Human-AI teams: learning from feedback may be the missing piece
Few experiments in a much-cited 2024 review let people learn from their AI partner, and learning conditions shifted how well human-AI teams performed.
By
Ethan Reid
In brief
- A 2024 review of 74 studies had found human-AI teams doing worse than the better partner alone.
- A new analysis finds that only 10 of those studies offered any feedback to participants.
- Teams that had both feedback and explanations tended to beat the stronger partner working alone.
- The authors say experiments that directly vary learning opportunities are still needed.
Human-AI teams may look weaker than they are because most experiments never gave participants a chance to learn how the AI behaves. A fresh look at earlier data, directed by scientists from the Max Planck Institute for Human Development in Berlin, argues that the bleak picture painted so far was too gloomy.
A 2024 verdict gets a second look
Pairing people with AI is widely expected to beat either one working alone. Yet a review that made waves in 2024 concluded the opposite. It analyzed 74 studies on human-AI interaction, covering tasks such as classifying images, predicting house prices and writing code, and found a negative average synergy effect. The human-AI team performed worse than the better of the two alone, which fueled doubts about pursuing this kind of collaboration.
The new paper appeared in the Proceedings of the National Academy of Sciences. An international group headed by Julian Berger and Stefan Herzog, both at the institute’s Center for Adaptive Rationality, wrote it. Its central message, in Berger’s words: “The potential of human-AI collaboration is underestimated.” The findings were reported by Tech Xplore.
What the team looked for in the same 74 studies
Rather than collecting new data, the researchers returned to the existing studies. They asked two questions. Did results differ when the AI explained its output? And did they differ when people were told the right answer after each decision, so that they could learn and adapt to the algorithm?
Learning opportunities turned out to be scarce. Feedback of any kind reached participants in just 10 of the 74 studies, and not one of them manipulated it as an experimental variable. “Learning effects in human-machine interaction remain virtually unexplored,” said Berger, who explored this topic in his doctoral dissertation.
To separate the effects, the team ran a meta-regression, a statistical method. It set studies with feedback against those without, while accounting for whether the AI explained its recommendations.
Three conditions, three different outcomes
| Condition | Effect on synergy |
|---|---|
| Feedback and explanations | Team tended to outperform the better partner alone |
| Explanations only | Distinctly negative synergy |
| Feedback only | No clearly discernible difference |
Explanations on their own were associated with worse results, while feedback on its own made no clear difference. Only the combination pointed to teams doing better than the stronger partner alone. The authors write that explanations improve synergy mainly when humans can learn to verify the AI’s reliability through feedback.
Limits, chatbots and the risk of offloading
Because no study varied learning opportunities within a single experiment, the comparison rests on how the studies were designed. The authors stress that stronger causal conclusions would require experiments that manipulate those opportunities directly.
Whether the results carry over to generative AI and large language models remains an open question. Herzog points out that a chatbot can give wrong answers, so its replies are not independent feedback. In his view, conversation with a chatbot could make personalized feedback and explanations easier to provide, but only if people can find out what is actually correct.
The authors ask for a change of paradigm in research on human-AI interaction, one that deals openly with human learning. They describe outcome feedback and AI explanations as a promising but understudied lever. Herzog sums up the practical lesson: “When designing AI systems and interactions with them, we need to consider humans as learning beings from the outset.”
Photo: Anna Shvets / Pexels
