Research

AI at Your Side: The Student's Guide to Smarter Learning

with Sebastian Galiani. Oxford University Press, 2026.

AI at Your Side is a guide for university students on how to use generative artificial intelligence in academic work without eroding the understanding that work is meant to build. Its central argument is that AI should augment human intelligence rather than replace it. The book develops this through the concept of a harness, a structured set of rules, routines, and constraints that keep an AI model focused, transparent about its uncertainty, and accountable to the student's reasoning. Successive chapters construct discipline-specific harnesses for ideation, research, writing, mathematics, coding, data analysis, economics, and finance. Grounded in learning science, it treats each task as an occasion to deepen understanding and retain intellectual ownership.

Measuring Efficiency and Equity Framing in Frontier Economics Research: LLM-Based Evidence from 1950 to 2021

with Sebastian Galiani, Ramiro H. Gálvez, and Franco Mettola La Giglia. Revise and Resubmit, The Economic Journal (Special Issue on AI Measurement in Applied Economics).

We measure how frontier economics research frames what is normatively at stake along the efficiency–equity dimension. We classify 27,464 top-five articles from 1950 to 2021 with a validated, rubric-guided LLM pipeline. Efficiency-oriented framing rises through the late 1980s, then declines as equity-related framing expands after 1990, especially in applied work and policy evaluation. By 2021, papers with an equity component are about as common as purely efficiency-framed ones. Equity enters alongside efficiency rather than displacing it, yet formal cost–benefit analysis remains rare in policy evaluations. Transmittal letters in the Economic Report of the President show the same post-1990 shift.

PDF

Deep Research on a Loop: Agent-Driven Construction of Auditable Structured Datasets

with Santiago Afonso, Sebastian Galiani, and Ramiro H. Gálvez.

Constructing datasets from primary sources remains a bottleneck in empirical research. Deep-research agents can search the web, read documents, and synthesize evidence, but their default output is a narrative memo rather than a comparable, auditable dataset. We introduce Deep Research on a Loop (DRIL), a protocol-driven methodology that turns open-ended agent research into structured data. A design-stage agent helps specify a research instrument, mapped unit space, and evidence policy. These choices are frozen into a protocol that implementation-stage agents apply across units under fixed coding and citation rules. Recorded values must be supported by verbatim evidence, with uncertainty and gaps documented explicitly. We deploy DRIL on a 2025 partial update of a public tax-expenditure database covering 86 analytic jurisdictions, about 40% of the database's universe. The run returns 1,040 sources, 1,113 evidence records, 103 quantitative estimates, and complete qualitative coverage on 22 fields, at about USD 22.50 in equivalent API cost. Quantitative gaps decompose into absent public reporting (71%), document-parsing limits (16%), and residual design-sensitive cases. The results suggest that partial automation can materially reduce the cost of empirical dataset construction when the process is standardized, evidence-bound, and auditable.

PDF

AI Agents and Prompt Engineering in Econometric Coding

with Sebastian Galiani and Federico A. López.

We study how large language models write code for econometric analysis. We compare three dimensions of AI-assisted coding: statistical software (Stata, R, or Python), prompting (zero-shot versus few-shot), and the degree of agency, from a chatbot that writes a single script to an agent that executes and revises its own code. On a benchmark of applied econometric and statistical tasks, moving from the chatbot to the constrained agent raises task success from 74 to 96 percent, at about eight additional cents per run. For Claude Sonnet 4.6 and GPT-5.4 through Codex, few-shot prompting improves the chatbot far more than the constrained agent, indicating that prompting and agency act as substitutes. For these models, differences across statistical software are sizeable under the chatbot but largely disappear under the constrained agent.

PDF

How Institutions Reweight Evidence: Climate Science from the IPCC to the Press

with Sebastian Galiani and Franco Mettola La Giglia.

A climate finding grows more severe as it is summarized for the public, most of it added by the press, not the IPCC. Using three language models, we score this displacement at the claim level across about 1,000 internal and 25,305 media pairs. Both steps lean toward the severe end while staying within accepted ranges. The press step, near +0.17 on a [−1,+1] scale, runs three times the internal one. Left- and right-leaning outlets converge rather than diverge, which fits shared institutional incentives better than partisan demand. The pattern is the systematic outcome of incentivized summarization, not a verdict on the science.

