A recommendation system can predict what a viewer or listener will choose next and still make a streaming service less satisfying over time, according to a recent economic model.
The research, published in the Journal of Cultural Economics, examines what happens when a platform pursues immediate engagement while repeated exposure changes a person’s tastes.
Author Samsun Knight, an assistant professor of marketing at the University of Toronto’s Rotman School of Management, found that a system can become highly accurate at predicting individual choices while repeatedly promoting content that is already too familiar.
The paper presents a theoretical model, not an analysis of subscriber behaviour at Netflix, Spotify or any other platform. It does not calculate how many users cancel because their recommendations feel repetitive.
How familiar content becomes a trap
Knight’s model assumes that exposure initially helps people appreciate a style of music, film or television. Enjoyment reaches a peak and then falls as the material becomes overfamiliar.
A short-term recommendation system may miss both ends of that curve. It can abandon unfamiliar content before a user has had time to develop a taste for it, then keep promoting a proven favourite after the user is becoming tired of it.
The system also influences the behaviour it later measures. If it repeatedly shows one genre, the user has more opportunities to select that genre. Those selections then appear to confirm the earlier recommendation.
Economists call this a self-confirming equilibrium: the system keeps finding evidence for its own choices because those choices helped produce the evidence.
“There’s a real risk of being very good at a short-term metric while slowly degrading the long-term product,” Knight said in a Rotman School research release.
Accuracy can reduce discovery
Recommendation systems must balance exploitation and exploration. Exploitation means showing material that available data already indicates a user will like. Exploration means testing less certain choices to learn whether the user’s interests extend further.
In Knight’s simulations, a moderate amount of prediction error sometimes improved consumer welfare because it introduced variety. A perfectly accurate system focused on immediate engagement could keep returning to the safest option, while occasional errors exposed users to something new.
The result applies only under the model’s assumptions. It is not a general finding that inaccurate recommendations are better. Too much prediction noise made the simulated experience worse, and a platform could introduce variety deliberately instead of relying on mistakes.
Knight argues that longer evaluation periods and continued testing of unfamiliar material could give new tastes time to form. Human editors may also persist with a difficult or unproven work after weak early engagement, a choice that a short-term optimiser may reject.
Streaming firms are testing wider discovery
Recent work from Spotify shows how the industry is trying to move beyond habitual choices. In August, its researchers described GLIDE, a generative recommendation system for podcasts designed to combine established preferences with a listener’s current context.
Spotify reported that online tests involving millions of users increased non-habitual podcast streaming on its home screen by up to 5.4% and discovery of new shows by up to 14.3%. The company produced those results, and the tests did not examine Knight’s overfamiliarity model directly.
Spotify has separately said that clicks, streams and session length may be imperfect substitutes for long-term satisfaction. Its researchers have tested systems that use early signals to estimate whether a listener will continue with a newly discovered podcast over a two-month period.
Netflix also says its recommendations learn from viewing history, ratings, watch time and the behaviour of members with similar tastes. Recent viewing gradually outweighs the preferences supplied when a user creates a profile, according to the company’s explanation of its recommendation system.
Knight’s research does not establish that either company is making its service stale. It sets out a testable prediction: systems optimised for short-term engagement should produce narrower recommendations and faster boredom than human or editorial curation.
The paper proposes comparing algorithmic playlists with curator-made selections and tracking how people respond to different levels of exposure over time. Such tests could show whether a recommendation that wins today’s click also helps keep the customer interested months later.
