Social media feeds are populated by algorithms designed to keep users engaged. Now, researchers at Princeton are studying what happens when users can create their own feeds from scratch.
“There’s a lot of questions about the harms and consequences of algorithmic feeds,” said Manoel Horta Ribeiro, assistant professor of computer science and co-author of the research, which will be presented at the ACM Conference on Human Factors in Computing Systems on April 16.
Horta Ribeiro and his collaborators have created a tool that gives some users control over the content in their feed. “What happens when they have to deal with this agency, with its messiness and the challenges associated with it?”
The tool, called Bonsai, was built and tested on Bluesky, a decentralized social media platform. The researchers recruited 15 Bluesky users to test Bonsai over an average of 12 days. Most found Bonsai allowed them to curate social media feeds, which they enjoyed. But they also noted that considerable effort was required to construct a feed from scratch.
“They felt liberated from the algorithm, but there was also a trade-off,” said Marianne Aubin le Quéré, a co-author and postdoctoral researcher at the Center for Information Technology Policy. “It takes more cognitive effort sometimes when you have to be intentional about things.”
Making the system more accessible to casual social media users is one avenue for future research. “Not all users are ready to put in the effort to be intentional about their feeds,” said Omar El Malki, the paper’s first author and a former visiting student researcher at Princeton.
In theory, a tool to customize social media feeds could work on any social media platform. But Bonsai can only be deployed on a decentralized platform like Bluesky or Mastodon because these platforms have built-in application programming interfaces. These interfaces allow outside developers to access the platform’s data and use it to build third-party software tools.
To set up Bonsai, the user is first asked to describe their ideal feed in plain English. For example, one might write: “I want to see funny posts about dogs and updates on the English football club Arsenal. I don’t want to see anything distressing or political.”
A natural language model analyzes this answer and generates suggestions for accounts, hashtags and feeds that might be of interest. The user can edit the suggested list and add additional sources. Subsequent prompts ask the user to rank subject preferences and decide whether they prefer posts that are recent, popular or relevant. The tradeoffs between these choices are explained as the user navigates the system.
Using a natural language model was essential, said Aubin le Quéré. With natural language, users could be specific, telling the system to show only posts about cats but no posts about dogs, for example. Directions could also be subjective, requesting posts that are funny or not depressing. “That’s something that is very hard to convey via keywords, but is much easier to do via natural language,” she added.
While Bonsai only works on decentralized platforms, the researchers hope centralized platforms like YouTube, X and Meta will offer users more control over their feeds as the social media landscape evolves. With more platforms and more choice, more experimentation is possible.
“My hope is that many flowers bloom in a social media ecosystem that is more open,” said Andrés Monroy-Hernández, associate professor of computer science and a co-author on the paper. “But even on the closed platforms, I hope we’ll see many different kinds of uses.”
The research, Bonsai: Intentional and Personalized Social Media Feeds, will be presented at the Association for Computing Machinery Conference on Human Factors in Computing Systems (CHI) in Barcelona on April 16.