PDF

The Empathy Channel in Fertility

with Sebastian Galiani. Submitted.

Being around children makes people want children of their own, a response documented in human biology. We build this empathy channel into a quantity–quality overlapping-generations model, in which each birth generates a durable social externality: it adds to the surrounding stock of children, and that stock lifts the desire of others for as long as the children remain present. The externality creates a social multiplier, since a fall in births means less exposure, which depresses desired fertility further. When the response is sharp enough, the economy admits a low-fertility trap. Because no household internalizes the effect of its own children on others, the decentralized economy is sub-fertile, and the remedy is a pro-natal subsidy set by the capitalized social value of a child and tilted toward high-exposure groups. An illustrative calibration sizes the channel and the subsidy. The result is a microfoundation for the social interactions long studied in the demography of fertility, and a rationale for pro-natal policy that runs through the desire for children rather than the income and price mechanisms of the standard model.

PDF

Composition Beats Collapse: Insights from the Bisin–Verdier Model on Endogenous Fertility Reversal

with Sebastian Galiani. Submitted.

Fertility rates have fallen below replacement in most countries, fueling predictions of demographic collapse. We show these forecasts overlook a crucial fact: societies are not homogeneous. Using the Bisin–Verdier model of cultural transmission with endogenous fertility and direct socialization, calibrated to U.S. and global data, we find that high-fertility, high-retention groups persist, gain share, and lead the total population to grow. Even if fertility remains below replacement in every country, extinction is unlikely. Simulations imply continued growth with pronounced compositional change, driven especially by religious communities with high fertility. In our ten-generation world calibration, Muslims become the largest tradition.

PDF

Easy Like Sunday Morning: Organizational Design and Physician Agency in Birth Timing

with Jimena S. Ferraro and Franco M. Vazquez. Submitted.

Cesarean rates differ sharply between private and public hospitals in Buenos Aires, even though payments do not vary with delivery mode in either sector. We argue this gap reflects organizational design rather than procedure prices. In the Buenos Aires metropolitan area, individualized private and team-based public maternity care coexist within the same labor market. Using ex ante preferences and medical records for low-risk first-time mothers, we document a preference margin and a timing margin. Stated preferences for cesarean predict scheduled procedures strongly in private hospitals and not at all in public ones. Elective cesareans in private hospitals cluster on boundary working days (the Monday or Friday adjacent to a weekend or holiday), and the intrapartum cesarean rate is about 13 percentage points higher on working than nonworking days in private hospitals, but indistinguishable from zero in public ones, a gradient that reflects selection induced by the timing margin. A simple model rationalizes both margins through the opportunity cost of physician time, internalized only when physicians are residual claimants of their own schedule. Organizational design, not procedure prices, is the primary lever behind the discretionary component of cesarean use.

PDF

Minimum Slaughter Weight Regulation in Cattle Markets: Evidence from Argentina

with Federico Sturzenegger and Franco M. Vazquez. Submitted. M.A. thesis.

A minimum slaughter weight forbids slaughtering animals below a legal threshold, a rule meant to raise meat production. We evaluate it in a Faustmann competitive equilibrium with endogenous breeding, a binding land constraint, and two technologies, calibrated to Argentina. With a single technology, discounting makes producers slaughter below the supply-maximizing weight, so a moderate floor raises supply. The supply gain is small, and once the transition through the cattle cycle is priced, the floor delivers no welfare gain. With pasture and feedlot sharing the market, equilibrium slaughter weights are 408 and 346 kg. A uniform 420–450 kg floor can cut welfare by 13.2–18.8% of revenue in present value where cattle is the dominant land use, because feedlot output losses dominate pasture gains. Argentina's floor did not bind the representative producer, so the 2026 repeal carried effectively zero welfare cost and a tighter one would have been harmful. As middle-income agriculture turns dual, a capital-intensive segment beside a land-intensive one, a uniform instrument written for a single technology misfires once that structure takes hold.

PDF