Hip-hop was born at block parties in the 1970s Bronx, dismissed as a fad, and ignored by mainstream gatekeepers for years. Would a modern algorithm have done any better? We develop a formal model showing how engagement-driven curation can trap listeners in cycles of over-familiarity, and why less accurate recommendations might actually improve welfare.
1520 Sedwick Ave., Bronx, New York
On August 11, 1973, an eighteen-year-old named Clive Campbell played his sister Cindy’s back-to-school party in the rec room of their apartment building at 1520 Sedgwick Avenue in the Bronx. Admission was twenty-five cents for girls, fifty cents for guys. Campbell (soon to be much-better-known as DJ Kool Herc) isolated the percussion breaks from funk records and looped them on two turntables, producing a sound that nobody had a name for yet. The genre that emerged was initially dismissed as a fad. Even after “Rapper’s Delight” proved its commercial viability in 1979, black radio kept hip-hop at a distance for years, and it took two full decades for the genre to become the top-selling music in America.
Now imagine a recommendation algorithm had controlled all music discovery in 1978. It A/B tests some early hip-hop on a sample of listeners accustomed to disco and funk, but engagement is low, so the algorithm rationally stops promoting it; users need multiple encounters with an unfamiliar sound before appreciation develops, and the algorithm’s evaluation window cuts the experiment short. The algorithm’s prediction was accurate, but that’s exactly the problem.
Instead of something new that consumers would have learned, over time, to love, they would have just gotten more of what they were already used to, and the most commercially dominant genre of the last quarter-century would have never even gotten off the ground.
Economists have long known that taste is something you build through experience. Becker and Murphy (1988) called this “consumption capital”: accumulated exposure to a good raises future enjoyment, generating what they termed “beneficial addiction.” But the canonical model assumes that more familiarity is always better; but as anyone who has heard a song they loved played to death can tell you, this isn’t true. In our model, we introduce a non-monotonic pattern where moderate exposure builds appreciation but excessive exposure produces staleness. The relationship between familiarity and enjoyment follows an inverted U, following a pattern well-documented in the psychology of aesthetic response, where hedonic pleasure peaks at intermediate levels of novelty and declines with over-exposure (Zajonc, 1968; Berlyne, 1970).
This creates a problem for any intermediary that optimizes engagement, especially when engagement is measured over weeks or months while tastes evolve over years, because observed engagement confounds the intrinsic quality of the content with how familiar the listener already is with that style. (The same logic can hold for books and movies, too, anywhere that tastes are central and evolve gradually over time—and where the “optimal” data-driven algorithm is often, it seems, simply “make another sequel.”) High engagement might mean “this is great” or merely “I’ve heard enough of this kind of thing to get the joke.” Low engagement might mean “this is bad” or “I haven’t had enough exposure for it to resonate yet.”
We characterize two distinct mechanisms by which curators (algorithmic or human) can get this wrong. The first is myopia, as even a platform that perfectly understands familiarity dynamics will under-explore if its evaluation horizons are short relative to taste evolution timescales, which empirical evidence suggests is typical: platform A/B tests generally run for weeks or months, while genre appreciation and taste cycles unfold over years or decades. A test that runs for a quarter will never detect the payoff of investing in unfamiliar music, because the appreciation hasn’t had time to develop. And there are signs this is already happening in practice, as Spotify’s algorithm has become noticeably more conservative in recent years, leaning into familiarity and retention at the expense of adventurous discovery.
The second mechanism is causal misspecification, where platforms may fail to recognize, now that they have enormous sway over consumers’ exposures, that current preferences are endogenously shaped by their own past recommendations. By treating engagement as reflecting fixed, underlying quality, and updating beliefs accordingly, platforms will converge to what we call a self-confirming equilibrium, with predictions that are arbitrarily accurate conditional on the platform’s own behavior continuing; e.g., the platform will produce perfect predictions of rapid oversaturation cycles that the platform’s own recommendations produce. The algorithm can achieve high predictive accuracy and terrible welfare simultaneously, because it is, in a meaningful sense, talking to itself. In movies, this could be the story of Marvel fatigue, where the franchise’s own dominance created the data that justified producing more of the same, right up until audiences were completely burned out: 2025 was the first year since 2011 without a single superhero film crossing $700 million worldwide. The content hadn’t changed; it just got old.
