From Code to Composition: Personalized Recommendations and Product Innovation (joint with Maria Trejos, Diana Van Patten, and Kai Hao)

This paper studies how personalized recommendation algorithms (PRAs), a central application of artificial intelligence, shape product innovation. We develop an equilibrium model in which consumers have heterogeneous tastes, producers choose product characteristics subject to costs, and a platform recommends products while learning preferences from noisy signals. Personalization changes producers' incentives by allowing them to target different consumer segments. This response can increase market-wide differentiation and make new products more distinct from the existing stock, even when personalization narrows individual exposure. We examine these implications in the music industry, where PRAs have been widely adopted since the mid-2010s. Using tools from music information retrieval applied to raw audio files, we find that songs have become more original, both within a given year and relative to past releases, following the widespread adoption of PRAs. We extend the model to CES preferences and combine our originality estimates with listening-time shares to quantify the welfare implications. The implied consumer-surplus gains associated with additional originality range from 2.2% to 2.4% in 2023 relative to a counterfactual with originality fixed at its 2013 level.