Case Study 7.1: The Algorithm That Decided Your Feed

A Deep Dive into Social Media Recommendation Systems

The Setup

In September 2021, a former Facebook data scientist named Frances Haugen walked out of the company with tens of thousands of internal documents and brought them to the Wall Street Journal, the U.S. Congress, and regulatory bodies around the world. What those documents revealed was not that Facebook's recommendation algorithm was broken. It was that the algorithm was working exactly as designed — and that design had consequences no one had fully reckoned with.

At the heart of the story was Facebook's News Feed algorithm, one of the most influential recommendation systems ever built. By 2021, Facebook had nearly 3 billion monthly active users. The News Feed algorithm determined what each of those users saw when they opened the app — which friends' posts appeared first, which news articles were surfaced, which groups were suggested, which ads were served. Every day, the algorithm made billions of recommendation decisions, each one a tiny nudge shaping what a person saw, read, and believed.

How the Algorithm Works

Facebook's News Feed algorithm (and its equivalents on Instagram, TikTok, YouTube, and X) is a recommendation system built on a deceptively simple principle: show users the content they're most likely to engage with.

"Engagement" is measured through signals like: - Whether you click on a post - How long you spend looking at it - Whether you like, comment, or share it - Whether you click through to an external link - Whether you watch a video and for how long - Whether you return to the app after seeing a particular type of content

The algorithm learns from these signals using the techniques described in this chapter — collaborative filtering (what do users similar to you engage with?), content-based filtering (what features does engaging content have?), and deep learning models that identify complex, nonlinear patterns across hundreds of variables.

Over time, the algorithm builds a detailed model of each user's engagement patterns and serves content accordingly. If you engage with political content, you see more political content. If you engage with cooking videos, you see more cooking videos. If you engage with outrage, you see more outrage.

The Core Tension

Facebook's internal research, revealed in the Haugen documents, showed that the company's own data scientists had identified a structural problem: content that provokes strong emotional reactions — especially anger and outrage — generates more engagement than content that informs, entertains, or connects.

This created a tension between the algorithm's optimization goal (maximize engagement) and the platform's stated mission (connect people). Divisive political posts got more comments than nuanced ones. Sensational health misinformation got more shares than accurate health information. Inflammatory group content attracted more participation than constructive group content.

The algorithm wasn't "choosing" to amplify harmful content. It was optimizing for the metric it was given (engagement), and that metric happened to correlate with content qualities that were harmful.

The Feedback Loop

The recommendation algorithm created a powerful feedback loop:

  1. Emotional content gets more engagement (people click, comment, share)
  2. The algorithm observes the engagement and recommends similar content
  3. Creators learn what the algorithm rewards and produce more emotional content
  4. Users' feeds become more emotionally charged, which shifts their expectations and behavior
  5. The engagement data confirms the pattern, and the algorithm reinforces it

This feedback loop didn't just affect individual users — it reshaped the entire content ecosystem. News outlets learned to write more provocative headlines. Political operatives learned to craft messages that triggered algorithmic amplification. Content creators learned that nuance was punished and outrage was rewarded.

The International Dimension

The Haugen documents revealed that the algorithm's effects were particularly severe outside the United States and Western Europe. In countries like Myanmar, Ethiopia, and India, Facebook's recommendation system amplified content that incited ethnic violence, spread dangerous misinformation about COVID-19, and fueled political polarization — with far fewer content moderation resources available to catch the harms.

In Myanmar, a UN investigation found that Facebook had played a "determining role" in the spread of anti-Rohingya hate speech that preceded a military campaign later classified as genocide. The recommendation algorithm had amplified viral posts containing dehumanizing language and calls to violence, reaching millions of users who might never have encountered them in an organic, chronological feed.

The disparity was stark: Facebook had thousands of content moderators for English-language content and a handful for languages spoken by hundreds of millions of users. The recommendation algorithm operated at the same scale everywhere; the safeguards did not.

What Changed (and What Didn't)

In response to internal findings and public pressure, Facebook (now Meta) made several changes:

  • In 2018, the company modified the News Feed algorithm to prioritize "meaningful social interactions" — content from friends and family over publishers and brands. However, researchers found this change actually amplified divisive content in groups.
  • In 2021, the company introduced an "oversight board" to review content moderation decisions, but the board has limited authority over algorithmic recommendation.
  • In 2023, Meta introduced chronological feed options on both Facebook and Instagram, giving users the ability to see posts in time order rather than algorithmic order. However, the algorithmic feed remains the default, and most users never change the setting.

The fundamental business model — advertising revenue driven by user engagement — remains unchanged. As long as engagement is the primary optimization target, the structural tension between engagement and wellbeing persists.

Discussion Questions

  1. Optimization targets matter. Facebook's algorithm optimized for engagement. What alternative optimization targets could a social media platform use? What would the trade-offs be? (Consider: time well spent, diversity of content, user satisfaction surveys, informed citizenry.)

  2. The creator response. Content creators adapted their strategies to align with what the algorithm rewarded. Is this the algorithm's "fault," the creators' choice, or a systemic outcome that neither party individually controls? What does this suggest about where regulation should intervene?

  3. The default problem. Meta now offers chronological feeds, but the algorithmic feed remains the default. Research consistently shows that most users accept defaults. Is offering a chronological option sufficient, or should the default itself change? What are the arguments on each side?

  4. International equity. The recommendation algorithm operated at global scale, but content moderation resources were concentrated in English-speaking markets. What responsibility does a platform have to ensure its algorithm doesn't cause harm in markets where it has fewer safeguards? Should a platform decline to operate in markets where it can't moderate effectively?

  5. Connecting to the chapter. This case study illustrates recommendation, classification (content moderation), feedback loops, and the accuracy-interpretability trade-off. Identify one specific example from the case that illustrates each concept.

Key Takeaway

The Facebook/Meta case demonstrates that recommendation systems are not neutral conduits of information. They are active shapers of the information landscape, and their optimization targets — chosen by humans — determine what billions of people see, think about, and believe. Understanding how recommendation algorithms work is not a technical curiosity; it is a prerequisite for informed citizenship in a digital society.