Does error diversity matter more than accuracy for human replacement?

Within the same ensembling framework, Choudhary, Marchetti, Shrestha and Puranam derive a second, more specific result about composition: when an ensemble must be built by either augmentation (adding AI to an existing human team) or automation (AI replacing a human in it), the human most likely to be displaced is not the least accurate one. "It is not necessarily the H with the lowest accuracy who is at greatest risk of being replaced in the ensemble but, rather, the one who adds the least diversity of prediction error relative to the AI."

This follows from the same diversity-prediction logic underlying Can humans and AI improve decisions by predicting together?: because ensemble accuracy depends on the dispersion of member errors rather than on any one member's accuracy, a highly accurate human whose errors closely track the AI's contributes little beyond what the AI already supplies, while a less accurate human whose errors diverge from the AI's can still raise the ensemble's overall performance. The authors frame this as a reason automation (replacement) and augmentation (addition) — categories borrowed from Raisch and Krakowski (2020) — are not simply a referendum on whether "AI is better than this person"; they function as a referendum on error overlap between the two.

This complicates How should AI agents and humans divide research tasks?, which finds humans retaining final decision authority regardless of relative accuracy: Choudhary and colleagues suggest that even where humans keep formal final say, the economic case for keeping a specific human inside an ensemble turns on how differently that person errs from the AI, not on titles or final-say arrangements. It also stands apart from Do language model groups mimic human group reasoning patterns?, where the ensemble members are all LLMs and the diversity that matters is unique information surfaced during discussion — here the diversity that matters is statistical divergence in prediction error between a human and an algorithm, with no discussion involved at all.

As with the parent proposition, this is a derived theoretical result (the paper's "Proposition 2"), not an observed pattern from real replacement decisions inside organizations — the excerpt offers no field or experimental data testing whether managers who replace low-diversity humans actually see the predicted gains. The implication the paper supports is narrower than "the AI judges who to cut": it is that if a firm wants to decide which human to retain alongside an AI ensemble, measuring how that human's errors correlate with the AI's is a more diagnostic signal than measuring the human's standalone accuracy.