Research

1. Saharsh Agarwal, Ronak Vishnoi. Creating Safe Digital Spaces.

(Book chapter) Research Handbook on Platform Ecosystems. Edited by Peng Huang and Sunil Mithas, Edward Elgar Publishing

Digital spaces created by modern day platforms have come to shape the way we connect today and live as a global community. However, with their increasing roles in our lives, it is becoming more and more evident that they also come with a new set of risks of harm. In this chapter, we look at how these harms can be mitigated to create digitally safe spaces. We argue that platforms need to make changes at the institutional design level rather than just police content. We first propose a framework to categorize digital harm, based on its nature and scope. Then, we develop a multi-layered sociotechnical model to analyze both the sources of harm and the points of intervention (at the micro, meso and macro level). This is followed by a discussion on key challenges and trade-offs that the platform leaders and policymakers face. Overall, the chapter presents a new framework for understanding how digital safety can be embedded into the very foundation of platforms-shifting the focus from reactive responses to proactive design. 

4. Saharsh Agarwal, Ananya Sen. Google AI Overviews and Publisher Traffic: Evidence from a Field Experiment

Online search engines depend on external content to respond to user queries, while content providers rely on search engines for visibility and user acquisition. The integration of GenAI-based summaries into search interfaces (eg, Google’s AI Overviews (AIOs)) may reduce users’ incentives to visit downstream publishers. Despite growing regulatory and industry attention, causal evidence on the effects of AI-generated summaries on search behavior remains limited. This paper provides causal evidence on the impact of AIOs on user behavior and downstream engagement using a field experiment that randomly assigns users to either the standard Google Search experience with AIOs or a version where AIOs are removed. We develop a custom Chrome browser extension that passively tracks detailed browsing behavior while also enabling real-time modifications to the search engine results page (SERP), allowing us to precisely manipulate AIO exposure in a natural browsing environment. When they appear, AIOs reduce outbound organic clicks by 38% and increase the probability of a zero-click search by 33%. We find no effects on sponsored clicks or overall search frequency, implying that AIOs reduce total external traffic by lowering organic clicks per search. Heterogeneity analyses show that effects are concentrated when AIOs appear at the top of the page and among the highest-ranked organic results. Using survey evidence, we find that despite these behavioral effects, removing AIOs does not affect users’ overall search experience, including perceived ease of finding information and the quality of results. Exploratory evidence from a third AI Mode condition, where users are redirected to a Gemini-powered conversational AI interface, suggests even stronger reductions in engagement with downstream publishers. Overall, the findings suggest that AIOs divert traffic away from publishers without delivering measurable improvements in user experience, raising important implications for competition policy, platform power, and publisher sustainability.

3. Saharsh Agarwal, Uttara Ananthakrishnan. Impact of Short-Form Videos on Screentime and Sleep: Evidence from TikTok

Minor revision, Management Science

Short-form video platforms have become central to debates about digital overuse and sleep disruption, yet causal evidence on their behavioral impacts remains limited. We study the effects of adopting short-form video in the context of TikTok, which pioneered the format and accounts for a large share of global short-form video consumption. Using granular smartphone touchstream data from a large opt-in panel of U.S. Android users, we track app usage, screen time, and inferred sleep behavior before and after TikTok adoption. We employ a Difference-in-Differences (DiD) framework leveraging a high-frequency app usage dataset from nearly 7,000 TikTok adopters. Our findings reveal a significant increase in overall screen time---averaging 221 minutes per week for heavy TikTok users---following TikTok adoption, alongside notable declines in usage of other social media and gaming apps, suggesting a strong substitution effect. The screen time effects are highest during late night hours. We also observe shifts in sleep patterns, including delayed bedtimes and a moderate reduction in sleep duration, among high-intensity users. We observe no effects for low-intensity users, who engage minimally with TikTok post-adoption. Our findings show that short-form video adoption expands total digital attention rather than merely reallocating it, with important implications for platform design, regulation, and digital well-being. As a side result, our analysis provides clear evidence of substitutability between TikTok and traditional social media platforms, which has immediate implications for ongoing antitrust proceedings. 

2. Saharsh Agarwal (primary author), Uttara Ananthakrishnan, Catherine Tucker. Infrastructural Gatekeeping and its Spillovers

Under third round review, Information Systems Research

 Prior research on online governance has largely focused on fine-grained, within-platform content moderation, where platform operators and community members remove individual posts or users. Infrastructural Gatekeeping is a distinct and emerging class of intervention in which third-party infrastructure providers (e.g., cloud hosts and app stores) remove entire platforms from the digital ecosystem due to content related reasons. Unlike within-platform content moderation, infrastructural gatekeeping eliminates entire communicative spaces, raising new questions about how affected communities reorganize across the broader digital ecosystem. We examine the Parler takedown in January 2021, when multiple infrastructure providers terminated the platform’s web and app hosting. Using large-scale mobile touchstream data from over 46,000 Parler users, we analyze how individuals redistributed their activity across alternative platforms. We introduce a governance framework to classify platforms based on permissiveness (lenient/strict) and observability (open/closed) and show that Infrastructural Gatekeeping produces substantial spillovers across all quadrants of this space. Users increase activity not only on ideologically or functionally similar lenient-open platforms but also on lenient-closed and strict-open alternatives. Migration responses differ sharply by prior engagement: heavy users rapidly consolidate into closed, less observable environments, while light users exhibit broader and more transient experimentation. Infrastructural Gatekeeping interventions also heighten exposure to low-quality and misleading content for the most engaged users. These findings show that infrastructural gatekeeping is a qualitatively distinct governance mechanism with ecosystem-level consequences, potentially shifting harmful activity into less observable environments rather than eliminating it.

1. Li Ding, Saharsh Agarwal, Vivek Choudhary. The Impact of Payment on Delivery on Gig Drivers' Behavior in Developing Economies

Problem definition: In developing economies, gig platforms are widely viewed as engines of income generation, yet a prevalent and central operational feature of these markets, Payment-on-Delivery (POD), has received little attention. Under POD, drivers collect payments from customers and temporarily retain the full order value before settling with the platform. This provides workers with immediate liquidity but also creates frictions, including cash-handling risks and reconciliation responsibilities. Despite its scale and operational importance, we lack empirical evidence on POD’s impact on gig drivers, the primary operational asset of these platforms. We investigate how POD shapes drivers’ decisions about when to stop working, whether to return the next day, and whether to remain active in the longer run. Methodology/Results: We analyze transaction-level data from a leading food delivery platform in India, including 4,776 drivers and ~2.8 million orders over six months. We exploit quasi-random POD order allocation to causally identify how POD affects drivers’ labor supply decisions. We find that receiving a POD order increases the likelihood that a driver ends the shift by four percent, consistent with cash-targeting behavior in which workers treat collected cash as salient and immediately usable. The effect becomes stronger when the opportunity cost of stopping is low (e.g., off-peak hours) and when liquidity needs are elevated (e.g., month-end). Although POD induces shift-ending, it increases next-day participation and lowers attrition, resulting in a more stable labor supply over time. Managerial implications: POD is not only as a mechanism for expanding consumer access but also as a lever for improving driver retention. By providing immediate liquidity, POD helps stabilize participation, reduce turnover, and smoothen labor supply in settings where workers face significant financial constraints. Platforms that recognize and manage this liquidity channel can strengthen operational reliability in highly liquidity-dependent developing country markets.