Recommendation algorithms might be making your entertainment boring, new research suggests

A recent study published in the Journal of Cultural Economics suggests that highly accurate content recommendation algorithms might accidentally make our entertainment feel boring over time. The theoretical model indicates that injecting a small amount of randomness into these systems tends to improve long-term user satisfaction. This mathematical imperfection helps people discover new tastes before they grow tired of their usual favorites.

Today, computer programs dictate the discovery of music, movies, and videos for billions of people. Platforms design these systems to maximize immediate user engagement. But researcher Samsun Knight noticed a paradox in this modern setup.

Knight is an assistant professor at the University of Toronto’s Rotman School of Management and a faculty affiliate at the University of Toronto School of Cities. He is also a novelist and a graduate of the Iowa Writers’ Workshop. His second novel, Likeness, was published in July 2025 and was named a People magazine best new book.

“I read Bourdieu’s The Rules of Art and loved it, and that helped me put a name to a number of seemingly disconnected things that I’d previously noticed about the creative algorithmic ecosystem, but hadn’t had the language to put together before,” Knight said. “For example, I’d had this rather odd experience of really loving many of Spotify’s algorithmic song recommendations at first, but then was surprised to notice how Spotify belligerently kept recommending those same oh-what-a-great-find songs, until I couldn’t stand listening to them anymore.”

He noticed similar patterns in his other profession. “I’m also a novelist, and had heard from a number of publishing industry professionals that the application of data-analytics tools seemed to have coincided with a massive increase in trend-chasing among publishing houses, and at the same time, many readers were complaining that a lot of big-publishing-house fiction all sounded strangely similar.”

Free daily newsletter

“Given that publishing houses presumably want to be making their readers happy and Spotify wants me to keep loving the songs they recommend, I wondered why such well-resourced companies might get stuck in bad equilibria,” Knight said. “This paper is the final form, so to speak, of that wondering.”

A key concept in this research is what economists call consumption capital. This idea simply means that the more you consume a specific type of art, the more you build an appreciation for it. Human enjoyment of art follows an inverted curve. Moderate exposure to a style makes you like it more, but excessive exposure eventually causes boredom or satiation.

“The central idea is that because aesthetic tastes evolve slowly over multiple years, algorithms that predict ever-more-perfectly what you want to watch or listen to today may incidentally prevent us from discovering what we’d otherwise learn to love tomorrow,” Knight told PsyPost. He explained that it takes a certain amount of exposure to know how to appreciate a style, while too much exposure can make a person sick of a whole type of song or show.

“Put another way, the algorithm that can find you exactly the song you want tonight might be quietly narrowing the set of songs you’ll ever want at all,” Knight said. “The most concrete example I usually mention is hip-hop.”

“It took many listeners a long time to learn how to listen to hip-hop, which at first sounded abrasive and off-putting to people who were only used to listening to rock and roll,” Knight said. He noted the same thing happened for rock and roll a few decades earlier. “The idea of the paper is that if Spotify was as dominant in the 1980s as it is today, listeners’ initial distaste would have pushed hip-hop far down in their algorithmic recommendation rankings, and the genre may have never gotten off the ground at all.”

Recommendation algorithms usually test content over a few weeks or months. Human tastes evolve over ten or twenty years. Knight built a mathematical model to see what happens when short-sighted computer systems control all exposure to art. Because this topic involves decades of taste evolution, Knight did not recruit human participants.

Instead, the author constructed a dynamic mathematical model. A theoretical model uses mathematical equations to simulate complex human behaviors under controlled conditions. The model included two primary components. First, it simulated how human appreciation rises and falls with repeated exposure to a specific style.

Second, it simulated a curator that chooses which content to show users to maximize engagement. The researcher 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 a piece of content had high underlying quality.

This flawed algorithm failed to realize that its own past recommendations caused the familiarity driving that high engagement. The other type of simulated curator correctly understood how familiarity works. However, this second algorithm only planned for short-term engagement, acting with extreme short-sightedness.

