Social media algorithms and radicalization: what the evidence shows

Short answer: Social media ranking systems reward engagement, and outrage spreads fast. A 2018 Science study of about 126,000 stories on Twitter found that false news was 70 percent more likely to be retweeted than true news. Whether recommendation algorithms radicalize users is contested. A 2021 PNAS study of YouTube found no evidence that recommendations systematically cause far-right engagement.

EasWrk is a hosting platform, not a research institute. This essay summarises published sources, which are listed at the end. It connects to our essays on AI journalism and trust and AI dynamic pricing, and to the research hub.

Social Media Algorithms and Outrage: The Engagement Engine

The modern social media platform is not, at its core, a communications tool. It is a prediction machine: an enormously sophisticated system designed to forecast, with increasing precision, which piece of content will keep you scrolling for the next thirty seconds. Every tap, pause, like, and share feeds a continuously updated model of your psychological profile. That model has one purpose: to maximize the amount of time you spend inside the platform, because time spent translates directly into advertising revenue.

Engagement maximization is not an incidental feature of social media design. It is the central engineering objective. In this essay's view, platform phrases such as "meaningful social interaction" describe engagement maximization in softer language. The metric that matters is time on platform, and design choices such as the infinite scroll and the timing of notifications are oriented toward it.

The machine learning systems behind this optimization are technically impressive. YouTube's own engineers described its recommender in 2016 as ranking videos by expected watch time (Covington, Adams, & Sargin, 2016). Most other platforms do not publish their ranking objectives in comparable detail, so claims about how they rank are harder to test.

Critics often compare notification feeds to variable-ratio reinforcement, the schedule behind slot machines: you do not know when the next like, post or message will arrive. This is an analogy that helps explain compulsive checking. It is not a measured result of any study cited here.

The economic logic is straightforward. Social media companies earn their revenue primarily from advertising. Advertising rates are determined by audience size and engagement levels. An engaged user who spends three hours per day on a platform is worth far more to advertisers than a user who checks in briefly and leaves. This creates a direct financial incentive to engineer maximum engagement, independent of any consideration of whether that engagement is positive, healthy, or accurate. A 2018 study by MIT researchers Soroush Vosoughi, Deb Roy, and Sinan Aral, published in Science, found that false news diffused farther, faster, deeper and more broadly than the truth on Twitter in all categories of information (Vosoughi, Roy, & Aral, 2018). The study covered about 126,000 stories tweeted by about 3 million people from 2006 to 2017. False stories were 70 percent more likely to be retweeted, and true stories took about six times as long to reach 1,500 people (MIT News, 2018). Robots spread true and false news at the same rate, so humans, not bots, were more likely to spread false news.

Why? The authors found that false news was more novel than true news, and replies to false stories showed more fear, disgust and surprise. The study measured how stories spread, not what platform ranking systems did with them. A ranking system that rewards engagement would be expected to favour whatever people react to, but that link is an inference.

Engagement also rewards moral outrage. Brady and colleagues (2017) analyzed 563,312 social media messages on three polarizing moral and political issues. Each additional moral-emotional word increased a message's diffusion by a factor of about 20 percent. The effect was stronger within liberal and conservative networks than between them. In the author's view, a ranking system tuned to clicks and watch time has no built-in reason to prefer accurate content over content that provokes.

Algorithmic Identity and the Filter Bubble

In 2011, internet activist Eli Pariser published "The Filter Bubble," a book that introduced the concept of algorithmic personalization as a form of epistemic segregation (Pariser, 2011). Pariser argued that as platforms learned individual user preferences and tailored content accordingly, they would create self-reinforcing information environments in which users were progressively insulated from perspectives that challenged their worldview. The filter bubble, in Pariser's framing, was not merely a political phenomenon. It was a fundamental restructuring of how individuals encountered reality.

More than a decade later, the empirical picture is more complicated than Pariser's original formulation. A 2019 study by Guess, Nagler, and Tucker linked a survey to Facebook sharing data from the 2016 US presidential campaign. It found that sharing articles from fake news domains was relatively rare, and that it was concentrated among conservatives and users over 65 (Guess, Nagler, & Tucker, 2019). That is hard to square with a story in which filter bubbles push everyone toward falsehood.

