Accepted Papers

“Human–AI teaming” has become the default vocabulary for expert–AI collaboration, and is increasingly used to justify design and procurement decisions in domains far from the crews it was built on. We found that the construct carries three unstated preconditions: jointly attributable outcomes, feedback fast enough to calibrate reliance within the working life of the collaboration, and accountability that can genuinely be shared. We test them against expert financial judgment, a domain that adopts AI advisors at scale and, unusually, audits every decision and every outcome. First, we define the calibration horizon τ, the calendar time before outcome evidence alone can establish that an AI advisor helps, and evaluate it across twelve professions using power analysis and the statistics of Sharpe-ratio inference. It ranges from about a week in flight operations to 196 years for a discretionary equity portfolio manager, while a two-hundred-manager firm running a randomised-withholding programme reaches the same evidence in about four. Identification is available to the institution and not to the individual, which inverts the relation the construct assumes, since its prescriptions are addressed to teammates. Second, we show that the dyad is the limited unit of analysis. When M experts share one advisor, privately optimal reliance cuts each expert’s squared error by 67% while making the consensus 22× worse, and we derive a correlation budget of roughly M/4 independently-erring models. We conclude that in this domain “teaming” describes the work less accurately than it redistributes, in language, an accountability that law makes non-delegable, and we propose delegation under agency, whose four primitives are contractible rather than psychological.

Human–AI decision systems are often called “teams” when artificial intelligence (AI) assistance improves human performance. That comparison cannot carry the label: a system can help the people using it and still be worse than the AI working alone. We propose TEAM-Audit, a two-baseline audit for the performance version of the claim: evaluate the final, frozen system on data not used to build it; compare it separately with human-only and with AI-only performance; and only when both comparisons are supported with uncertainty call the result performance complementarity — “beats both.” For systems that route cases between a human and an AI, the audit corrects historical logs that record only the acting party’s outcome, and it reports human workload beside accuracy. In simulation, ignoring development-stage uncertainty produced false “beats both” verdicts in 0.504 of null replications, and reusing one sample to both select and evaluate a policy produced 0.103. Reanalyzing a public human–AI study, all 14 assisted conditions beat both baselines when the study’s questions are treated as fixed, but only 2 of 14 when the uncertainty also covers the broader item population those questions stand in for. The verdict turns on a choice that is almost never stated.