To what extent can the strength of a local urban community impact neighborhood safety? We construct measures of community vibrancy based on a unique dataset of block party permit approvals from the City of Philadelphia. Our first measure captures the overall volume of block party events in a neighborhood whereas our second measure captures differences in the type (regular versus spontaneous) of block party activities. We use both regression modeling and propensity score matching to control for the economic, demographic and land use characteristics of the surrounding neighborhood when examining the relationship between crime and our two measures of community vibrancy. We conduct our analysis on aggregate levels of crime and community vibrancy from 2006 to 2015 as well as the trends in community vibrancy and crime over this time period. We find that neighborhoods with a higher number of block parties have a significantly higher crime rate, while those holding a greater proportion of spontaneous block party events have a significantly lower crime rate. We also find that neighborhoods which have an increase in the proportion of spontaneous block parties over time are significantly more likely to have a decreasing trend in total crime incidence over that same time period.

Gig economy companies benefit from labor flexibility by hiring independent workers in response to real-time demand. However, workers' flexibility in their work schedule poses a great challenge in terms of planning and committing to a service capacity. Understanding what motivates gig economy workers is thus of great importance. In collaboration with a ride-hailing platform, we study how on-demand workers make labor decisions; specifically, when to work and for how long. Our model offers a way to reconcile competing theories of labor supply regarding the impact of financial incentives and behavioral motives on labor decisions. We are interested in both improving how to predict the behavior of gig economy workers and understanding how to design better incentives. Using a large comprehensive dataset, we develop an econometric model to analyze workers' labor decisions and responses to incentives while accounting for sample selection and endogeneity. We find that financial incentives have a significant positive influence on the decision to work and on the work duration-confirming the positive income elasticity posited by the standard income effect. We also find support for a behavioral theory as workers exhibit income-targeting behavior (working less when reaching an income goal) and inertia (working more after working for a longer period). We demonstrate via numerical experiments that incentive optimization based on our insights can increase service capacity by 22% without incurring additional cost, or maintain the same capacity at a 30% lower cost. Ignoring behavioral factors could lead to understaffing by 10-17% below the optimal capacity level. Lastly, inertia could be a potential sign of workers' loyalty to the platform.

Workers spend a significant amount of time learning how to make good decisions. Evaluating the efficacy of a given decision, however, can be complicated --- e.g., decision outcomes are often long-term and relate to the original decision in complex ways. Surprisingly, even though learning good decision-making strategies is difficult, they can often be expressed in simple and concise forms. Focusing on sequential decision-making, we design a novel machine learning algorithm that is capable of extracting "best practices" from trace data and conveying its insights to humans in the form of interpretable "tips". Our algorithm selects the tip that best bridges the gap between the actions taken by the human workers and those taken by the optimal policy in a way that accounts for which actions are consequential for achieving higher performance. We evaluate our approach through a series of randomized controlled user studies where participants manage a virtual kitchen. Our experiments show that the tips generated by our algorithm can significantly improve human performance relative to intuitive baselines. In addition, we discuss a number of empirical insights that can help inform the design of algorithms intended for human-AI interfaces. For instance, we find evidence that participants do not simply blindly follow our tips; instead, they combine them with their own experience to discover additional strategies for improving performance.

Increasingly, work happens through human collaboration with generative AI (e.g., ChatGPT). In this paper, we conduct a qualitative study of this collaboration for real-life work tasks. We focus our study on US public school teachers (N = 24) who must regularly complete text-generation tasks including creating quizzes, slide decks, word problems, reading passages, lesson plans, classroom activities, and/or projects. In one-on-one video- and audio-recorded virtual sessions, we observe each teacher use ChatGPT-4 for work tasks of their choosing for 15 minutes, and debrief their experience. Analyzing 201 prompts inputted by the 24 teachers, we uncover four main modes in which the teachers request support from ChatGPT: (1) make for me (55% of prompts), (2) find for me (15%), (3) jump-start for me (10.5%), and (4) iterate with me (15.5%). When we analyze the data at the conversation level, where a conversation is one or more prompts about the same topic, we find 66 out of the 80 teacher-ChatGPT conversations in our data employ only one support mode. In only 14 conversations did teachers combine different modes of support. A follow-up survey of the same 24 teachers reinforces our evidence of the prompt- and conversation-level variety in ChatGPT use "in the wild." Our findings contribute to an emerging theory of generative AI use in practice.