The filter bubble critique does not end the argument. In this essay's view, the more useful question is whether ranking systems shape preferences over time and not only reflect them. The evidence on that is mixed, as the next section shows.

In the author's view, identity shaping is intertwined with commercial logic. Platforms profit from the creation of stable, predictable audience segments. A user with a clearly defined identity is a more valuable advertising target than an undifferentiated one. A passionate supporter of a particular political candidate, a devoted follower of a particular dietary philosophy, an enthusiastic consumer of a particular genre of outrage content: all are easier to sell to. The platform has a financial interest in helping you become more certain about who you are, even if what you are becoming is a more polarized, more tribal, more advertisable version of your previous self.

The sociologist Zeynep Tufekci wrote in The New York Times on March 10, 2018 that YouTube "may be one of the most powerful radicalizing instruments of the 21st century" (Tufekci, 2018). Her column drew on what she saw while following the 2016 election on YouTube: recommended videos grew more extreme and inflammatory whatever political content she searched for. She also drew on analysis by former YouTube engineer Guillaume Chaslot, who concluded that recommendations were overwhelmingly contentious and included conspiracy theories, because such content keeps viewers watching (Columbia Journalism Review, 2018). These are a columnist's observations and one engineer's analysis, not controlled studies.

One explanation, in the author's view, is that a system trained on watch time may find more extreme content engaging and keep recommending it. The audit studies in the next section test whether that happens in practice.

Teen use raises extra concern because identity formation is central in adolescence. The mental health evidence is covered in the teen mental health section below.

How Does Social Media Contribute to Radicalisation?

In this essay, online radicalization means a person moving toward more extreme views through what they see and who they talk to online. The pipeline metaphor says that a recommendation algorithm drives that movement. Researchers disagree on how much it does.

The YouTube recommendation algorithm and radicalization: what audits found

Ribeiro and colleagues (2020) audited 330,925 videos on 349 YouTube channels, along with more than 72 million comments and over 2 million recommendations. They found that users consistently moved from milder to more extreme channels, and that Alt-right videos were reachable through channel recommendations. They used commenting users as a proxy for radicalization, which is a limit of the method.

Hosseinmardi and colleagues (2021) reached a different conclusion with different data. They used a panel of more than 300,000 Americans and their browsing on and off YouTube from 2016 to 2019. News consumption on YouTube was dominated by mainstream and largely centrist sources, and consumers of far-right content were a small and stable share. The authors found no evidence that engagement with far-right content was caused by YouTube recommendations systematically. They concluded that political consumption on YouTube appears to reflect individual preferences that extend across the web.

In the author's reading, these studies do not settle the question. Recommendations may help some people find extreme content, while the broader population shows little sign of being pushed there. The choice of measure, comments or browsing, shapes the answer.

What Facebook's internal research showed

Frances Haugen, a former Facebook employee, testified on October 5, 2021 before a Senate Commerce subcommittee at a hearing titled "Protecting Kids Online: Testimony from a Facebook Whistleblower" (Haugen, 2021). She had given internal company documents to regulators and journalists. She alleged that company leaders disregarded evidence of harm to democracy and to children's mental health.

One internal study in those documents was "Carol's Journey to QAnon". According to NBC News (2021), a Facebook researcher created a test account in summer 2019 for a politically conservative mother in Wilmington, North Carolina, who followed Fox News and Donald Trump. Within two days Facebook recommended that she join QAnon groups, and within a week her feed was full of groups and pages that violated Facebook's rules on hate speech and disinformation. The account never joined the recommended groups. This was one test account, which shows what recommendations can do and not how often they do it.

In the author's view, the argument over radicalization matters for regulation. If the problem lies mostly in recommendation design, content moderation alone would not fix it. If audit evidence such as Hosseinmardi and colleagues (2021) is right that preferences matter more, moderation and user choice carry more of the weight. The evidence supports caution about both claims.

Social Media and Teen Mental Health: The Evidence

The relationship between social media use and adolescent mental health has drawn heavy public attention and scientific controversy. The evidence base is large, contested, and politically charged.

