6월 1주차

Organizations increasingly use machine learning (ML) and artificial intelligence (AI), including large language models, to generate variables for regression models that inform business and policy decisions. For example, practitioners may use AI to predict review sentiment, ad aesthetics, or emotional expressions, and then estimate their causal effects on outcomes such as sales or engagement. However, because AI predictions are imperfect, directly using these AI-generated variables as regressors introduces measurement error that can systematically bias causal estimates, potentially leading to over- or underinvestment in business strategies. We develop new estimators that correct this bias in partially linear regression models, which are widely deployed in experimental systems at major platforms, including Tencent, Amazon AWS, and Microsoft. Our approach requires only a small human-annotated subsample alongside the large AI-labeled data set to achieve unbiased and efficient estimation. We demonstrate that our methods work with both traditional ML algorithms and LLM-based predictions. Our framework can be directly integrated into existing analytics and experimental systems, enabling practitioners to leverage the scalability of AI-generated data while maintaining reliable causal conclusions. This work also has implications for AI fairness, as our approach can help correct biases from any source in AI predictions.

Digital platforms often share publicly visible behavioral data, such as ratings and reviews, to improve service quality. Yet, these seemingly innocuous data can also reveal hidden social ties among users. We develop a two-stage framework that first diagnoses this leakage risk by inferring latent social interactions from behavioral data and then mitigates the risk while preserving data utility. We characterize when hidden ties are identifiable from observed actions and show, using public Yelp data from Louisiana and Pennsylvania, that an attacker can recover about half of true social ties at a 10% false-positive rate and more than 60% at a 20% false-positive rate. These inferred ties can materially increase cyber risk: when used for spear phishing, the estimated return on attack rises from 109% for a campaign with 500 impersonation attempts to 1,098% for one with 10,000 attempts. To reduce this leakage, we propose two perturbation mechanisms with formal differential privacy guarantees on the released representations: one adds Gaussian noise directly to the released action matrix, and the other adds Laplace noise to the learned representation before resynthesizing the released data. Both mechanisms reduce the link-inference accuracy and substantially lower estimated attacker returns; under our main protection settings, returns become negative for smaller campaigns and remain much lower at larger scales. Overall, the paper provides a practical framework for diagnosing social interaction leakage and evaluating the privacy-utility trade-off when platforms share behavioral data. History: Peiyu Chen, Senior Editor; Heng Xu, Associate Editor. Funding: Y. Leng is supported by the U.S. National Science Foundation (NSF) [Grant IIS-2153468]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/isre.2024.1469 .

This study examines how similarity, social identity, and platform design jointly shape user experiences in two-sided digital platforms. Drawing on theories of power asymmetry, social identity, and digital platforms, we analyze more than 500,000 reviews on Airbnb to evaluate how shared marginalized identities between guests and hosts influence perceived experiences. Contrary to expectations, similarity alone does not consistently yield more positive outcomes; however, platform features that enable power sharing, such as flexible cancellation policies and instant booking, significantly increase the likelihood of positive reviews among marginalized groups. These findings contain practical and policy implications. For platform designers and policymakers, the results highlight the opportunity of structural features that redistribute control and flexibility within transactions to foster inclusive and positive experiences. For platform owners and regulators, these results suggest a shift from representation-focused interventions toward feature-based governance that embeds power sharing, and potentially other shared values, into platform architecture. More broadly, the study underscores the importance of aligning technological features with the values and lived experiences of marginalized users, offering a scalable pathway to promote both inclusion and marketplace performance.

There are two different types of brand loyalty, and they work in different ways. Our study shows that loyalty created by switching costs, such as proprietary standards, data frictions, or restrictive interfaces, has very different market consequences from loyalty created by preference shifts, where repeated use makes customers genuinely prefer a product’s design or features. Although both increase retention, switching costs protect incumbents, soften competition, raise prices, and reduce consumer and social welfare. Preference shifts can have the opposite effect; they can lead to a reduction in differentiation between rival products, intensify competition, lower prices, and under some conditions, improve welfare by better aligning products with what customers want. The managerial message is clear; investments in user experience, reliability, and product design are not strategically equivalent to investments in lock-in. The policy message is equally important; regulators should not treat all customer stickiness as harmful. Remedies, such as interoperability, data portability, or limits on proprietary standards, are best targeted at switching barriers, not loyalty created by genuine product improvement.