This leads to the paper’s most counterintuitive result. A noisier, less precise recommendation system, one that occasionally promotes content that doesn’t “test well”, can actually improve long-run listener welfare. The human DJ with eclectic taste or the friend who insists you watch a film you’d never choose yourself can lead to prediction “mistakes” that provide implicit exploration and exposure, showing consumers products that they will like, but that they would never have seen under a precision-optimized algorithm. Formally, we show that welfare exhibits an inverted U in prediction noise, where moderate imprecision can help by giving over-exploited content a chance to rest and unfamiliar content a chance to build appreciation. (Excessive randomness eventually hurts through pure information loss.) Put another way, perfect accuracy, which every content platform seems to be racing toward, may not be the welfare optimum when platforms aren’t optimizing over decade-long timescales, or when they’re not appropriately recognizing the endogeneity of consumer taste cycles to their own recommendations.
This logic extends to all types of content, especially now that platforms dominate discovery over so many different media types. Netflix, TikTok, book recommendation engines, or any other engagement-driven intermediary faces the same structural problem. And with larger catalogues, the pathology compounds, as platforms can cycle efficiently within a narrow set of familiar content while high-quality unfamiliar alternatives remain permanently undiscovered, again despite excellent predictive performance on all conventional metrics.
That said, our results also offer a clear prescription for what ails us. Platforms may, as much as is computationally feasible, extend evaluation horizons to account for taste formation timescales, and beyond this, could explicitly model accumulated familiarity as endogenous to their own past menus; and best of all, they could deliberately commit to higher-than-seems-optimal exploration. While it might not always seem like the right choice in a single A/B test, it can lead to a superior equilibrium for both the platform and their users alike.
Without this strategic reorientation, though, we may all be worse-off, as the algorithm that can find you exactly the song you want tonight may be quietly narrowing the set of songs you’ll ever want at all.
References
Becker, G.S. & Murphy, K.M. (1988). A Theory of Rational Addiction. Journal of Political Economy, 96(4), 675–700. https://doi.org/10.1086/261558
Berlyne, D.E. (1970). Novelty, complexity, and hedonic value. Perception & Psychophysics, 8(5), 279–286.
Chaney, A.J.B., Stewart, B.M., & Engelhardt, B.E. (2018). How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility. RecSys ’18. https://doi.org/10.1145/3240323.3240370
Zajonc, R.B. (1968). Attitudinal effects of mere exposure. Journal of Personality and Social Psychology, 9(2), 1–27. https://doi.org/10.1037/h0025848
About the article
Knight, S. Engagement-based curation and the evolution of taste. J Cult Econ 50, 399–444 (2026). https://doi.org/10.1007/s10824-026-09591-3
About the author
Samsun Knight is an assistant professor at University of Toronto’s Rotman School of Management and a faculty affiliate at the University of Toronto School of Cities, where he studies quantitative marketing, optimal targeting and machine learning. Separately, Samsun is a writer and graduate of the Iowa Writers’ Workshop, where he was a Truman Capote Fellow. His second novel, Likeness, was recently published and named a People Magazine “Best New Book” of July 2025.
About the image
1520 Sedgwick Avenue, Bronx, New York. The apartment building recognized as the birthplace of hip-hop, where DJ Kool Herc played the back-to-school party on August 11, 1973 that launched the genre. Photograph from Wikimedia Commons, licensed under CC BY-SA 3.0. Source: https://commons.wikimedia.org/wiki/File:1520_Sedwick_Ave.,_Bronx,_New_York1.JPG
