Can humans and AI improve decisions by predicting together?

Choudhary, Marchetti, Shrestha and Puranam (Journal of Management, online 2023, issue 2025) argue that human and AI decision-makers can be usefully combined through "ensembling" — aggregating separate predictions on the same task — even when "neither has a clear advantage over the other in terms of predictive accuracy, and even if neither alone can attain satisfactory accuracy in absolute terms." This is distinct from the dominant mode of human-AI collaboration in the literature, which the authors call "division of labor with specialization": splitting a task into sub-tasks (screening resumes versus interviewing candidates) and assigning each sub-task to whichever of human or AI is better at it, or using humans as gatekeepers or trainers of AI. Ensembling instead has both human and AI "tackle the same prediction problem," with their separate outputs combined by averaging (estimation problems) or voting (classification problems).

The theoretical basis is the "ambiguity decomposition" (Krogh & Vedelsby, 1995), popularized by Scott Page as the "diversity prediction theorem" and by Surowiecki as "wisdom of the crowd": the error of an averaged crowd estimate is systematically lower than the average individual error whenever diversity across individual errors — not individual accuracy — is positive. The authors stress this diversity is "necessary but not sufficient": it guarantees the ensemble beats the average member, but not that it beats the single best individual predictor, nor that adding more diversity always raises accuracy further. They locate the needed diversity between humans and AI in "H-specific and machine inaccessible data," including humans' capacity to transfer learning from unrelated tasks to the task at hand — a capacity the authors note could erode if AI transfer-learning research advances.

This formalizes the boundary condition that When do human-AI teams actually outperform either alone? leaves open empirically: that review finds decision-task pairings tend to underperform the better of human-alone or AI-alone, and this paper's theorem specifies that such underperformance follows precisely when the two predictors' errors lack diversity, not merely because the task is a "decision task." It also sits in tension with Can aggregated LLM rankings match human expert judgments?, which attributes aggregation gains among multiple LLMs mainly to scale rather than to diversity of judgment — a different mechanism, for a different kind of ensemble (AI-only rather than human-AI). And it extends the same formal-conditions approach as Why do accurate predictions lead to poor decisions?, which likewise argues that whether prediction helps a decision depends on specific structural conditions rather than being automatic.

The paper is a theoretical and conceptual contribution — it derives propositions from existing results (the diversity prediction theorem, prior human-AI collaboration literature) rather than testing them empirically on real managerial decisions. The authors explicitly restrict the theory to prediction-based decisions such as project evaluation, hiring, or investment choices, and state it does not apply to "ensembling of qualitative outputs between H and generative AI" (text, images, music) or to managerial tasks like motivating employees or complying with legal requirements. The practical implication, at the strength the paper supports, is that a manager weighing whether to combine personal judgment with an AI model's prediction should look for evidence that the two sources of error diverge, not for evidence that either source is individually reliable.