The Editor and the Algorithm: Recommendation Technology in Online News

We run a field experiment to study the relative performance of human curation and automated personalized recommendation technology in the context of online news. We build a simple theoretical model that captures the relative efficacy of personalized algorithmic recommendations and curation based on human expertise. We highlight a critical tension between detailed, yet potentially narrow, information available to the algorithm versus broad (often private), but not scalable, information available to the human editor. Empirically, we show that, on average, algorithmic recommendations can outperform human curation with respect to clicks, but there is significant heterogeneity in this treatment effect. The human editor performs relatively better in the absence of sufficient personal data and when there is greater variation in preferences. These results suggest that reverting to human curation can mitigate the drawbacks of personalized algorithmic recommendations. Our computations show that the optimal combination of human curation and automated recommendation technology can lead to an increase of up to 13% in clicks. In absolute terms, we provide thresholds for when the estimated gains are larger than our estimate of implementation costs.