As gig economy platforms increasingly rely on algorithms to manage workers, understanding how algorithmic recommendations influence worker behavior is critical for optimizing platform design and improving worker welfare. In this paper, we investigate the dynamic interactions between gig workers and platform algorithms, with a particular focus on how workers learn to improve their strategy and performance over time. Using multiple quantitative methods, including two-way fixed-effects regression and multinomial logit modeling, we analyze over one million orders completed by gig workers on a retail delivery platform. Our findings reveal a clear learning curve: workers progressively improving their efficiency and on-time delivery performance with increased experience. We also find that while newcomers heavily rely on algorithmic recommendations for task selection, more experienced workers tend to deviate from these recommendations, developing and employing personalized strategies. This shift suggests that experienced workers may perceive algorithmic recommendations as less beneficial or misaligned with their evolved preferences, highlighting the necessity for adaptive recommendation systems. Our research underscores the importance of designing human-centric recommendation algorithms that accommodate workers' learning trajectories, incorporate their feedback, and offer flexibility to support personalized strategies, ultimately enhancing collaborative dynamics and outcomes for both workers and platforms.

Consumers often perceive higher-priced products to have higher quality. Less is known on how quality perception is affected by price markdowns. In addition, it is an open question whether and how consumers' ex-ante expectation on a future markdown affects their quality perception as well as purchase decisions. We answer these questions in a markdown setting under a fixed inventory. This paper adds to the growing literature that incorporates consumers' behavioral regularities in revenue management by studying the new dimension of quality perception and generates new insights absent in the current literature. Our results offer insights on how retailers should adapt their markdown strategy in the presence of price-based quality perception. We develop a consumer model that incorporates quality perception and emotional loss when the expected markdown is too optimistic as compared to the actual markdown. We embed this model into the retailer's markdown optimization and examine the impact of consumers' behavioral factors on the retailer's optimal strategy. Finally, we design and conduct a consumer study to calibrate our model and validate the functional relationships among key factors. Consumers' quality perception increases with the products full price while it decreases with the (expected) markdown. We show that the retailer's optimal markdown is nonmonotone in these quality perception parameters. The nonmonotonicity is driven by the nontrivial tradeoff of trying to maintain a higher perceived quality by the consumers while controlling potential loss emotion that could arise if consumers observe a smaller-than-expected markdown, particularly when total market demand is not very large. Furthermore, we find that it is beneficial for the retailer to pre-announce and commit to a markdown strategy to prevent a mismatch between consumers' expectation and the actual markdown. This approach benefits the retailer by eliminating the negative effect on sales of the consumers' loss emotion due to an optimistic expectation. Ignoring these behavioral factors can substantially hurt the retailer's payoff. When inventory is tight, it is critical to correctly capture consumers quality perception (38% average loss in payoff if ignored). When instead inventory is sufficient, the retailer should be mindful of the potential emotional loss that its markdown could create among its consumers.

Gig economy platforms increasingly compete to source labor from common pools of multihoming workers, who dynamically allocate their services between competing platforms. Therefore, the question of how platforms can design pay and other levers to attract labor has gained significance. However, standard gig economy data are often incomplete from a choice modeling perspective, which impedes platforms' understanding of workers' multihoming preferences and choices. We study this problem using data shared by a major ride-hailing platform integrated with public data revealing the drivers' outside options in NYC. We structurally estimate a dynamic work-or-switch model using a novel combination of simulation and adversarial machine learning to overcome the empirical problem of contextually incomplete choice data. We find drivers to be significantly focused on short-horizon intraday payoffs while displaying significant heterogeneity in their costs of working. We offer prescriptions for platforms based on counterfactual analyses. We find that at a fixed labor cost, offering dynamic but guaranteed hourly pay rates can increase labor capacity over paying workers on a per-trip basis. A substantial segment of drivers are rarely observed to use pay-per-work platforms and would demand substantially increased compensation to accept that mode of pay. Next, we conduct a counterfactual policy exercise based on New York City's Driver Income Rules; this out-of-sample exercise validates recent empirical studies showing that higher posted pay worsens congestion while achieving muted gains in realized earnings. As alternative levers, platforms can reduce multihoming via offering streak bonuses, but impeding quits induces drivers to switch earlier to other platforms. Overall, managing multihoming requires aligning pay timing with workers' decision horizons, and instituting wage floors requires complementary measures to manage increased congestion.

