Can humans and AI teams achieve performance neither reaches alone?

The paper develops a conceptual framework for "complementary team performance" (CTP) — "a level of performance that neither of them can attain individually" — built around two measures, complementarity potential (CP) and complementarity effect (CE), each of which the authors split into an "inherent" and a "collaborative" component. The inherent component "captures decision-making synergies that, for each task instance, can be attributed to the individually more accurate team member within the human-AI team"; the collaborative component "captures decision-making synergies that only emerge through collaboration resulting in team decisions which are more accurate than each of the individual ones." The authors identify "information and capability asymmetry as the two key sources of complementarity" that drive both components, and note that CTP "has rarely been observed" in the literature, which they attribute to an "insufficient understanding of the principle."

The mechanism is illustrated with a worked example of 25 decision instances: the AI gets 13 wrong, the human gets 15 wrong independently, and 5 instances neither can solve alone. If the team simply picked whichever member was individually correct on each instance, it would resolve 20 of the 25 — reducing the AI's own error count from 13 to 5, an "improvement potential of 8 decisions" that the paper attributes to inherent potential. The remaining 5 instances, where neither party is individually correct, represent a further, collaborative potential — illustrated with a radiologist who initially calls a scan "malign" and an AI that calls it "no cancer," each wrong alone, but who can reach the correct "benign" diagnosis if the human learns the AI's rationale (e.g., which regions of the X-ray drove its decision) or the AI's output is corrected by patient-history information the human alone holds. Information asymmetry concerns which data each party can access (the paper gives lidar sensor data unavailable to a human driver, or patient history unavailable to an AI model); capability asymmetry concerns differences in how each processes shared information, since "AI models encode relationships inferred from training data, whereas humans employ compositional mental models."

This gives a possible structural account for why When do human-AI teams actually outperform either alone? finds decision-task combinations underperforming on average: that review's own prescription — split the task so each party does only the parts it is better at — is close to what this paper's inherent-potential mechanism formalizes, and its failure in most studies matches this paper's claim that CTP "has rarely been observed." The relationship to How should humans and AI systems divide decision-making tasks? is one of parallel conceptual framing rather than overlap: Puranam's Type C condition (independence, modest and roughly equal individual accuracy) describes when aggregation can beat either member alone, while this paper's inherent/collaborative split describes two different routes once aggregation is attempted — one a byproduct of picking the already-better answer per instance, the other a product of exchanging information or capability through interaction. Can humans and AI improve decisions by predicting together? offers a formal result adjacent to the collaborative component here, since both describe value that does not reduce to the team just deferring to whichever member was right.

The excerpt is explicit that its evidence comes from "two empirical studies" using "laboratory-based experiments and vignette methodologies," which the authors flag as a limitation on "generalizability and its practical implications" — the framework itself is not validated outside these controlled settings. It also states plainly that the work does not address "motivation... engagement... self-efficacy... or behavior patterns within teams," factors the authors say also shape human-AI team effectiveness. The implication the excerpt supports is narrow: the CP/CE decomposition offers a vocabulary for diagnosing why a given human-AI team over- or under-performs, not yet a demonstrated recipe for producing CTP in real-world, non-lab decision settings.