This dissertation studies learning and information, using theoretical models in the first two chapters and an empirical analysis in the third. The questions are motivated by computer science applications and by an interest in human–AI collaboration.The first chapter models human–algorithm collaboration using tools from information design and delegation. A principal designs an algorithm that generates a publicly observable prediction of a binary state. She must decide whether to act directly based on the prediction or to delegate the decision to an agent with private information but potential misalignment. I study the optimal design of the prediction algorithm and the delegation rule in such environments. I find that providing the most informative algorithm may be suboptimal even if the principal can act on the algorithm’s prediction. Instead, the optimal algorithm may provide more information about one state and restrict information about the other. Moreover, well-intentioned policies aiming to provide more information, such as keeping a “human-in-the-loop” or requiring maximal prediction accuracy, could strictly worsen decision quality compared to systems with no human or no algorithmic assistance. These findings predict the underperformance of human–machine collaborations if no measures are taken to mitigate common preference misalignment between algorithms and human decision-makers. The second chapter, joint work with Nicole Immorlica, Brendan Lucier, and Clayton Thomas, quantifies the gap between learning aggregated information (e.g., LLM summaries) and learning raw data points. We often learn from multiple sources that convey information in different ways. How informative is it to know the source of a signal, and how is this informativeness shaped by the distribution of sources? We extend the standard (binary-state, binary-signal) Blackwell experiment model by introducing a commonly known distribution over signaling schemes, representing the distribution of information sources. We compare learning under two information models: source-aware, where decision makers observe a signaling scheme and its realization (for example, raw reviews, search results), and source-blind, where only the signal realization is observed (for example, aggregate ratings, generated summaries). We show that a mean-preserving spread in the distribution of signaling schemes translates into Blackwell dominance for source-aware decision makers, implying they are “risk-loving” in information sources. In contrast, it has no impact on source-blind decision makers. When learning from repeated draws of signaling schemes, source-blind decision makers learn more slowly. However, as long as the average signaling scheme is epsilon away from being completely uninformative, source-blind learning can match source-aware learning by using at most O(1/epsilon) times more data. The third chapter, joint work with Germán Reyes and Evan Riehl, studies empirically how the stakes of a signaling mechanism affect the informativeness of the signals. We examine a natural experiment in Brazil in which similar students took the same standardized test as either a low-stakes school accountability exam or a high-stakes admission exam for the country’s top universities. Using administrative data and a difference-in-differences design, we find that test score gaps between high- and low-income students expanded on the high-stakes exam, consistent with wealthy students engaging in test prep. Yet the increase in stakes made scores more informative for students’ college outcomes. Thus the “muddling” of information on natural ability and test prep improved the quality of the score signal, although it also exacerbated inequality.