The alternative is a formula: add the scores up, or weight them based on what the research supports.
People strongly resist the formula. Psychologists call it algorithm aversion. Yet decades of studies, often run by researchers trying to prove the opposite, keep finding the same thing.
"In a battle between an algorithm and judgment, the algorithm wins."
Sackett called it a candidate for the most well-established finding in all of psychology. The algorithm wins for two simple reasons. It treats every candidate the same. And it isn't distracted by irrelevant information, like finding out that a candidate's child goes to the same school as yours.
This is where Sackett parted ways with parts of his own field. Researchers often gather at conferences and complain that hiring managers ignore the evidence. He thinks that's the wrong take.
Managers want control because they live with the decision every day. They also want credit for the teams they build. Sackett respects both.
"Autonomy is crucial."
So he spent the rest of his keynote on strategies that preserve autonomy and still raise decision quality.
1. Show the prediction, let the manager decide. Tell the hiring manager they have full discretion, and share what the algorithm predicts for each candidate. In randomised experiments, managers who see the prediction move much closer to it, while still bringing in their own information.
2. Involve hiring managers in the design. When managers help decide how much each part of the process should weigh, they buy into the result.
3. Let a manager set their own weights. Even if their weights are imperfect, they get applied consistently across every candidate. In one University of Minnesota doctoral study, even random weights applied consistently outperformed unaided human judgment.
4. Use multiple independent judgments. Several decision makers average out noise, but only if each one judges independently before the scores are combined. A consensus meeting loses that advantage. The most senior person, the loudest voice or whoever speaks first ends up deciding.
5. Shortlist with the algorithm, choose with judgment. In US public sector hiring, the law requires a job-related procedure applied consistently to every applicant. The top three go to the hiring manager, who has full discretion. Autonomy stays intact, and the final three are the strongest candidates.
Sackett ended with three messages:
"Do these three things. I can guarantee you you're going to make better hiring decisions."
None of it requires a large corporation with unlimited resources. Any team can start on the next role.
In the Q&A, the host asked about candidates preparing with AI. Sackett's answer: impression management is as old as the interview itself. People have studied common interview questions for as long as there have been interviews. It adds some noise. The structured interview still predicts best on average.
That's the standard Alva is built to. Validated methods, the same structure for every candidate, and an agent that holds that consistency across every role, so hiring managers can focus on the decision.
Next on stage was Malcolm Burenstam Linder, CEO and founder of Alva Labs. He took the principles from Sibony and Sackett and asked what hiring looks like when you put them into practice with agents. → [Read: Malcolm Burenstam Linder on the future of hiring]