Here is an uncomfortable finding from AI research.

When researchers combine all the experiments in which a person and an AI work on a task together, the team often does worse than the better of the two working alone.

A 2024 meta-analysis by Vaccaro and colleagues found exactly that. On average, human–AI combinations failed to beat the stronger partner.

A new paper in PNAS asks a simple question about those experiments: did the people ever get a chance to learn?

The answer, mostly, is no.

Almost none of 74 human–AI studies let people learn

A team led by Julian Berger and Stefan Herzog at the Max Planck Institute for Human Development went back through the 74 studies in the earlier meta-analysis.

They looked at how the experiments were designed.

  • Only 10 of the 74 studies gave people outcome feedback, meaning they were told whether their answers were right.
  • None experimentally tested what feedback does.
  • The typical study had about 30 trials, and most had no practice rounds.

A separate review the authors cite found outcome feedback in just 6 of 124 studies.

So in most of these experiments, people were handed an AI tool, given a short set of tasks, and never told how they or the AI were doing.

Under those conditions, it’s hard to know when to trust the AI and when to overrule it. And that judgment is the whole point of a human–AI team.

With feedback, teams beat the stronger partner alone

The authors compared studies with and without feedback.

Without feedback, human–AI teams tended to underperform. With feedback, they tended to do slightly better than the stronger partner alone.

The clearest result involved AI explanations, where the AI explains why it made a recommendation.

Explanations are often presented as the solution to the trust problem. But the re-analysis suggests they help only when people also receive feedback.

  • Explanations with feedback: teams tended to outperform.
  • Explanations without feedback: teams did noticeably worse.

The interpretation is intuitive. An explanation tells you what the AI is thinking. Feedback tells you whether that kind of thinking turns out to be right. Without the second, the first may just make people more confident in advice they can’t evaluate.

Learning from feedback is nothing new, and that is the point

None of this would surprise a teacher, a coach or anyone who has trained new employees.

People get good at a task by doing it, seeing the result, and adjusting. Professions that have fast, clear feedback tend to develop well-calibrated judgment. Professions without it often don’t.

Working with AI is a skill like any other. Knowing when a tool is reliable, and when it’s out of its depth, is learned.

What the paper shows is that much of the research on human–AI collaboration has quietly left that learning out, and then concluded that humans and AI don’t combine well.

The science is now catching up with something long known about how people build expertise.

The evidence is suggestive, not conclusive

It’s suggestive, not conclusive, and the authors say so.

This is a re-analysis of existing studies, not a new experiment. Because feedback was never varied deliberately within a study, the comparison is between different studies that happened to include or omit it. Those studies may differ in other ways too.

The number of studies with feedback is also small, and most of the statistical estimates have uncertainty ranges that include zero. The difference between feedback and no feedback was positive with about 84% probability, which is encouraging but far from certain. The explanation-plus-feedback result was the most robust.

The studies also involved structured tasks with clear right answers. Open-ended creative or strategic work is harder to give feedback on.

Helping people use AI well gets far less attention than better models

The technology industry spends enormous sums making AI models more capable. Far less attention goes to helping people use them well.

This paper doesn’t show that training people beats improving models. It wasn’t designed to compare the two.

What it shows is that the human side of the partnership has been under-studied and under-supported, and that the simplest learning tool there is, telling people how they did, is missing from most of the research.

For organizations, the practical lesson is old-fashioned: give people practice, show them when the AI was right and wrong, and let their judgment develop.

A better AI helps.

A person who has learned when to trust it may help more.