Research

As social media usage reaches record highs, personalization algorithms risk radicalizing users by reinforcing existing beliefs. However, evidence on how algorithms and user behavior jointly shape harmful online engagement is limited. In this paper, I conduct an individually randomized experiment with 8 million users of a prominent TikTok-like platform in India, replacing the feed-ranking algorithm with random content delivery. I focus on hateful content targeting minority groups, given its prominence on Indian social media and establish a trade-off: random post recommendation lowers exposure to anti-minority ("toxic") content by 27%, but at a substantial cost to the platform as overall platform usage falls by 35%. Strikingly, treated users share a larger proportion of the toxic posts they view, mitigating the decline in the number of toxic posts shared from the platform. Users with a higher interest in toxic content at baseline drive this result as they seek out posts the algorithm does not show them. I rationalize these findings with a model of a revenue-driven algorithm that faces heterogeneous users choosing which posts to consume. Counterfactual simulations evaluate alternative interventions that target toxicity in the algorithm's recommendations. Finally, I collect survey evidence to trace users' behavior beyond the platform and show that the most affected users substitute away to other platforms. These results underscore the limits of piecemeal algorithmic regulation intended to moderate harmful content online.

The internet, once hailed as a global space for free expression, is increasingly being restricted through government-imposed shutdowns. Despite the internet’s critical role in facilitating business operations, providing access to social protection, and coordinating protest, little is known about when, where, and why these shutdowns occur. We leverage a novel, high-resolution dataset to systematically document the extent and targeting of shutdown in India, the country with the most frequent internet shutdowns. We present five main findings. First, internet shutdowns are common and impacted more than 44% of India’s population in 2018-2022. Second, while shutdowns occur often, they are highly targeted. An average shutdown in our sample period and states lasts only 2 days and affects 4 districts. Third, we find that shutdowns exacerbate inequalities as they are disproportionately concentrated in poorer areas and in areas with larger Muslim populations. Fourth, shutdowns are more likely to occur immediately following both peaceful protests and violent riots. Fifth, on days when there were large scale protests against the national government, shutdowns are substantially more widespread in states that are politically aligned with the national government.

This paper investigates how ethnic violence and subsequent residential segregation shape children's lives across social groups. Using variation in communal violence due to a Hindu nationalist campaign tour across India, I show that violence displaces Muslims to segregated neighbourhoods. I exploit exogenous differences in the planned and actual route of the campaign trail to show that communal violence is associated with an increase in residential segregation of communities threatened by violence. Surprisingly, I find that post-event, Muslim primary education levels are higher in cities that were more susceptible to violence. For cohorts enrolling after the riots, the probability of attaining primary education decreases by 2.3% every 100 kilometres away from the campaign route. I interrogate the role of neighborhood effects in driving primary education outcomes across social groups in India.

Local news is widely viewed as essential civic infrastructure, yet it has been in structural decline. Most research investigates its decline and proposes remedies focusing on supply rather than demand. This paper studies whether users engage with local news when it is produced and amplified at scale on a major social media platform in India. We implement a field experiment in which 65.3 million users were randomly assigned to receive sub-district local-news content in either the main feed or a dedicated local-news tab, or to remain in the control group with near-zero local-news exposure. The intervention successfully increased local-news consumption, but there is an engagement trade-off: non-local-news engagement significantly fell and total sharing in the treatment month declined by about 1% in both treatment arms, implying 1.8 million fewer monthly shares on the platform. Heterogeneity and content-genre analyses suggest that the decline is more pronounced among male users, several large language groups, and social content categories. These results highlight a demand-side constraint in revitalizing local news: supply-side efforts alone may not be sufficient if users do not have strong demand for local news on digital platforms.

FinTech lending has the potential to expand credit access by replacing relationship-based lending, where human discretion can lead to bias, with automated credit scoring based on large-scale data. Yet the effect of algorithms on bias is ambiguous in settings where many borrowers lack thick credit histories: when individual-level data are sparse, models may rely more heavily on aggregate signals that reflect past exclusion. This paper studies algorithmic bias in fintech consumer lending in India. Using a spatial difference-in-differences design and a market-level digitisation shock, we show that areas with historically thin credit bureau coverage receive significantly less fintech credit. This constraint is severe for first-time borrowers living in neighbourhoods with limited prior access to formal mortgage and consumer loans. These findings identify a distinct channel of algorithmic bias: unlike bias encoded in training data, this type of disparity arises when individual-level data are missing. Consistent with this mechanism, algorithmic bias is mitigated when thicker credit histories of minority groups in historically excluded neighbourhoods become more readily available. Repayment behaviour provides suggestive evidence of misallocation: borrowers from excluded areas who receive credit repay at higher rates than comparable borrowers elsewhere. Our results are consisted with an imputed credit-scoring algorithm wherein lenders place greater weight on social identity markers when assessing borrowers from historically excluded geographies. Together, these results suggest that fintech algorithms misinterpret historical exclusion as current risk, reproducing disparities in credit access.