Research

A central prediction of human capital theory is that individuals with the highest marginal returns to education should be the most likely to invest in it. We document a stark violation of this prediction in financial literacy education: individuals with the lowest baseline knowledge—and thus the highest potential returns—have the lowest demand for education. Using motivating evidence from a large university personal finance course and a controlled online experiment (n=1,047), we document that demand for financial education is positively correlated with baseline financial literacy, potentially exacerbating pre-existing inequalities. In the lab setting, we identify the mechanism driving this inefficiency. Low-ability individuals systematically miscalibrate their beliefs, overestimating their baseline knowledge while underestimating the returns to education. To test whether reducing perceived cognitive costs can correct this distortion, we vary the format of the education material, comparing a traditional lecture-style education to video- and AI-based alternatives. While the AI and video versions significantly increase aggregate demand, they fail to correct the adverse selection on ability. Together, the results suggest that supply-side interventions alone may be insufficient to close financial literacy gaps without addressing underlying belief distortions.

Profit-maximizing firms often find ways to benefit from inattentive consumers. An example is in insurance markets, where insurers facing limited consumer mobility have an incentive to strategically cycle their risk pools by using low advertised rates to attract new business while raising rates on existing consumers. When a cohort of healthy new enrollees ages and sickens, the insurer closes the block to new entrants, continues extracting above-cost premiums from the trapped cohort, and opens a fresh block to attract the next healthy cohort. We call this practice “deadpooling.” This paper formalizes the mechanism as a general property of open price insurance contracts with imperfect consumer switching, then uses the Medigap market, a setting with state-level variation in switching rules, to provide the first empirical documentation of deadpooling at scale. We find that 16.7% of Medigap life-years are spent in a plan that deadpools, resulting in an estimated $293.8 million annual transfer from seniors to insurer profits, and $5.6 billion over our sample period. Our mechanism test distinguishes strategic markup from passive adverse selection: within states with no switching rights, deadpooled blocks show premium growth exceeding claims growth, while non-deadpooling closures do not. A calibrated micro-simulation evaluates policies aimed at curbing this practice, finding that many demand-side interventions are not effective without a highly attentive consumer base.

Despite dramatically expanded access to selective U.S. colleges, first-generation students persistently trail continuing-generation peers in GPA, internship attainment, and early-career outcomes. We identify a key mechanism: the hidden curriculum—unwritten strategies like cold-emailing alumni or strategically engaging faculty, essential for success yet unknown and costly to discover without guidance. Leveraging survey and administrative data from 100,000+ undergraduates across 20 public universities, we document stark disparities: first-generation students invest 14–26% less in these high-return hidden actions while over-investing in formal tasks. Standard explanations—income, ability, or preferences—do not fully explain these gaps. Through a field experiment at UC Berkeley, we isolate causal channels by randomizing information on action availability (awareness) versus returns (beliefs): awareness treatments close the 30% baseline gap almost entirely. Finally, we develop an AI college advisor to expose underlying search frictions in an online experiment; first-generation students allocate just 11% (versus 16%) of queries to hidden topics and follow up about 48% less on hidden curriculum nudges. However, an “active” AI that increases awareness narrows these search gaps and follow-up behaviors. By formalizing the hidden curriculum as dual informational frictions, we demonstrate that overcoming these invisible barriers requires more than equal access.

This project investigates a novel friction in household finance: the absence of “self-knowledge” or self-awareness of financial behavior. While existing fintech budgeting apps typically target information gaps through dashboards or self-control failures through commitment devices, we explore whether AI-mediated self-reflection can causally reduce discretionary spending by encouraging self-reflection and helping users construct persistent narratives around their financial lives. In partnership with an AI-backed personal finance application, we utilize an observational event study using high-frequency transaction data alongside a randomized controlled trial. Our experimental design isolates the mechanism of self-knowledge by comparing a self-reflection, narrative-framed AI intervention against both an active, information-only AI control and a delayed-access group. By analyzing chat transcripts and longitudinal surveys, the study seeks to determine if self-reflection and increasing self-knowledge can improve saving and budgeting behavior and financial well-being more than traditional information- or commitment-based approaches.

Individuals frequently make costly financial mistakes while expressing high confidence in their financial knowledge. Does overconfidence inhibit information acquisition and thereby lead to suboptimal financial decisions? We study this question in a field experiment. First, we document that participants are substantially overconfident about their knowledge related to debt. Then, we experimentally reduce confidence in a randomly selected group using an Illusion of Explanatory Depth (IOED) intervention that reveals gaps in their knowledge, while a control group receives a placebo exercise that does not affect confidence. After the intervention, all participants are offered access to an educational resource. We test whether lowering confidence increases information take-up and leads to improvements in financial outcomes.

Technological progress has historically favored high-skill labor, widening income inequality. Generative AI offers a potential counter-force: it can de-skill tasks like writing cover letters or optimizing resumes, theoretically benefiting low-skill job seekers the most. However, the realized impact depends on adoption. We combine observational data with a field experiment among active job seekers to investigate whether those who need AI assistance the most are the ones who use it. Similar to other surveys on AI usage, we expect to find an “adoption mismatch” where low-income and less-educated individuals—who face higher turnover and structural vulnerability—are significantly less likely to adopt AI job search tools than their high-skill counterparts. To separate selection from treatment effects, we plan to employ a two-stage design: we first offer an AI tool to all participants, then randomize financial incentives to induce uptake among those who initially declined. This allows us to estimate treatment effects for “reluctant adopters.” We will then examine whether exposure to AI job search tools changes beliefs about the usefulness of AI. We also test whether AI assistance reduces unemployment duration and increases wage offers and compare differences between the “voluntary adopters” group and the “forced adoption” group. Finally, we analyze the interaction logs of the AI job search tool to examine the extent to which high and low-skill job seekers override high-quality algorithmic suggestions, reverting to suboptimal text. Our results will highlight how differential selection may lead AI tools to widen labor market inequalities rather than narrow them.

Increasing transparency is one of the first and most common recommendations from international financial institutions to policymakers in countries that experience economic crises. Despite the widespread prescription of this elixir, disagreements persist about its efficacy during crises. Much of the existing literature suggests that increasing transparency decreases information asymmetries, increases policy predictability and the credibility of policy commitments, improves the effectiveness of monetary policy, and bolsters public confidence. Each of these effects could plausibly shorten the duration of economic crises. Critics counter, however, that effects of transparency are ambiguous and may increase policy uncertainty, raise volatility, increase the prospect of collectively self-destructive behaviors, and decrease the effectiveness of monetary policy—effects that could prolong crises. These debates persist in part because related empirical research tends to focus primarily on the transparency of central banks and its impact on market expectations regarding short-term interest rates without considering the transparency of national governments and how the availability of credible data about the national economy from sources other than the central bank affects public and market expectations. We argue that greater transparency of national governments—often inferred from, yet independent of, the transparency of central banks—will decrease the duration of inflation and currency crises by providing information about existing economic conditions, increasing the predictability and credibility of national economic policy, and increasing confidence in the efficacy of policy choices by demonstrating the degree to which the policy positions of national politicians and central bankers align. We operationalize government transparency in terms of the government dissemination of credible macroeconomic information using the Hollyer, Rosendorff, and Vreeland (HRV) index. Our analyses of 125 countries from 1980 through 2010 indicate that higher levels of government transparency are strongly correlated with shorter durations of inflation and currency crises and that the level of transparency is negatively correlated with the severity or size of inflation crises.