In recent research, Journal of Cultural Economics suggests that highly accurate content recommendation algorithms can mistakenly make entertainment seem boring over time. Theoretical models show that injecting small amounts of randomness into these systems tends to improve long-term user satisfaction. This mathematical imperfection helps people discover new flavors before they get bored with their usual favorites.
Today, computer programs influence the way billions of people discover music, movies, and videos. The platform designs these systems to maximize immediate user engagement. However, researcher Samsan Knight noticed a contradiction in this modern setting.
Knight is an assistant professor at the University of Toronto’s Rotman School of Management and an affiliated faculty member at the University of Toronto’s Urban Graduate School. He is also a novelist and a graduate of the Iowa Writers’ Workshop. his second novel, Uniquenesspublished in July 2025; people Magazine best new book.
“I read Bourdieu’s book.” rules of art “It helped me name a lot of seemingly unconnected things that I had previously noticed about the creative algorithmic ecosystem but didn’t have the words to summarize. For example, I had the rather strange experience of initially really liking many of Spotify’s algorithmic song recommendations, and then being surprised to find out how aggressively Spotify kept recommending the same ‘great discovery’ songs until I couldn’t anymore,” Knight said. Don’t listen to them anymore. ”
He noticed similar patterns in other professions. “I’m also a novelist, and what I’ve heard from many publishing industry professionals is that the application of data analysis tools seems to have coincided with a huge increase in trend-following among publishers, at the same time that many readers were complaining that many of the big publishers’ novels sounded strangely similar.”
“Given that publishers probably want their readers to be happy, and Spotify wants me to continue to love the songs they recommend, I wondered why such a resource-rich company would end up in a bad balance,” Knight said. “This paper is, so to speak, the final form of that speculation.”
A key concept in this study is what economists call consumed capital. This idea means that the more you consume a certain type of art, the more you appreciate it. Human enjoyment of art follows an inverse curve. Moderate exposure to a style will make you like it more, but overexposure will eventually lead to boredom or satiation.
“The central idea is that aesthetic tastes evolve slowly over many years, so an algorithm that predicts what we want to see or hear today more perfectly than ever before may, by chance, prevent us from discovering what we’re supposed to like tomorrow,” Knight told SciPost. He explained that a certain amount of exposure is necessary to know how good a style is, but too much exposure can lead to people getting bored with that type of song or the show as a whole.
“In other words, the algorithm that helps you find exactly what songs you want tonight may be silently narrowing the set of songs you really want,” Knight says. “The most specific example I often give is hip-hop.”
“It took a long time for many listeners to learn how to listen to hip-hop. Hip-hop initially sounded unpleasant and unpleasant to people who were only used to listening to rock and roll,” Knight said. He pointed out that the same thing happened with rock and roll decades ago. “The idea of this paper is that if Spotify had been as dominant in the 1980s as it is today, listeners’ initial dislike might have pushed hip-hop so far down the algorithmic recommendation rankings that the genre might never have taken off.”
Recommendation algorithms typically spend weeks or months testing content. Human taste evolves over 10 to 20 years. Knight built a mathematical model to see what would happen if a myopic computer system controlled all access to art. Because the subject involves decades of the evolution of taste, Knight did not recruit human participants.
Instead, the authors built a dynamic mathematical model. Theoretical models use mathematical formulas to simulate complex human behavior under controlled conditions. The model included two main components. First, we simulated how people’s ratings would rise or fall as they repeatedly exposed themselves to a particular style.
Second, we simulated a curator choosing what content to show to users to maximize engagement. The researchers tested different types of algorithmic curators within the model. One type of simulated curator operated with a flawed understanding of the world. This model assumed that high engagement simply meant that the underlying quality of the content was high.
This flawed algorithm was unaware that its past recommendations were responsible for the familiarity that led to high engagement. The other type of mock curator got the right idea of how familiarity works. However, this second algorithm only plans for short-term engagement and operates very short-sighted.
Knight analyzed how these different algorithmic approaches affect overall user satisfaction over time. The simulation included tracking variables such as expected discount utility, a mathematical measure of the total enjoyment experienced by users over time. We also tracked the exploration rate in the simulation.
This ratio determines how often a computer program tries to show you something completely new and how often it tries to show you known favorites. To verify the mathematical proof, the authors performed a Monte Carlo computer simulation. This involved running the equation through 1,000 separate trials to observe the average result.
This model provides evidence that high-precision algorithms cannot fully explore new content. The flawed algorithm recorded low engagement signals when simulated users ignored unfamiliar genres. The genre was considered inherently bad because it lacked a broad timeline.
A mathematical proof shows that the search rate of such an algorithm eventually drops to zero. You become completely closed off to new possibilities. Instead, the system repeatedly recommended familiar content until the simulated user became completely bored.
This algorithm essentially created a self-fulfilling prophecy of monotonicity. Calculations show that the algorithm is stuck in a loop where incorrect assumptions appear to be completely correct based on the data collected. This study provides evidence for a phenomenon called straddling.
In a straddling scenario, the algorithm moves back and forth between two bad choices. Show high-quality items so often that users get fed up with them, and show low-quality items just enough to confirm that they’re not very good. The system never realizes that resting a quality item will restore the user’s enjoyment.
Even if the simulated algorithm correctly understood the change in preferences, it was unable to introduce sufficient diversity. The evaluation period was too short to understand the long-term benefits of increasing appreciation for new genres. As a result, simulated users experienced out-of-date conditions for extended periods of time.
Interestingly, computer simulations showed that less accurate recommendation systems actually perform better when it comes to long-term user satisfaction. When Knight introduced moderate prediction errors into the simulation, this noise caused the algorithm to occasionally recommend unfamiliar content. These serendipitous recommendations allowed the simulated users to develop an awareness of the new style.
Noise in the system gave users a break from their usual favorites. When the model was expanded to include more than two items, the benefits of the slightly flawed algorithm became even more apparent. In a perfectly accurate system, a new and extremely fun item would never receive enough exposure for users to get a taste of it.
Systems with a bit of randomness can result in unfamiliar items appearing in users’ feeds. Over time, this chance exposure causes the item to cross a threshold from unfamiliar to valuable. “On the contrary, a noisy, imperfect, creative product discovery system, or at least a system that engages in more exploration than seems optimal in the short term, could make life better for all of us,” Knight said.
A potential misconception of this research is the idea that all platform algorithms are completely broken or malicious. Mathematical models simplify complex human psychology and isolate specific mechanisms. Actual results may vary depending on individual user habits and the design of the particular platform.
Additionally, this study relies on theoretical simulations rather than tracking the actual viewing habits of live users over a 20-year period. Testing these ideas in the real world presents significant challenges. Platforms need to run experiments over many years to observe a full cycle of artistic familiarity, which is generally not realistic for technology companies.
Future research could test these predictions by comparing different types of platforms. Scientists can compare the long-term engagement of users who receive highly personalized algorithmic recommendations with those who receive more random, human-curated suggestions. Tracking how quickly different groups tire of a genre would help validate mathematical models in natural settings.
Researchers can also examine historical data from streaming services. By examining the period before and after highly targeted algorithms were introduced, scientists may be able to find evidence that artistic burnout occurred earlier. This would provide real-world data to support the theory that extreme accuracy destroys long-term entertainment value.
To address these issues, platforms can explicitly program their algorithms to understand familiarity as a change in state. In addition to responding to recent clicks, the platform may track how often users interact with certain art styles over several years. “We hope this research will contribute to the development of a healthier creative ecosystem for artists and art lovers,” Knight said.
The study, “Engagement-based curation and the evolution of taste,” was authored by Samsun Knight.
