Cowles Foundation Discussion Papers

Industry–academia ML collaborations routinely fail to launch—not for scientific reasons, but because academics must publish while companies must protect models trained on proprietary data, and no standard contract framework resolves this tension. Because contracts are negotiated by legal departments alone, many apparent legal disputes are incentive misalignment problems that only scientists at the table can correctly diagnose. We propose PBOS (Protect-the-Business / Open-Source-the-Science), a community-adoptable contract template anchored to a single technically-grounded boundary: pre-training artifacts (architectures, training code, benchmarks, untrained weights) are open science; post-training artifacts (weights trained on proprietary data) are business IP. This boundary is technically meaningful, legally clean, and auditable—and could not have been drawn correctly without scientists at the negotiating table. We argue the ML community should adopt PBOS as its default contract for such collaborations.

This paper examines the theoretical and empirical consequences of rank-based reward systems in schools in which students’ performance and effort are evaluated relative to their peers. In such environments, classmates act simultaneously as competitors—due to rank-determined rewards—and as educators through peer learning and assistance. Using nationally representative panel survey data from U.S. high schools, combined with administrative information on the location assignments of new refugee student cohorts, we exploit variation in school competition policies and class ability compositions to identify empirically their dual effects on student effort and peer learning. We develop a theoretical tournament model with heterogeneous students who adjust their effort in response to the effort of similar peers and in which students learn from peers. The model predicts that when rewards depend on relative standing, adding higher-ability students to a cohort will reduce both incumbent academic effort and peer assistance, particularly in schools emphasizing rank-based awards, while adding lower-ability students has the opposite effects. Empirical tests of the model confirm these predictions. In schools with strong rank-based reward policies, the addition of stronger peers reduces high-performing incumbent students’ homework time and eliminates the positive spillovers from peer learning observed in less competitive settings. The adverse effects are concentrated among high-ability incumbents, while lower-ability students—who are less likely to win competitive awards—are largely unaffected. The results indicate that performance-based competition undermines cooperative peer learning and reduces student effort and overall academic performance, especially in institutions with high-ability students that explicitly emphasize relative ranking in determining academic recognition.

In GMM estimation it is well known that if the number of moment conditions grows with the sample size, GMM asymptotics differ from the standard case with moment size fixed as the sample size tends to infinity. The present work explores infinite dimensional GMM estimation under various conditions on the moment conditions and the weight matrix. Our approach employs a partial sum process formed by the moment conditions to represent high dimensional moments and an invariance principle to capture the infinite dimensional asymptotics as the moment size grows. Next, the GMM weight matrix is assumed to converge to one of two kernels at the limit: a continuous kernel or the Dirac delta function. Combining these different conditions enables development of a large sample theory for most efficient GMM estimation. The effects of permuting the moment conditions on GMM efficiency are also explored. The resulting theory is applied to weak instrumental variable estimation and the Angrist and Krueger (1991) data are re-analyzed in an empirical application of the new methods.

Solar Radiation Modification (SRM) has been proposed as a potential tool to limit increases in global or regional temperatures caused by anthropogenic greenhouse gas emissions. While previous research has extensively examined the climate system’s response to various SRM strategies, as well as their aggregate economic consequences, the regional distribution of economic impacts has received less attention. In this study, we use NorESM2–DIAM—an Earth System Model coupled to a high-resolution integrated assessment model—to assess the economic impacts, measured in GDP per capita, in an idealised SRM scenario where incoming solar radiation is reduced by 1%. Our results suggest that, relative to a baseline without SRM, most countries experience economic gains under SRM, with only a few countries facing negative impacts. Low-income countries tend to see the largest benefits, reducing global economic inequality relative to the baseline. However, reduced damages and lower inequality are accompanied by higher emissions under SRM, potentially leading to additional adverse effects not captured here. These findings highlight potential trade-offs between economic benefits, reduced inequality, and increased emissions relevant for SRM governance.

Signaling is wasteful. But how wasteful? We study the fraction of surplus dissipated in a separating equilibrium. For isoelastic environments, this waste ratio has a simple formula: β/(β + σ), where β is the benefit elasticity (reward to higher perception) and σ is the elasticity of higher types’ relative cost advantage. The ratio is constant across types and is independent of other parameters, including convexity of cost in the signal. We show that the directional effects of β and σ on waste extend to non-isoelastic environments.

AI/ML methods are increasingly used in economics to generate binary variables (or labels) via classification algorithms. When these generated variables are included as covariates in regressions, even small misclassification errors can induce large biases in OLS estimators and invalidate standard inference. We study whether the bootstrap can correct this bias and deliver valid inference. We first show that a seemingly natural fixed-label bootstrap, which generates data using estimated labels but relies on a corrupted version in estimation, is generally invalid unless a strong independence condition between the latent true labels and other covariates holds. We then propose a coupled-label bootstrap that jointly resamples the true and imputed labels, and show it is valid without this condition. Two finite-sample adjustments further improve coverage: a variance correction for uncertainty in estimated misclassification rates and a Hessian rotation for near-singular designs. We illustrate the methods in simulations and apply them to investigate the relationship between wages and remote work status.