AI-Assisted Decision Making with Human Learning

TL;DR

This work analyzes AI-assisted decision-making where an algorithm selects which features a learning human can use to predict outcomes. By modeling the human's evolving beliefs with a $\phi$-convergent learning dynamic and optimizing a discounted loss with horizon $\delta$, it reveals that there exists a stationary optimal feature-sequence that can be computed efficiently in $\Theta(n \log n)$ time. The authors characterize when the algorithm should favor more informative versus less divergent features, showing that higher patience and faster human learning push toward informative feature sets, and that early learning yields greater long-run benefits. They also develop misspecification bounds showing that the impact of model errors is limited when the algorithm only guides what the human uses, not the final prediction, with practical implications for dashboard design and decision-support systems.

Abstract

AI systems increasingly support human decision-making. In many cases, despite the algorithm's superior performance, the final decision remains in human hands. For example, an AI may assist doctors in determining which diagnostic tests to run, but the doctor ultimately makes the diagnosis. This paper studies such AI-assisted decision-making settings, where the human learns through repeated interactions with the algorithm. In our framework, the algorithm -- designed to maximize decision accuracy according to its own model -- determines which features the human can consider. The human then makes a prediction based on their own less accurate model. We observe that the discrepancy between the algorithm's model and the human's model creates a fundamental tradeoff. Should the algorithm prioritize recommending more informative features, encouraging the human to recognize their importance, even if it results in less accurate predictions in the short term until learning occurs? Or is it preferable to forgo educating the human and instead select features that align more closely with their existing understanding, minimizing the immediate cost of learning? This tradeoff is shaped by the algorithm's time-discounted objective and the human's learning ability. Our results show that optimal feature selection has a surprisingly clean combinatorial characterization, reducible to a stationary sequence of feature subsets that is tractable to compute. As the algorithm becomes more "patient" or the human's learning improves, the algorithm increasingly selects more informative features, enhancing both prediction accuracy and the human's understanding. Notably, early investment in learning leads to the selection of more informative features than a later investment. We complement our analysis by showing that the impact of errors in the algorithm's knowledge is limited as it does not make the prediction directly.

Figures (2)

• Figure 1: The ratio of discounted losses of the more informative sequence over the less informative sequence for the example in Section \ref{['sec:example-learning']}.
• Figure 2: The transition curve for the example in Section \ref{['sec:example-learning']}: For $\delta$ values to the right of this curve, selecting the more informative feature (of the two) is preferred, while for $\delta$ values to the left of this curve, the less divergent feature is preferred. The plot demonstrates that the empirical curve in Figure \ref{['fig:w-delta-heatmap']} coincides with the theoretical curve derived in \ref{['clm:educating-two-features']}.

Theorems & Definitions (48)

• Claim 3.1
• proof
• Definition 3.2
• Definition 4.1
• Lemma 4.2
• proof
• Corollary 4.3
• Proposition 4.4
• Definition 5.1
• Claim 5.2