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.

Artificial intelligence (AI) is increasingly used to support human decision making in high-stake settings in which the human operator, rather than the AI algorithm, needs to make the final decision. For example, in the criminal justice system, algorithmic risk assessments are being used to assist judges in making pretrialrelease decisions and at sentencing and parole [20, 69, 65, 18]; in healthcare, AI algorithms are being used to assist physicians to assess patients' risk factors and to target health inspections and treatments [63, 26, 77, 49]; and in human services, AI algorithms are being used to predict which children are at risk of abuse or neglect, in order to assist decisions made by child-protection staff [79, 16]. In such systems, decisions are often based on risk assessments, and statistical machine-learning algorithms' abilities to excel at prediction tasks [60, 21, 34, 68, 62] are leveraged to provide predictions as advice to human decision makers [45]. For example, the decision that judges make on whether it is safe to release a defendant until his trial, is based on their assessment of how likely this defendant is, if released, to violate his release terms, i.e., to commit another crime until his trial or to fail to appear in court for his trial. For making such risk predictions, judges in the US are assisted by a "risk score" predicted for the defendant by a machine-learning algorithm [20, 69].

Congestion is a common failure mode of markets, where consumers compete inefficiently on the same subset of goods (e.g., chasing the same small set of properties on a vacation rental platform). The typical economic story is that prices solve this problem by balancing supply and demand in order to decongest the market. But in modern online marketplaces, prices are typically set in a decentralized way by sellers, with the power of a platform limited to controlling representations -- the information made available about products. This motivates the present study of decongestion by representation, where a platform uses this power to learn representations that improve social welfare by reducing congestion. The technical challenge is twofold: relying only on revealed preferences from users' past choices, rather than true valuations; and working with representations that determine which features to reveal and are inherently combinatorial. We tackle both by proposing a differentiable proxy of welfare that can be trained end-to-end on consumer choice data. We provide theory giving sufficient conditions for when decongestion promotes welfare, and present experiments on both synthetic and real data shedding light on our setting and approach.