When a human and an AI know different things, does that gap make the team smarter — or just create new ways to fail?

This looks at what happens when a human and an AI each know different things, and whether that gap helps the team, hurts it, or does both depending on how the work is set up. The short answer from the corpus is that information asymmetry is the main reason human-AI teams can beat either partner alone. It is also one of the main reasons they usually don't. One framework splits team gains into two sources: picking whichever partner has the better answer on a given task, and actually trading information or ways of reasoning so the combined answer improves Can humans and AI teams achieve performance neither reaches alone?. Differences in what each side knows feed both sources. But the same work finds that real complementarity is rare, because the conditions that let a team use those differences are seldom in place.

The common failure is that the human doesn't use their own information. They anchor on the AI's answer instead. One study tried 'cognitive forcing functions', which are deliberate points of friction that make people think before seeing the AI's answer. These did reduce overreliance, but the teams still did worse than the AI working alone, and participants disliked the friction Do forcing functions reduce AI overreliance without harming team performance?. Another approach changes what the AI sends. In 'Learning to Guide', the AI doesn't hand over a verdict. It points out which parts of the input deserve attention, so the human's own knowledge stays involved in the decision instead of being overridden Can AI guidance reduce anchoring bias better than AI decisions?. So the asymmetry pays off only if the setup keeps both kinds of knowledge active.

This makes the problem partly one of organizational design. Puranam treats a human-AI pair as a small organization in which someone has to decide who handles which kind of task and how each side learns from feedback How should humans and AI systems divide decision-making tasks?. Agrawal, Gans and Goldfarb make a related point at company scale: AI improves individual decisions but also makes them vary more, so linked decisions need new communication channels to stay coordinated Does AI adoption require rethinking how organizations coordinate?. How information moves also matters. In a small experiment, changing how humans and agents communicated changed how much people trusted the agent and how well they tracked the shared workspace How do communication modalities shape human-agent collaboration patterns?.

The less obvious finding comes from multi-agent research. When one team member's goals quietly drift, information asymmetry makes the damage much worse. In adversarial team games, a single misaligned agent hurt outcomes most when roles were specialized and information was unevenly spread, because teammates trusted what they couldn't check Does one misaligned agent harm a team in adversarial settings?. That trust can build up over time. In repeated partner-selection games, people came to prefer AI partners because the AI behaved more consistently Do humans learn to prefer AI partners over time?. Put together, these suggest that the asymmetry that makes a partner valuable is the same thing that makes its failures hard to see.

A caveat: the corpus has a lot on the parts around this question, such as complementarity, overreliance and coordination. It has little that measures information asymmetry directly as a variable in human-AI teams. The main takeaway is that a knowledge gap between partners is useful only if the team is set up to bring it out. Otherwise the human defers to the AI or the team trusts it blindly, and the gap goes to waste.