Jean Twenge, a psychologist at San Diego State University, was among the first researchers to document population-level shifts in adolescent mental health that correlated with smartphone adoption. In her 2017 book "iGen," Twenge presented data on rising depression, anxiety, loneliness, and suicide ideation among American teenagers beginning around 2012, which she linked to the spread of smartphones (Twenge, 2017).

Jonathan Haidt, a social psychologist at NYU's Stern School of Business, is a prominent academic voice arguing that social media causes significant mental health harm in adolescents, particularly girls. In "The Anxious Generation" (2024), Haidt argues that the shift to a phone-based childhood has harmed teen mental health, citing social comparison, cyberbullying, sleep disruption, and the displacement of in-person interaction (Haidt, 2024).

Social comparison is the mechanism Haidt emphasizes: adolescents, especially girls, measure their appearance, social status, and life quality against the curated highlight reels of peers and influencers. Instagram is often singled out in this account because its visual format makes appearance comparison central.

Internal Facebook research, reported by The Wall Street Journal in September 2021, included a 2019 slide stating "We make body image issues worse for one in three teen girls." Among teenagers who reported suicidal thoughts, 13 percent of British users and 6 percent of American users traced the desire to kill themselves to Instagram, according to another presentation slide (Al Jazeera, 2021). The research was not public until the reporting. Facebook's leaders disputed the reading: Mark Zuckerberg told Congress that research suggests social apps "can have positive mental-health benefits," and Instagram chief Adam Mosseri said the effects on teen mental health are "quite small."

The scientific debate is genuinely contested. Critics argue that effect sizes in most studies are small, that correlational evidence cannot establish causation, and that teen mental health has several possible causes. Orben and Przybylski (2019) analyzed three large datasets (355,358 participants in total) and found that the association between digital technology use and adolescent well-being was negative but small, explaining at most 0.4 percent of the variation in well-being. They concluded that such effects were too small to warrant policy change. Haidt, above, argues the harm is larger.

Both sides accept that the association is negative. They disagree on its size, its mechanism and its cause. In the author's view, the uncertainty does not absolve platforms of the duty to study and act on harms that their own research flags.

Surveillance Capitalism and Data Harvesting

In 2019, Shoshana Zuboff published "The Age of Surveillance Capitalism," a book that argues the digital economy runs on a new logic (Zuboff, 2019). Her argument was that the leading technology platforms had developed a new form of economic logic, distinct from industrial capitalism, in which human experience itself, the granular data of daily life, thought, and behavior, was appropriated as raw material for the production of behavioral prediction products.

In Zuboff's framing, social media platforms are not primarily in the business of providing communication services. They are in the business of harvesting behavioral surplus, the data that your interactions generate beyond what is needed to provide the service, and using machine learning to extract from that surplus predictive intelligence about your future behavior. This predictive intelligence is then sold to advertisers, political campaigns, and other interested parties who wish to influence your behavior in particular directions.

In April 2018 Facebook said that data on up to 87 million people, mostly in the US, may have been improperly shared with Cambridge Analytica (Facebook, 2018). In the author's view, that episode was an extreme case of a business model built on behavioral data, and not an aberration.

For children, the US Children's Online Privacy Protection Act (COPPA) generally requires verifiable parental consent before an operator collects personal information online from a child under 13 (15 U.S.C. 6501 and 6502). Whether platforms comply in practice is a question this essay does not answer.

In the author's view, the surveillance capitalism model is reinforced by the structure of platform data practices in ways that make regulation difficult. The behavioral data that platforms collect is not merely stored passively. It is processed by machine learning systems that continuously refine behavioral models, which are then used to make real-time decisions about content ranking, advertising targeting, and product recommendations. This creates a feedback loop in which more engagement generates more data, which enables better prediction, which drives more engagement. Breaking this loop requires not just data protection regulation but fundamental changes to the algorithmic recommendation architecture. Platforms have strong financial incentives to resist those changes.

Regulatory Responses and Their Limits

The regulatory response to social media's documented harms has been ambitious but limited in effect. The European Union's Digital Services Act (DSA), Regulation (EU) 2022/2065, entered into force on 16 November 2022 and applied to designated very large platforms from August 2023 (European Union, 2022). It is one of the most comprehensive regulatory frameworks applied to social media in any major jurisdiction. The DSA requires very large platforms to assess and mitigate systemic risks, including risks to civic discourse, public health and minors, to submit to independent audits, and to explain how their recommender systems work. The Commission can impose fines of up to six percent of global annual turnover.

