Why do AI-assisted teams often make worse decisions together than either the person or the AI would alone — yet do better work when creating something new?

This explores why human-AI teams often do worse than the stronger partner working alone on decisions, yet tend to do better when making things. The main evidence is a meta-analysis of more than 100 studies. On average, human-AI combinations scored below the best of human or AI alone, but content-creation tasks were the exception When do human-AI teams actually outperform either alone?. The authors put the difference down to the structure of the task rather than to what either partner can do. A decision task has one right answer, so each partner's contribution is either a vote for it or noise. A creation task is open-ended, so each partner can add something the other wouldn't have come up with.

In decision tasks, the human usually reviews what the AI suggests, and that is where things break. People anchor on the AI's answer. They also fall into a set of cognitive traps that make each other worse: treating fluent output as if it were the truth, mistaking quick intuition for careful reasoning, and having their existing beliefs confirmed back to them Why do people trust AI outputs they shouldn't?. The outcome is the worst of both: the person defers when the AI is wrong and overrides it when it's right. One proposed fix changes the AI's role. Instead of handing over a decision for the person to accept or reject, the AI points out which parts of the input deserve attention, and the person still makes the call. This reduces anchoring and keeps the human's judgment active Can AI guidance reduce anchoring bias better than AI decisions?.

The less obvious point is that combining human and AI judgments can work on decisions, as long as they aren't run as a review chain. If people and AI make independent predictions on the same cases and you average them, the result beats either one alone whenever the two get different cases wrong. Neither needs to be especially accurate Can humans and AI improve decisions by predicting together?. So the meta-analysis losses may come partly from how collaboration is usually set up, with one partner checking the other, which destroys the independence that makes a crowd smarter. Puranam makes this an organization-design question: how the work is divided and how feedback reaches each partner decide whether the pair learns or gets stuck How should humans and AI systems divide decision-making tasks?.

Creation gains also have conditions. In multi-agent ideation, a mix of perspectives improves idea quality only when the members have real domain expertise. A diverse team without that expertise does worse than one competent agent, because stimulation without knowledge just creates coordination overhead Does cognitive diversity alone improve multi-agent ideation quality?. Time matters as well. AI agents outscore human experts 4× on two-hour research tasks, but humans catch up and pull ahead as the time budget grows When do AI agents outperform human research experts?. This suggests the partners are good at different stages of the work, which suits open-ended making better than one-shot judging.

The takeaway: whether human-AI collaboration helps may depend less on the task category than on whether the setup keeps the two partners' contributions different. Creative work does that naturally. Decision workflows usually remove it by having one partner check the other. They can recover it by collecting judgments independently or by having the AI guide rather than decide.