Report - Belfer Center for Science and International Affairs, Harvard Kennedy School

Technology Primer: Social Media Recommendation Algorithms

| Aug. 25, 2022

Executive Summary

The use of social media platforms like Facebook, Twitter, YouTube, and TikTok is increasingly widespread, currently amounting to billions of users worldwide. In 2021, a majority of adults in the U.S. reported using at least one social media site, with the most popular ones being YouTube (81 percent) and Facebook (69 percent), and TikTok being especially common among 18–24-year-olds (55 percent).

Billions of pieces of content are published daily by users including individuals, public figures, interest groups, and organizations. Content ranges from personal updates to entertainment, tutorials, and news stories. In 2021, about one in two Americans reported getting their news from social media at least sometimes.

Social media companies deploy proprietary recommendation algorithms to automate the selection, ranking, and presentation of content on the platform’s “feed” or recommended content section, every time a user opens or refreshes the site or app. YouTube estimates that over 70 percent of views on the platform come from the company’s recommended section, as opposed to self-directed searches or shared links.

These algorithms leverage complex, distributed machine-learning models, such as deep neural networks, to identify, rank, and serve the subset of all available posts that are predicted to be “relevant” to each user based on how likely the user is to engage with it via views, clicks, likes, shares, and others.

To make these engagement predictions accurate and personalized to each user and at each point in time, these algorithmic recommendation systems are trained on billions of data points drawn from the user’s prior activity history and inferred interests (as well as those of other “similar” users) and adjusted to a particular context (e.g., time of day, device being used, etc.)

Recommendation systems are a crucial tool to drive and retain user engagement. Companies insist that their recommendation systems seek to connect users with people and content that matter to them, but critics argue that their business model prioritizes content that lures users to stay on the site longer and come back often, even if the content is controversial, harmful, or otherwise low-quality.

Numerous bills recently introduced in Congress reveal interest in regulating social media platforms due to their large influence over users’ online and offline experiences and mounting evidence of their downstream harms, including the amplification of misinformation and harmful content, worsening mental health, and the perpetuation of bias and inequality.

Regulatory approaches to social media recommendation algorithms include legislation, self-regulation, and external oversight. Legal scholars and commentators have suggested privacy and antitrust legislation (e.g., for algorithmic interoperability) as more feasible regulatory avenues than those related to content hosting and amplification liability, since they avoid imposing content preferences (which might face constitutional challenges) and would instead focus on increasing users’ agency by having greater control over their own data and a greater variety of recommendation algorithms to choose from.

Technology-driven solutions include the design of alternative recommendation systems that optimize for human values like fairness, well-being, and factual accuracy. In practice, aligning algorithms with complex human values is challenging, and usually involves trade-offs and unforeseen consequences. Sustainable solutions will likely require a better understanding of how these algorithms operate and how their benefits and harms manifest, underscoring the need to provide external researchers and regulatory agencies greater access to data on social media platforms’ algorithmic practices and outcomes. Ultimately, a successful approach to the regulation of social media recommendation algorithms will require a combination of government regulation, self governance, and external oversight to facilitate value alignment across these diverse actors and tackle the various challenges associated with this technology.

For more information on this publication: Belfer Communications Office
For Academic Citation: M. Vidal Bustamante, Constanza , Joaquin Quiñonero Candela, Lucas Wright, Leisel Bogan and Marc Faddoul. “Technology Primer: Social Media Recommendation Algorithms.” Edited by Ariel Higuchi and Amritha Jayanti. Belfer Center for Science and International Affairs, Harvard Kennedy School, August 25, 2022.

The Authors

Liesel Bogan headshot

The Editors