Reinforcement learning

Recommender Systems

 
Recommendation systems are typically AI systems based on machine learning. Through the collection of data about people’s preferences, behaviors (likes, clicks, followers, and so on), and past usage, the AI model actively learns and then predicts users' interests in certain content and personalizes the displayed content. Modern recommendation systems are highly sophisticated and often use hybrid machine learning approaches, including supervised, unsupervised, and reinforcement learning.
Today, recommendation systems are among the most widely used AI technologies and are implemented in many well-known social networks such as Facebook and TikTok, e-commerce sites such as Amazon, and video and movie streaming services such as YouTube and Netflix.
 
These systems offer numerous advantages, such as:
  • On the personal level, recommender systems can provide genuine utility through filtering content that is more catered to the needs and desires of the users.
  • On a societal and democratic level, recommender systems can be designed to push more diverse content, raise the political voice of marginalized social groups, and facilitate companies’ interactions with consumers.
 
Despite these advantages and the usefulness of personalized content for marketing purposes and user engagement, there are also various challenges and risks associated with them.
 

Psychological Effects

The effects of AI-based recommendation systems on cognitive autonomy are particularly evident among younger users. Features like endless scrolling, autoplay, or algorithmic suggestions exploit existing cognitive biases and often keep users on a platform longer than intended.

On TikTok and Instagram, for example, users can immediately swipe from one short video to the next. Combined with autoplay, this creates a continuous usage experience that keeps many people engaged for extended periods, creating the "rabbit hole effect". TikTok illustrates the scale of this engagement: globally, users spend an average of around 95 minutes per day on the app.

A group of vintage figures ride a circular machine, symbolising the endless, repetitive cycle of algorithmic scrolling and shallow interactions.
Children and adolescents are particularly vulnerable because their prefrontal cortex—the area responsible for critical thinking and impulse control—is still developing. Studies have associated prolonged use of algorithmically curated content with shorter attention spans, increased anxiety, and higher susceptibility to digital addiction, although much of this evidence is correlational and remains debated among researchers.

Alfano, M. et al. (2021). Technologically scaffolded atypical cognition: The case of YouTube’s recommender system. Synthese, 199(1–2), 835–858. https://doi.org/10.1007/s11229-020-02724-x

 

Polarization

Since content on social media is heavily filtered according to users' interests and preferences, the user experience is often repetitive and self-reinforcing. This encourages the creation of so-called filter bubbles: users do not receive all relevant information, but predominantly content that aligns with their existing interests.

A technological filter bubble is a decrease in the diversity of a user’s recommendations over time as a result of decisions made by different stakeholders in the recommendation system.

Michiels et al. (2022). What Are Filter Bubbles Really? https://doi.org/10.1145/3511047.3538028

An echo chamber describes an environment in which people mostly interact with like-minded individuals and predominantly hear opinions that confirm their beliefs. Recommendation systems can amplify this effect by suggesting similar content and reducing exposure to opposing perspectives.

Such phenomena can be problematic because they affect individual autonomy and have societal consequences, including increased polarization and a narrowing of public discourse. When people consistently consume confirming information, their willingness to engage with alternative perspectives decreases.

In political contexts, this can increase the risk of radicalization and reduce the ability to form balanced judgments. Without diverse information, decisions become more biased and polarized.