We develop a dynamic model of how engagement-driven curation, such as Spotify’s algorithmic curation of playlists or Netflix’s personalized menus of TV shows, shapes the evolution of aesthetic tastes. Building on the canonical consumption capital model of Becker et al. (1988), we introduce two innovations: non-monotonic returns to familiarity, where moderate exposure to a style enhances appreciation but excessive exposure causes boredom; and curator-controlled exposure, where engagement-maximizing intermediaries determine which content consumers encounter. We show that when curators treat consumers’ familiarity as exogenous, failing to recognize it as shaped by their own past promotional choices, they can become trapped in inefficient equilibria with self-confirming beliefs, even though they may have high predictive accuracy for any given recommendation. Second, even when curators track familiarity as an endogenous state variable, they structurally underexplore when their evaluation horizons are short relative to long-run taste evolution timescales. We derive welfare implications, characterize comparative statics, and show that in both cases, lower prediction accuracy in curation can be welfare-improving.
