Does Algorithmic Recommendation Weaken Consumers' Original Preferences? -- An Economic Analysis Based on Information Constraints

Abstract

The paper explores the topic of whether algorithmic recommendation systems radically change the preferences of consumers by constraining information. We construct a theoretical model of search costs, dynamic learning and endogenous preference formation, in which algorithms act as information intermediaries that selectively filter information about available options in a systematic manner. On the basis of extensive e-commerce purchases records of 35,874 purchases made by 1,000 participants over a period of 12 months, we apply difference-in-differences techniques to determine causal impacts. Our findings indicate that algorithmic exposure enhances concentration of consumption by 52% (measured using Herfindahl Index) and decreases the diversity of preferences by 91 percent (measured using Shannon entropy) and both are statistically significant at p<0.001. The analysis of mechanisms shows that feedback loops have 45% of the total effect, the information filtering has 35%, and cognitive dependence characterizes 20%. The effect of treatment has a heterogeneous search cost sensitivity and accrual over time due to the accumulation of user history over time among algorithms. We do not make any evidence of other readings like refinement of tastes or supply limitation. The results of this research contradict the traditional perspective that algorithms only support the expression of preference, pointing rather to its active involvement in the formation of preferences with unclear welfare implications. Some policy recommendations suggested are platform design that promotes diversity, compulsory impact evaluation of algorithms, and consumer empowerment interventions.