Karthik Srinivasan
Welcome!
I am an Assistant Professor at the University of Michigan's School of Information.
I work on topics in behavioral economics, labor economics, and political economy. Currently, I'm interested in the economics of social media.
Feel free to reach out to me at ksrini@umich.edu or take a look at my CV.
Thanks for visiting my website!
Working Papers
Humans are social animals. Is the desire for attention from other people a quantitatively important non-monetary incentive? I consider this question in the context of social media, where platforms like Reddit and TikTok successfully attract a large volume of user-generated content without offering financial incentives to most users. Using data on two billion Reddit posts and a new sample of TikTok posts, I estimate the elasticity of content production with respect to attention, as measured by the number of likes and comments that a post receives. I isolate plausibly exogenous variation in attention by studying posts that go viral. After going viral, producers more than double their rate of content production for a month. I complement these reduced form estimates with a large-scale field experiment on Reddit. I randomly allocate attention by adding comments to posts. I use generative AI to produce responsive comments in real time, and distribute these comments via a network of bots. Adding comments increases production, though treatment efficacy depends on comment quality. Across empirical approaches, the attention labor supply curve is concave: producers value initial units of attention highly, but the marginal value of attention rapidly diminishes. Motivated by this fact, I propose a model of a social media platform which manages a two-sided market composed of content producers and consumers. The key trade-off is that consumers dislike low-quality content, but including low-quality content provides attention to producers, which boosts the supply of high-quality content in equilibrium. If the attention labor supply curve is sufficiently concave, then the platform includes some low-quality content, though a social planner would include even more.
Can making decisions in extreme cases bias subsequent decisions? I study this question in a high-stakes field setting: felony sentencing. I estimate the effect of sentencing a first-degree murder on the length of sentences issued to subsequent defendants. I use data on the universe of felony sentencing decisions in Cook County to estimate a difference-in-differences design comparing judges in the same courthouse who have and have not recently sentenced a first-degree murder. Judges issue sentences that are 13% longer in the 10 days after they sentence a first-degree murder. Effects are twice as large for defendants who share the same race as the murderer and defendants who face high-class felony charges. A back- of-the-envelope calculation suggests that this bias affects 6% of defendants on an ongoing basis, because judges regularly sentence first-degree murders.
I study the scope of a principal-agent problem in the field. I analyze news firms and journalists with possibly misaligned preferences over the partisan slant of content, and find that the firm's ability to exert control is limited. I construct a dataset that links 2,700 journalists to firms, news articles, and Twitter profiles. I measure article slant with a machine learning algorithm I train to identify partisan phrases. Using a movers design, I find that firm ideology does not change the slant of a journalist’s writing. In contrast, journalist ideology, estimated using the following decisions of Twitter users, is strongly correlated with article slant.