Organizations increasingly deploy algorithmic tools to support complex operational decisions, raising a practical design question: how should these tools be built when designers care not only about immediate performance, but also about preserving and building human skill that remains valuable when advice is unavailable, imperfect, or requires genuine oversight? We study how the \emph{precision} of algorithmic advice shapes this trade-off. We develop a stylized model of advice-taking and learning. The model characterizes a reward-learning frontier: precise, action-level advice is easier to implement and improves payoffs while available through higher compliance, whereas broad, strategic advice requires interpretation, induces greater exploration, and generates knowledge that is portable, even when decision environments differ. We test the model's predictions in two online experiments in an electric-vehicle routing and charging task, representing typical characteristics of sequential decision tasks. Consistent with the theory, precise numerical advice delivers the strongest gains during the advice phase, whereas broader advice can yield more robust performance after advice is removed, specifically if the new environment differs substantially, but not completely. We use inverse reinforcement learning to recover interpretable latent objective components from action traces, distinguishing transient compliance from persistent internalization. Our results provide design guidance for advice systems that balance short-run operational efficiency with the development of long-run human capability. They also help validate inverse reinforcement learning as an effective tool for estimating human behaviors in complex sequential tasks.

Backward planning is an effective and efficient operational process when working towards a goal: work backward from the desired outcome to figure out the steps needed to accomplish them in the time allowed. While many organizations and workers might use it, backward planning is a universal practice among US K12 teachers. The emergence of generative AI has stimulated many conversations about its impact on teacher work, but it is still unclear if and how generative AI fits within the backward planning approach adopted by most every teacher. Given backward planning is standard workflow process in K12 education, we ask: How are teachers using generative AI to support their teaching work? Our methodology is a case study of 24 US public school teachers, sampled to vary by subject area and grade level, during the 2023-2024 school year. We conduct interviews, observations, and surveys at different points in time to understand their evolving generative AI use. In fall 2023, all teachers were novice users or had never tried generative AI. By spring 2024, the teachers separate into three distinct groups: (1) those who seek generative AI input (i.e., thoughts or ideas about learning plans) and output (i.e., quizzes, worksheets), (2) those who only seek generative AI outputs, and (3) those not using generative AI. The teachers in the first group-but not the second group-report productivity gains in terms of workload and work quality. Our findings have implications for understanding how to integrate generative AI into backward, goal-oriented workflows.

Gig platforms increasingly use algorithmic recommendations to steer workers toward tasks, but recommendations are only effective when workers both accept the suggested tasks and can execute them well. This paper studies how a platform should recommend a batch of tasks when workers are heterogeneous, improve with experience, and accept tasks strategically. We develop a stylized model in which each worker's latent service quality follows a learning curve and task acceptance follows a logit response to expected performance. The platform's per-round objective trades off the expected quality of the executed task against the risk of market failure when no worker accepts. We first show that the platform's expected reward is monotone submodular, which motivates a scalable greedy oracle benchmark. We then propose C-UCB-LR, a learning-aware combinatorial UCB policy that exploits submodularity while learning worker-specific quality parameters from noisy service outcomes, and we derive a learning-theoretic sublinear regret bound. To address operational reliability, we also introduce a safe variant that enforces a user-specified lower bound on the probability of at least one acceptance. Finally, we calibrate the model using large-scale operational data from a U.S. retail delivery platform and run counterfactual simulations. The calibrated environment highlights sizable gains from learning-aware recommendations relative to baselines that ignore worker learning or acceptance behavior.

Prior research on human-AI systems has focused on simple adopt/reject decisions on AI-generated tips, leaving sequential decision-making contexts less explored. In sequential tasks, tips suggest best practices that humans must operationalize across multiple decisions towards a problem-solving objective. In this paper, we explore human non-adoption of such tips in a virtual kitchen management task where disruptions necessitate a change of strategy. A qualitative analysis reveals diverse views of tips: as rules, directional principles, experimental options, or initially ignorable advice. Tips can still benefit those who reject them through creating focal points influencing worker sense-making. The challenge of operationalizing tips can lead to diverse barriers not just related to trust, but also to tip usability and environmental factors. A follow-up quantitative study confirmed three prominent barriers impacting participant intent to use tips. It also found that problem solving style, specifically an "Orientation to Change", may influence one's experience of barriers and tip adoption.