How Social Media Algorithms Shape What You See and Think

The Feed Is Not Neutral. It Was Designed.

When Facebook launched in 2004, its feed showed posts in reverse chronological order—the newest content appeared first. In 2009, it introduced the "Like" button. In 2011, it introduced the algorithmic news feed, ranking content by predicted engagement rather than time. That decision—to show users not what was most recent but what an algorithm predicted they would react to most strongly—set the template for how two billion daily active users now experience reality. The algorithm's objective function was engagement: time spent, clicks, likes, comments, shares. It was not designed to maximize accuracy, wellbeing, or democratic participation. It was designed to maximize the amount of time people spent on the platform, because that time could be sold to advertisers.

How Recommendation Algorithms Work

Modern social media recommendation systems are collaborative filtering systems combined with deep neural networks. They predict, for each user-content pair, the probability that the user will engage with that content. "Engagement" is operationally defined as any measurable interaction: likes, comments, shares, saves, time-on-content, and in more sophisticated versions, emotional reactions and repeat views.

The inputs to these predictions include:

• Explicit signals: Past likes, shares, comments, follows, and search history
• Implicit signals: Time spent watching a video before scrolling past; whether a user returns to a post; whether they click a profile after seeing a post
• Social graph signals: What accounts the user follows, what their network engages with, what accounts similar users follow
• Content features: Topic classification, presence of images vs. video vs. text, detected emotion, presence of faces, recency
• Context: Time of day, device type, geographic location

TikTok's recommendation system differs from Facebook's in a fundamental design choice. Facebook's algorithm is primarily a social graph system: it prioritizes content from accounts you follow and from accounts your connections follow. TikTok built an interest graph: it recommends content based on inferred interests regardless of who you follow. New users see algorithmic recommendations immediately, before they follow anyone. This design allows TikTok to surface content from unknown accounts to massive audiences overnight—making virality more democratic but also making misinformation more easily amplified.

Filter Bubbles: The Evidence Is More Complex Than the Headline

Eli Pariser's 2011 book The Filter Bubble popularized the concern that algorithmic curation would trap users in information cocoons, exposing them only to content confirming their existing beliefs. The concept became widely cited. The empirical evidence is more nuanced.

A landmark 2023 study published in Science and Nature, using data from a randomized experiment conducted by Facebook and Instagram (Meta), found that:

• Reducing algorithmic amplification (switching to chronological feeds) increased exposure to content from like-minded sources, not less—because users' social connections are already politically homogeneous
• Seeing less political news reduced political knowledge but did not demonstrably reduce polarization
• The algorithm did amplify outrage-inducing and politically extreme content relative to a chronological baseline

A/B Testing at Scale: How Algorithms Are Tuned

Social media companies continuously refine their algorithms through A/B testing—running controlled experiments where a subset of users sees a modified algorithm and outcomes are measured across millions of users simultaneously. This is not hypothetical; it is the primary method of algorithm development.

In 2014, Facebook published a paper in the Proceedings of the National Academy of Sciences describing an experiment in which it manipulated the emotional content of 689,003 users' news feeds without their knowledge or consent—reducing positive or negative posts to test whether emotional contagion occurred online. The paper confirmed emotional contagion through algorithmic manipulation. The study generated significant criticism for its ethics but was published with institutional review board approval under a terms-of-service consent argument. It revealed, concretely, that emotional manipulation at scale was technically feasible and had already been conducted.

The Facebook Documents: Internal Evidence

In September 2021, Frances Haugen—a former Facebook data scientist—leaked tens of thousands of pages of internal research documents to The Wall Street Journal and the U.S. Securities and Exchange Commission before testifying before Congress. The documents became known as the "Facebook Papers" and revealed several relevant findings:

• Facebook's internal research found that Instagram (owned by Meta) made body image issues worse for approximately 32% of teenage girls who already felt bad about their bodies—and that Instagram was aware of and suppressed this finding
• A 2018 internal study found that a change to the news feed algorithm to reduce "viral misinformation" also reduced engagement among far-right users significantly—and the change was partially reversed under internal political pressure
• Internal researchers warned that the "reaction" emoji system (introduced in 2016)—particularly the "Angry" reaction—was being disproportionately applied to misinformation and outrage-inducing content, which the algorithm then amplified because high reaction rates correlated with engagement
• Facebook's internal research found that 64% of people who joined extremist groups on Facebook did so because the algorithm recommended the group

Algorithmic Amplification of Outrage: The Mechanism

The convergence of multiple studies and internal documents points to a consistent mechanism: outrage-inducing content tends to generate more comments, shares, and reactions than neutral or positive content—and engagement-maximizing algorithms therefore amplify it disproportionately. This is not a conspiracy or intentional harm. It is the unintended consequence of optimizing for engagement metrics that correlate with emotional intensity.

• A 2020 MIT Sloan Management Review study found that false news on Twitter spreads six times faster than true news, primarily because it is more novel and emotionally arousing
• A 2021 New York University / Cybersecurity for Democracy study found that Facebook pages sharing misinformation received six times more engagement per user interaction than mainstream news pages
• YouTube's recommendation system was documented by researchers at Google (YouTube's parent) to push users toward increasingly extreme content in the domain of conspiracy theories—a finding that led to algorithm changes in 2019, after which YouTube reported reducing recommendations of "borderline content" by 70%

You Are the Product—and Also the Researcher

Social media platforms contain within them the world's largest ongoing behavioral experiments, conducted with billions of participants who consented to participation in terms-of-service agreements almost no one reads. The output of these experiments is not shared publicly. It shapes what billions of people see, believe, and discuss. The algorithm is not a neutral curator. It is an optimization system with a specific objective function—and that function is not your wellbeing. Understanding the system is the starting point for navigating it deliberately.