ÃÛÌÒ´«Ã½Æƽâ°æÏÂÔØ

Facial-recognition systems often rely on a relatively simple safeguard. AI identifies a potential match, then a human reviews it before any action is taken.Ìý

But new research suggests the effectiveness of that safeguard may depend on who the human reviewer is.ÌýÌýÌýÌý

The study, published in February 2026 in theÌý, found that humans and facial-recognition algorithms generally agree most with people who scored highest on tests of face-recognition ability. But people vary widely in their ability to tell faces apart, meaning some human reviewers may be much better than others in evaluating an AI-generated match.Ìý

"We found that the AI systems are actually pretty good overall," saidÌýDavid Dobolyi, lead author of the study and assistant professor of organizational leadership and information analytics at theÌýLeeds School of Business. "They are able to perform similarly to higher-scoring individuals."

David DobolyiÌý

Those high-scoring people are sometimes called "super recognizers." People vary widely in how well they can tell faces apart, and the study found that stronger face recognizers were much more likely to see facial similarities the same way the algorithms did. Researchers measured those differences using a standard face-recognition test administered to nearly 4,000 participants.

That finding matters because many facial-recognition systems operate with a "human in the loop." The software generates potential matches, but a person reviews the results before any action is taken.

"I wouldn't say that having a human in the system solves the problem," Dobolyi said. "If you take the average human, that person may not be better than a specific model."

"I think just like an AI model, you need to have a measure of how good a person is at this task," he said.

That complicates a common narrative around facial recognition that AI makes the mistakes and humans catch them.

Dobolyi said the research reveals a more complicated reality: Both humans and algorithms have strengths, weaknesses and blind spots.

"People have their own biases. These systems have their own biases," he said. "It's not about one being better than the other. It's about understanding the pros and cons of each, or what one system is good at, what one system is bad at, including among people."

Different algorithms, different results

The researchers also found that not all facial-recognition systems see faces the same way.

When researchers compared six commercial and open-source systems, they found differences in how closely the algorithms aligned with human judgments. They also found that agreement sometimes differed depending on whether participants were comparing Black orÌýwhite faces. But those differences varied from one facial-recognition system to another.

For Dobolyi, that's a reminder that organizations shouldn't assume all facial-recognition tools work equally well.

"This is why it's important to do this kind of work, because you need to ensure that your system is fair across the board," he said. "You don't want to have any spot where you can't trust the outputs of the AI system."

Another challenge is that many commercial AI systems don't show their work.Ìý

"We don't really have explanations for why an AI system makes the decision it does," Dobolyi said. "So there's still this black box problem."

The study builds on another line of research Dobolyi has pursued on the role AI can play in high-stakes decisions. In a 2025 paper published inÌý, he and his co-authors found that AI tools could help reduce bias in how people evaluateÌýeyewitness testimony.

Both studies arrive at a similar conclusion. AI has the potential to improve decision-making, but it works best when people understand what it can do, what it can't and when to trust its recommendations.

As facial-recognition technology becomes more common, Dobolyi said he hopes organizations become more transparent about what these systems can and can't do and more rigorous about evaluating them before putting them to use.

"I think the study emphasizes the importance of more testing and more openness about what the capabilities of systems like this really are, what their limitations are, and the level of confidence and trust we should place in them," he said.