The DSA also lets vetted researchers request data from very large platforms to study systemic risks (Article 40). This is a departure from the voluntary, opaque approach before the DSA, in which platforms' claims about their own practices were hard to verify from outside. Practical barriers to access remain.

In the United States, Section 230 of the Communications Decency Act gives platforms broad immunity from liability for user-generated content. Critics argue that it has also insulated platforms from accountability for algorithmic design decisions. In Gonzalez v. Google, decided in May 2023, the Supreme Court declined to decide whether Section 230 covers YouTube's recommendation algorithm, leaving the question open (CNBC, 2023).

Regulatory responses have also struggled with the transnational nature of platform operations. A platform regulated in California reaches users in every country on earth. Regulations passed in the European Union apply to platforms based in California. This jurisdictional mismatch creates opportunities for regulatory arbitrage and limits the effectiveness of any single jurisdiction's approach. The DSA's approach of focusing on systemic risk at the platform level, rather than on individual pieces of content, represents a more structurally sophisticated regulatory strategy, but its ultimate effectiveness will depend on enforcement resources and political will that are not yet established.

In the author's view, content moderation has limits as a regulatory strategy. It is expensive, uneven across languages and regions, and prone to both over- and under-removal. More fundamentally, content moderation addresses symptoms while leaving the underlying engagement optimization architecture intact. Removing a piece of harmful content does not change the algorithm that would recommend the next piece of harmful content in its place.

What Comes Next

The trajectory of social media and AI is moving in several directions simultaneously, and not all of them are toward greater harm. The deployment of generative AI in social media content creation introduces new and potentially more serious versions of existing problems. In the author's view, synthetic media, including AI-generated text, images, audio, and video, can now be produced cheaply and at scale with enough fidelity to deceive casual viewers. The combination of generative AI content production with engagement-optimizing recommendation systems creates conditions for information environments that are not merely distorted by bias and misinformation but actively manufactured at scale by automated systems.

The specific risk is the commodification of synthetic influence. Political actors, commercial interests, and state-level adversaries can now generate coordinated inauthentic content at a scale and personalization level that was previously impossible. A synthetic influence operation can generate thousands of individually tailored pieces of content, each optimized for a specific audience segment's psychological profile, delivered through algorithmic recommendation systems that amplify based on engagement signals that the synthetic content is specifically engineered to trigger. Detecting such operations is a genuine technical challenge, and in the author's opinion platforms may not be equipped to address it at the scale it could reach.

At the same time, the political and regulatory momentum for genuine structural reform is building in ways that were not evident five years ago. The Haugen disclosures, the DSA, and the growing scientific literature on adolescent mental health have created conditions for more serious regulatory action than anything that has been implemented to date. The question is whether regulatory frameworks can be designed with sufficient specificity and technical sophistication to address the actual mechanisms of harm, rather than merely their most visible symptoms.

Some researchers and advocates have proposed structural interventions that go beyond content moderation and transparency requirements. These include: algorithmic choice requirements, which would allow users to opt out of engagement-optimized recommendation in favor of chronological or user-curated feeds; interoperability mandates, which would reduce platform lock-in and allow users to migrate to alternatives; mandatory risk assessments for algorithm updates; and data minimization requirements that would limit the volume of behavioral surplus platforms can collect and retain.

None of these interventions is sufficient on its own, and each faces implementation challenges. The evidence is mixed on how much recommendation algorithms drive radicalization, and the effect on teen mental health is disputed in size. The case for structural reform therefore rests on a narrower set of findings: false news spreads faster than true news (Vosoughi, Roy, & Aral, 2018), platforms collect behavioral data at scale, and internal research reported in 2021 pointed to harms for some teenage users. In the author's view, those findings justify transparency and independent research access even where causation is unsettled.

In the author's opinion, the engagement engine is unlikely to reform itself, because its financial logic is clear. What comes next will depend on whether regulators, researchers, and the public can maintain the clarity of analysis necessary to address the machine for what it actually is.