When Big Data Enables Behavioral Manipulation [pdf] (economics.mit.edu)

Acemoglu et al. present a dynamic model showing how AI-enabled big data can let platforms manipulate user behavior by exploiting a product’s “glossiness” — transient attributes that make low-quality goods appear attractive. In the model a platform repeatedly offers one of n products (each high- or low-quality) and users get noisy signals over time. Glossiness is modeled as a binary state that decays from glossy to revealing at rate ρ (a continuous-time Markov process); if non-glossy, bad-news signals arrive at rate γ. In a pre-AI world platform and users update beliefs symmetrically, but in a post-AI world the platform uses pooled data to estimate which items remain glossy for particular user types and can strategically present glossy, low-quality items while users fail to realize the platform’s informational advantage. The technical payoff: when glossiness is short-lived (large ρ) better targeting improves matches and raises consumer welfare; when glossiness is long-lived (small ρ) the platform profitably steers users toward durable, deceptive glossiness and consumer welfare falls. Increasing the product set amplifies manipulation under small ρ because more items offer exploitable short-term appeal. The paper formalizes an “endogenous privacy cost” and highlights a policy-relevant trade-off: more granular user data and wider offerings can either improve matching or enable systematic behavioral manipulation depending on how persistent superficial appeal is.