Does better prediction change which human judgment matters most?

Agrawal, Gans and Goldfarb's formal model of decision-making under uncertainty argues that improvements in machine prediction change which kind of human judgment is worth having, and the direction of that change depends on how cheap prediction becomes. They write that "we demonstrate that better prediction impacts the returns to different types of judgment in opposite ways. Hence, not all human judgment will be a complement to AI." In the model, a decision-maker chooses between a safe action with payoff S and a risky action whose payoff is R in a good state or r in a bad state, with R > S > r; judgment here means noticing "hidden attributes" — hidden opportunities that raise R, or hidden costs that lower it — that the decision-maker could not describe in advance.

Without any prediction, the decision-maker defaults to the safe action, so judgment is valuable only insofar as it surfaces hidden opportunities that would justify the riskier, more profitable choice — what the authors call judgment "regarding good news." Once a prediction technology lowers uncertainty and makes the risky action the default instead, the authors argue the valuable judgment flips: "the role of judgment is to move away from the riskier action to safer ones... if that judgment is directed at identifying hidden costs." Their conclusion states the trend directly: "as the cost of prediction falls and prediction is applied to more decisions, then the type of judgment that is valuable will move from judgment regarding good news to judgment regarding bad news."

This sharpens What makes accountable judgment scarce when AI cognition is cheap? by specifying which judgment becomes scarce: not judgment in general, but specifically the capacity to catch downside risk once prediction has already pushed decision-makers toward the risky default. It also complements Why do accurate predictions lead to poor decisions?: where that paper formalizes when a prediction model's fit-to-data objective supports optimal policy at all, this paper takes good predictions as given and asks what happens to the human judgment layered on top of them.

The model is a stylized two-action, two-state setup with a single consequential hidden attribute at a time (opportunities and costs are treated as mutually exclusive events); the excerpt does not generalize this to multi-action real organizational decisions, nor does it measure the effect empirically — it is a theoretical result about comparative statics in a formal model, not an observed shift in practice. The paper also only sketches, rather than works out, the related possibility that judgment interventions become training data that eventually lets machines learn to mimic the judgment itself.