Social media algorithms shape the beliefs people form

Even small changes to content curation algorithms can shape what people come to believe—for better or worse. 

The content people see on social media is typically determined by an algorithm. Researchers from the University of Copenhagen, the Max Planck Institute for Human Development, and TU Dresden have demonstrated in a controlled experiment involving approximately 1,500 participants that simple changes to how posts are sampled and ranked can affect the beliefs people form. The study suggests that alternative approaches to algorithmic content curation along with prosocial platform design may promote consensus and belief accuracy.  

What we see on social media often appears random: a brief comment here, an emotional post there, perhaps a post that many others have endorsed. Yet these posts do not appear by chance. They are carefully selected by algorithms that determine what appears in a user’s feed—and what does not. 

For years, researchers and policymakers have debated whether such algorithms contribute to polarization or lead people to form erroneous beliefs. However, answering this question has proven difficult. Real-world platforms are highly complex, and their recommendation systems are often opaque. 

Researchers from the University of Copenhagen, the Max Planck Institute for Human Development, and TU Dresden have now investigated this question in a controlled experiment. Their goal was to determine how different approaches to content curation influence the beliefs people ultimately form. 

A miniature social media experiment 

To examine the effects of recommendation algorithms as precisely as possible, the researchers developed a miniature version of a social media platform. In a first step, nearly 500 participants engaged with short argumentative posts on political and social topics. This created a content inventory showing how different groups of users responded to different posts. 

Subsequently, around 1,000 further participants were presented with feeds compiled from these posts. Different content curation algorithms were used to sample and rank the posts. Before and after viewing the feeds, participants reported their beliefs about the topics under discussion. This allowed the researchers to examine how different forms of algorithmic content curation influence belief updating and whether they promote consensus or belief accuracy.  

Why popular content is not necessarily the best 

One key finding concerns the widespread practice of engagement-based ranking. Posts that generated particularly high levels of engagement were often perceived by participants as interesting, high-quality and insightful. At the same time, however, they were more likely to lead participants to form more polarized and less accurate beliefs. The findings suggest that users’ perceptions of content and the actual quality of the information do not always align. 

“The platforms will say that their algorithms are just designed to help their users find the content they want to see,” says lead author Jason William Burton, Assistant Professor at the University of Copenhagen and Associate Scientist at the Center for Adaptive Rationality at the Max Planck Institute for Human Development. "But our study shows how the way the algorithms work can have potentially harmful effects on how users form beliefs about the world.”  

Burton stresses that the picture is not entirely negative. “Our results are, on the one hand, concerning because the social media platforms have become such dominant for a for civic discussion in our society, but our study could, on the other hand, be interpreted in a positive light as we show that it is possible to design algorithms that help users find more mutual understanding on these platforms.”  

Building bridges and improving belief accuracy 

The study demonstrated that alternative approaches are indeed possible. Under “bridging-based” ranking, content that receives approval across different groups is promoted. The idea is to surface content that helps to bridge divides and encourage positive interactions across diverse audiences. The results provide partial support for the hypothesis that bridging-based ranking can promote consensus.  

Another approach, “intelligence-based” ranking, entails promoting content that is likely to elicit belief updates that benefit collective accuracy. The researchers found partial support for the hypothesis that intelligence-based ranking can improve both collective accuracy and average individual accuracy, particularly for topics with an objective ground truth.  

However, neither approach proved superior in every situation. The effects were highly topic-specific and varied across issues.  

Why these findings matter 

The study shows that content curation algorithms are more than technical tools for organizing information. They actively shape the beliefs people form, even when the algorithms themselves are relatively simple.  

In the long run, the challenge will be to develop forms of prosocial platform design. “We are still at the very beginning,” says co-author Philipp Lorenz-Spreen, Head of the Computational Social Science research group at TU Dresden and Associate Scientist at the Center for Adaptive Rationality at the Max Planck Institute for Human Development. “In this study, we deliberately examined highly simplified versions of recommendation algorithms. But every development has to start somewhere. Technological innovation works the same way: You begin by testing approaches in the laboratory to see what works, and then gradually translate them into real-world applications.”  

The long-term objective is not to replace existing platforms or successful recommendation systems entirely. Rather, the challenge is to find ways of aligning them more closely with socially desirable goals. 

“The challenge is to achieve more prosocial effects without losing the positive aspects of today’s systems,” says Lorenz-Spreen. “There’s little value in developing an algorithm that could, in theory, enable better discussions if users just leave the platform. But societal harms such as polarization are too serious to keep ignoring.”  

The researchers therefore view their work as a first step in a broader process. New social networks and open platforms are already testing alternative recommendation algorithms that give users greater control over how content is curated. The findings presented here provide an initial experimental basis for systematically studying and further developing such approaches.  

At a glance 

• Algorithms shape the beliefs people form: Simple changes to content curation algorithms can affect beliefs, consensus, and belief accuracy.  

• Personalized engagement-based ranking can be problematic: Although users perceive these feeds positively, they tend to lead to less consensus and less accurate beliefs.  

• Consensus is possible: Bridging-based ranking can help promote consensus by surfacing content that receives approval across different groups.  

• Belief accuracy can be improved: Intelligence-based ranking can promote more accurate collective and individual judgments.  

• Small algorithmic changes can have large effects: Simple changes to how posts are sampled and ranked can substantially influence the beliefs people form. 

Original publication