Knight analyzed how these different algorithmic approaches affected overall user satisfaction over time. The simulation involved tracking variables like expected discounted utility, which is a mathematical measure of how much total enjoyment a user experiences over a long period. The simulation also tracked an exploration rate.

This rate determined how often the computer program would try showing the user something completely new versus showing them a known favorite. To verify the mathematical proofs, the author ran Monte Carlo computer simulations. This involved running the equations through one thousand separate trials to observe the average outcomes.

The model provides evidence that highly precise algorithms consistently fail to explore new content enough. When the flawed algorithm saw a simulated user ignore an unfamiliar genre, it recorded a low engagement signal. Because it lacked a broad time horizon, it assumed the genre was inherently bad.

The mathematical proofs show that such an algorithm’s exploration rate eventually drops to zero. It becomes entirely closed off to new possibilities. Instead, the system repeatedly recommended familiar content until the simulated users became entirely bored.

The algorithm essentially created a self-fulfilling prophecy of monotony. The math suggests that algorithms get trapped in loops where their bad assumptions appear completely correct based on the data they gather. The research provides evidence for a phenomenon called straddling.

In a straddling scenario, the algorithm bounces between two poor choices. It shows a high-quality item so often that the user becomes sick of it, and it shows a low-quality item just enough to confirm it is not very good. The system never realizes that resting the high-quality item would restore the user’s enjoyment.

Even when the simulated algorithm correctly understood that tastes change, it still failed to introduce enough variety. Its evaluation window was simply too short to see the long-term benefits of building appreciation for new genres. As a result, the simulated users experienced extended periods of staleness.

Interestingly, the computer simulations showed that a less accurate recommendation system actually performs better for long-term user satisfaction. When Knight introduced moderate prediction errors into the simulation, this noise forced the algorithm to occasionally recommend unfamiliar content. These accidental recommendations allowed the simulated users to build an appreciation for new styles.

The noise in the system also gave users a break from their usual favorites. When the model was expanded to include three or more items, the benefits of a slightly flawed algorithm became even more apparent. In a perfectly accurate system, a brand new, highly enjoyable item never receives enough exposure for users to develop a taste for it.

A system with a little bit of randomness occasionally bumps that unfamiliar item into the user’s feed. Over time, this accidental exposure pushes the item over the threshold from unfamiliar to appreciated. “Conversely, noisier, less-perfect systems of creative product discovery, or at least, systems that commit to more exploration than might seem optimal in the short-run, can make us all better off,” Knight said.

A potential misinterpretation of this work is the idea that all platform algorithms are entirely broken or malicious. The mathematical model simplifies complex human psychology to isolate specific mechanisms. Real world outcomes might vary depending on individual user habits and specific platform designs.

Additionally, the study relies on theoretical simulations rather than tracking the actual viewing habits of live users over a twenty year period. Testing these ideas in the real world poses significant challenges. Platforms would need to run experiments for years to observe complete cycles of artistic familiarity, which is generally not practical for technology companies.

Future research might test these predictions by comparing different types of platforms. Scientists could compare the long-term engagement of users who receive highly personalized algorithmic recommendations against those who receive more random, human curated suggestions. Tracking how quickly different groups get bored with a genre would help verify the mathematical model in natural settings.

Researchers could also look at historical data from streaming services. By examining periods before and after highly targeted algorithms were introduced, scientists might find evidence of faster artistic burnout. This would provide real world data to support the theory that extreme precision harms long-term entertainment value.

To address these issues, platforms could explicitly program their algorithms to understand familiarity as a changing state. Instead of just reacting to recent clicks, platforms might track how often a user has been exposed to specific artistic styles over several years. “I’d love for this research to contribute to the development of healthier creative ecosystems, for artists and art-lovers alike,” Knight said.

The study, “Engagement-based curation and the evolution of taste,” was authored by Samsun Knight.