When do human-AI teams actually outperform either alone?

A pre-registered systematic review and meta-analysis of "over 100 recent experimental studies reporting over 300 effect sizes" finds that "on average, human-AI combinations performed significantly worse than the best of humans or AI alone" (Hedges' g = −0.23, 95% CI −0.39 to −0.07). The paper distinguishes two bars: "human-AI synergy," where the combination beats both the human alone and the AI alone, and the weaker "human augmentation," where the combination beats the human alone. Synergy failed on average, but augmentation held — "the human-AI groups performed better than the humans working alone" — so the AI still helped, just not enough to clear the higher bar most people have in mind when they talk about combining humans and AI. Task type moderated the result sharply: "decision tasks associated with performance losses and creation tasks associated with performance gains." A third pattern cut across both: "when humans outperformed AI alone, we found performance gains in the combination, but when the AI outperformed humans alone we found losses."

The authors' explanation for the task-type split is structural, not about who is smarter. Creation tasks "often also involve substantial amounts of somewhat routine generation of additional content that AI can perform as well or better than humans" — the example given is generating an image or document, where creative direction needs a human but "filling in boilerplate or routine parts" doesn't. Decision tasks, by contrast, were mostly structured so that "both the human and the AI system make a complete decision, with the humans usually making the final choice" — each party does the whole task redundantly rather than splitting it by comparative strength. The paper's own prescription follows from this: "better results might have been obtained if the experimenters had designed processes in which the AI systems did only the parts of the task for which they were clearly better than humans."

This gives the decision-support literature's overreliance findings a structural explanation rather than just another instance. Do forcing functions reduce AI overreliance without harming team performance? found that even after cutting overreliance, "human+AI teams in our study continued to perform worse than the AI model alone" on a decision task — exactly the pattern this meta-analysis predicts for full-decision designs, and a sign that friction-based fixes may be addressing the wrong lever if the task structure itself (redundant whole-decision-making rather than split labor) is what caps performance. Why don't radiologists benefit from AI predictions? is a second decision-task case of the same loss pattern, located in a belief-updating error rather than a task-design one — the two accounts are compatible: a radiologist re-deciding the whole case redundantly is also a case of not splitting the task by comparative advantage. On the other side, Can generative AI replace the benefits of having a human teammate? sits in the creation-task category (open-ended idea generation), consistent with this paper's gain finding there. Does theory of mind predict who thrives in AI collaboration? offers a different, person-level moderator (perspective-taking) alongside this paper's task-level one (task type, relative skill); the two are not in tension, just operating at different grains.

The review's own limitations bound the claim. It only includes studies reporting all three conditions — human alone, AI alone, and the combination — so it says nothing about tasks neither party can do alone. Its effect sizes are pooled via Hedges' g across different outcome measures (accuracy, error, quality) and different populations ("some look at doctors … others at crowdworkers … and still others at students"), which the authors say "can limit the comparability of the effect size to a degree." The result is an average across a heterogeneous literature, not a guarantee for any specific task or team. What it does establish, at the strength a meta-analysis allows, is a design-relevant correlation: redundant whole-task designs cluster with losses, and designs (or task types) that split work by comparative advantage cluster with gains — a reason to look at how a human-AI task is divided before asking why the combination underperforms.