Economics, Decision-Making & AI
Intelligent agents — people, firms, institutions, and increasingly AI systems — make decisions using incomplete information and imperfect representations of the world. At the same time, the institutions around them try to shape those decisions: through the information they provide, the incentives they set, the monitoring they impose, and the constraints they enforce. My research asks how these two forces interact. I come to this question as an economist, using tools from structural econometrics, causal inference, and economic theory, and increasingly from machine learning. The same question that has motivated my work on consumer choice and healthcare markets — what does a decision-maker actually see and use, as opposed to what it could in principle optimize — turns out to be just as central to understanding AI systems.
1. Human and machine decision-making
A decision-maker rarely optimizes over the full choice set or the full information available to it. In my work on hospital choice, patients choose from a limited consideration set rather than every available option, shaped by what they have seen, heard, or been told. Modeling this limited consideration, and estimating how it responds to information and incentives, has been a central part of my empirical work on healthcare markets.
The same question arises, in a different form, for AI systems. A language model may encode information relevant to a decision — a comparison, a preference, a risk estimate — without that information actually driving its output. Interpretability research often asks whether a variable is represented inside a model; the question I am most interested in is whether that representation is behaviorally relevant, in the same sense that a piece of information is only economically relevant to a consumer if it changes what they choose. Bringing consideration-set and limited-attention models from economics into contact with interpretability tools is one direction I am pursuing.
2. Incentives, monitoring, and AI agents
Once a decision-maker's actions cannot be observed directly, whoever is training or evaluating it faces a classic principal-agent problem: which observable signals actually tell you something about the hidden behavior you care about, and which do not. This question has a long history in information economics — Holmstrom's informativeness principle, multitask contracting, and the broader literature on mechanism design under limited observability.
I think this literature has a lot to offer AI evaluation and monitoring, where similar structure recurs: models are trained and graded on observable outcomes and traces, evaluators face the same question about which signals are informative, and, under enough optimization pressure, models can learn to make their behavior look better on the measured signal without actually improving on the dimension the signal was meant to proxy for. I work through some of these connections informally in a personal essay series in my journal, Alignment × Econ — for example, using the informativeness principle to think about outcome versus process supervision, and using ideas from multitasking contracts to think about reward hacking and Goodhart's law. This is a research direction rather than a settled result: the aim is to see how far these classic contracting ideas travel, and where they break down, when the agent is a model rather than a person.
3. Markets, allocation, and distribution shift
A second strand of my work studies how markets and institutions allocate scarce resources — hospital capacity, subsidized health products, spectrum, spots in a matching market — and how the design of prices, subsidies, and rules shapes who gets what. This has mostly been empirical work in healthcare and Indian development contexts, using causal inference and structural methods to evaluate specific policies.
I am increasingly interested in optimal transport as a tool for this kind of question, because it gives a natural language for describing how a distribution of outcomes shifts in response to a policy, a price, or a change in the environment. That connects naturally to allocation problems in more classical economic settings, but also to AI settings, where systems increasingly participate in markets and decision pipelines and where the distributions they act on shift over time. Matching, allocation, and distribution shift are, at some level of abstraction, the same problem.
A common thread
The common thread across these three directions is a distinction I keep returning to: the objective a decision-maker could in principle optimize, the information or representation it actually uses, the incentives shaping its behavior, and the allocation or outcome that results. Economics has spent decades building tools for reasoning about the gaps between these — limited information, moral hazard, mechanism design, market design. My work asks how far those tools extend to a world where the decision-maker is sometimes a person, sometimes a firm or institution, and increasingly an AI system, and where the gaps between what could be optimized and what is actually used, observed, or rewarded matter more than ever.