DISC Research Seminar Series

We study how competing content platforms (e.g., TikTok, YouTube, Facebook, and X) moderate user-generated content while monetizing attention through advertising. Section 230 of the U.S. Communications Decency Act, enacted in 1996, delegates moderation to market forces on the premise that competition will discipline it. Since then, consumption-side competition has intensified and cross-posting tools have strengthened generation-side complementarity. We examine these shifts in a duopoly where sensitive and insensitive users jointly consume and generate content and platforms simultaneously choose moderation and advertising load. A platform's moderation unambiguously cleans its own content, but its effect on the competitor depends on the relative strength of consumption- and generation-side interactions. When platforms are strong generation-side complements-arguably the case today-moderation is a public good: it cleans the competitor and raises its demand, diluting private returns platforms under-moderate relative to centralized and welfare-maximizing benchmarks. Under weaker complementarity, moderation is a quasi-public good: it still cleans the competitor but steals its users on net, yet platforms can still under-moderate. When platforms are generation-side substitutes, moderation is business-stealing: it displaces harmful content onto the competitor and steals its demand, and platforms over-moderate. These spillovers can leave harmful-content generators better off and those moderation is meant to protect worse off. Proposed liability amendments to Section 230 (e.g., the SAFE TECH and EARN IT Acts) may improve welfare only up to a point, beyond which they worsen it. Under sufficiently strong generation-side complementarity, coordination on moderation alone, with advertising competition intact, is collectively profitable and raises moderation on both platforms.

To benefit from technology products, customers must learn to use them. The value of new technologies is often uncertain. So, customers must invest in learning while facing uncertain returns. These challenges hurt vendors' demand and profit. Software vendors can tackle this through three strategies: enhancing customers' learning rates, lowering the learning barrier, and reducing value uncertainty. These strategies are frequently deployed together, making it critical for vendors to understand how they interact to optimize benefits. Are they substitutes such that implementing one reduces the value of the others? Or are there synergies within a subset of strategies that should be deployed together? We investigate these strategic interactions and how the optimal combination of strategies may vary depending on whether the software is sold via perpetual or subscription-based licensing. We employ a Bayesian dynamic model incorporating two learning characteristics: customers’ learning progresses according to a learning curve, and the learning curve’s shape is unknown to customers but can be learned through experience. We show that without uncertainty, enhancing customers' learning rates and lowering the learning barrier are substitutes under both licensing schemes. Both strategies ease learning. An improvement because of one reduces the benefit from the other. In contrast, with uncertainty, the strategic interactions depend on the licensing scheme. Under perpetual licensing, enhancing customers' learning rates and lowering the learning barrier are complementary when customers’ prior beliefs are pessimistic, and learning is difficult, but substitutive otherwise. Reducing value uncertainty complements the other two strategies and should be deployed jointly with them. Uncertainty leads customers to consider an option value, which reshapes how the strategies interact. Conversely, under subscription licensing, reducing value uncertainty generally complements one of the other two strategies but substitutes for the second. Which strategy it complements varies, depending on parameter values. The recurring subscription price introduces a new mechanism that explains the different interactions across licensing schemes.

(updated) This seminar will cover two critical frontiers related to AI: the internal world of academic research production and the external world of public policy. First, "Fighting Fire with Fire: Infusing AI into Peer Review to Sustain Quality Scholarship," will address a "publication paradox" in academia. Increased submission volume due to generative AI will stress human peer-review capacity. We propose a hybrid editorial workflow where journal-specific LLMs act as first-line reviewers, allowing authors to iterate before human experts step in. This approach moves beyond the unvetted, isolated use of AI currently occurring "in the wild" toward a deliberate, institutionally governed framework that preserves epistemic standards. Second, "Exploratory Analysis of AI Legislation in the United States," will examine how society "puts the brakes" on AI technologies. We use text analysis and LLMs to categorize over 1,500 bills across 51 jurisdictions, via a custom seven-element taxonomy rooted in regulatory rationale—such as product safety and market structure. We then analyze the resulting curated corpus to provide empirical insights into state-level policy trends, political influences on legislative outcomes, and the evolving sophistication of AI governance.

Problem definition: This study explores the impact of initial and in-process waiting times in asynchronous telemedicine on subsequent online and offline visits, and the moderating effects of the consultation fee. Methodology/results: We focus on three measures of waiting time in asynchronous telemedicine: initial waiting time (the duration from a patient’s initial request to admission into an online consultation session), average in-process waiting time (the average time between a patient’s question and the doctor’s response during the session), and the variability of in-process waiting times, which together capture both the initial and ongoing responsiveness to patients during asynchronous care delivery. We use 42,111 patients’ online and offline consultation records from a primarily text-based, asynchronous telemedicine platform affiliated with a top-ranked hospital system (February 2021-April 2024). Our results show that patients with a longer (above median) average in-process waiting time (≥0.75 hours) have 14.53%, 16.47%, and 13.41% lower odds for subsequent all visits, online visits, and outpatient visits in the next 30 days, respectively. Patients with a higher (above median) variability of in-process waiting times (≥0.40 hours) have 9.70% and 12.72% lower odds of subsequent all visits and offline outpatient visits in the next 30 days, respectively. Surprisingly, the initial waiting time shows no significant effect. Results remain consistent when considering whether the subsequent visits are with the same doctor. Finally, the consultation fee negatively moderates the relationship between in-process waiting time and subsequent visits. Managerial implications: Average in-process waiting time in asynchronous telemedicine has the most significant impact on both subsequent online and offline patient visits among the three waiting time metrics. The findings highlight reducing in-process waiting in asynchronous telemedicine as a viable means of enhancing patient engagement and ensuring continuity of care across channels.

Firms increasingly use Artificial Intelligence (AI) enabled forecasting engines ("prediction machines") to augment their managers' own forecasting capabilities and thus improve sales-and-operations planning outcomes. Deployment of a prediction machine may cause an unintended reduction in a manager's own forecasting effort which in turn diminishes the value of machine adoption. We model a firm facing uncertain demand that delegates a procurement quantity decision to a human manager who can exert effort to generate a demand prediction. The firm deploys a machine that provides the manager with a demand-prediction signal. We establish the conditions under which managerial effort reduction occurs and thus reduces the machine's potential value. Adopting a Bayesian persuasion approach, we show that partially disclosing the machine's prediction, either downplaying high predictions or exaggerating low predictions, can be optimal, depending on the product's cost-to-revenue ratio. A strategy of minimal obfuscation (to achieve effort) is optimal if the machine is more accurate than the human however, maximal obfuscation (while maintaining effort) can be optimal if the human is more accurate. Our results imply that the firm may be better off tuning a machine to be less informative than its maximum capability.