Facebook proved its algorithm rewards anger in 2018. $164B in ad revenue kept it running. Here is the file.

In November 2018, a researcher inside Facebook’s integrity team finished a memo that answered a question the company had been asking itself. The document, titled “Does Facebook reward outrage? Posts that generate negative comments get more clicks,” reached a conclusion that executives could not claim they didn’t see: yes. The algorithm was systematically amplifying anger because anger drove engagement, and engagement drove revenue. The memo landed on the desks of senior leadership and then vanished into the company’s internal archives. It did not surface publicly until Frances Haugen walked out of Facebook’s offices three years later with thousands of internal documents on a hard drive [1].

By the time Haugen testified before the Senate in October 2021 [3], Facebook had generated $114.9 billion in advertising revenue that year [6], a direct product of an algorithm tuned to maximize time on platform, which turned out to mean maximizing the emotional states most likely to keep people scrolling. Anger was the most efficient of those states. The internal research confirmed it. The company kept going. The machine is still running.

The Official Story

Social media companies describe their algorithms as neutral recommendation systems designed to show users content they find relevant and valuable. Meta frames its feed ranking as an attempt to connect people with “meaningful” content from friends and family. X (formerly Twitter) presents its For You feed as a personalized discovery engine. YouTube describes its recommendation system as a tool for helping users find videos they want to watch. TikTok characterizes its For You page as a sophisticated personalization system that learns individual preferences. The companies argue that they are reactive to user behavior, showing people more of what they already engage with, and that harmful content represents a fraction of total platform activity that moderation teams work continuously to reduce.

This framing omits the part where they built the machine, measured what it amplified, and chose not to change it when their own research confirmed it was making things worse.

Follow the Money

Every major social media platform operates on the same revenue model: advertising. Advertisers pay for impressions and clicks; time on the platform generates both. More engagement equals more ad inventory equals more revenue. The algorithm’s function, stripped of the product language, is to maximize the duration of each user’s session.

Meta’s advertising revenue reached $114.9 billion in 2021 [6], the year Haugen blew the whistle. By 2024, it had grown to $164.5 billion [6]. By 2025, Meta’s total revenue crossed $200 billion [6], with advertising representing approximately 97 percent of income [6]. The company generated more than $800 billion in cumulative advertising revenue since 2019 [6]. This revenue, the entire financial foundation of Meta, is dependent on keeping users on the platform as long as possible. The algorithm is not incidentally responsible for this revenue. It is the revenue mechanism.

Facebook discovered empirically, through billions of A/B tests over years of product development, that certain emotional states reliably extended user sessions. Anger, fear, and moral outrage proved to be among the most powerful. In 2016 and 2017, as Facebook introduced emoji reactions [2], including the “angry” face, the company’s engineers had to decide how much weight to give each reaction in the engagement scoring that drove algorithmic amplification. They set the angry emoji at five times the algorithmic weight of a simple [2] “like.” One like was worth one point. One angry reaction was worth five.

This was not a rounding error or a miscalibration. Internal documents reviewed by the Washington Post, published in October 2021 [2], showed Facebook employees explicitly wrestling with the consequences of this formula, and that managers were slow or unwilling to act. By 2018, the company’s own integrity researchers had concluded that the five-to-one anger weighting was contributing to the spread of divisive content [1] and that posts generating angry reactions were more likely to contain misinformation. The information was there. The revenue incentive pointed the other way.

The Network

Mark Zuckerberg launched Facebook’s “Meaningful Social Interaction” algorithm overhaul in early 2018. The public framing emphasized connection, the platform would prioritize posts from friends and family over passive content consumption. Privately, the system worked differently. MSI, the internal name for the overhaul, measured “meaningful” by interaction volume, comments and reactions, which the angry emoji generated at outsized rates. Political parties across Europe and Asia reported back to Facebook [1] that year that they had discovered a dispiriting pattern: reaching audiences without publishing emotionally provocative, anger-generating content had become nearly impossible. The algorithm had set the price of visibility at outrage.

Frances Haugen joined Facebook in 2019 [3], specifically assigned to work on integrity issues, misinformation, election interference, and platform safety. What she found, as she told the Senate Commerce Committee on October 5, 2021 [3], was a company that had conducted rigorous internal research into these harms and had repeatedly chosen not to act on it. “Facebook has realized that if they change the algorithm to be safer, people will spend less time on the site [3], they’ll click on less ads, they’ll make less money,” she testified. Her central argument was not that Facebook executives were ignorant of the harm, it was that they had weighed it against revenue and chosen revenue.

The Nieman Journalism Lab, reviewing Haugen’s leaked documents, found that Facebook’s integrity teams had flagged the anger amplification problem repeatedly [1] across multiple years, and had been repeatedly deprioritized or overruled. The company disputed Haugen’s characterizations of specific decisions but could not dispute its own documents. One internal document from July 2020, reviewed by CNN, laid out proposals for adjusting the algorithmic weight of different types of reactions. The discussion had been ongoing for years. Facebook made no structural change to the anger weighting until outside pressure forced partial modifications.

Adam Mosseri, the head of Instagram, oversaw the platform that Facebook’s own researchers had studied for effects on teenage girls. The findings, leaked by Haugen and reported by the Wall Street Journal in September 2021, showed that Instagram made body image issues worse for one in three teenage girls who reported experiencing them. A separate internal study found that 13 percent of British teenage girls [1] traced suicidal thoughts to Instagram. These findings had been in internal circulation for years. Instagram continued to recommend beauty content, diet content, and influencer lifestyle content to teenage users, because that content generated engagement.

Musk’s takeover of Twitter, completed in late 2022, privatized the company and rebranded it as X. He expanded the algorithmically-ranked For You feed as the default experience, replacing the chronological feed that users could choose manually. A Knight Columbia study published in January 2024 found that 62 percent of political tweets [4] chosen by Twitter’s algorithm expressed anger, and 46 percent contained out-group hostility [4]. The algorithm was not selecting political content at random, it was selecting for anger. A November 2025 study published in Science, which experimentally reranked user feeds to reduce partisan animosity, found that even modest algorithmic adjustments measurably reduced users’ exposure to outrage content and measurably improved their reported sense of well-being.

The company knew it in 2018 [1]. It confirmed it in a hundred subsequent internal documents. It changed nothing.

What Was Buried

The most consequential thing buried in Facebook’s internal documents was not about advertising. It was about moderation.

At the time Facebook’s algorithm was amplifying anti-Rohingya content in Myanmar in 2017 [5], the company had approximately four content moderators who spoke Burmese, the primary language of the hate speech being spread. Facebook had 270 million users in Myanmar by that point. The Rohingya genocide, which UN investigators later determined warranted prosecution for crimes against humanity, organized and incited substantially through Facebook. The UN Fact-Finding Mission on Myanmar specifically cited Facebook as having “played a determining role” in spreading hate speech.

Haugen’s testimony before the Senate addressed the geographic inequality of Facebook’s moderation directly. She stated that Facebook concentrated 90 percent of its moderation resources [3] in English-language content. Most of Facebook’s users are not English speakers. In countries where Facebook is synonymous with the internet, including Myanmar, Ethiopia, and across sub-Saharan Africa, the platform’s algorithms operated with minimal moderation infrastructure while maximizing engagement through the same anger amplification that had been identified internally as harmful.

In September 2022, Amnesty International published a 200-page report titled “The Social Atrocity,” concluding that Meta’s systems had promoted violence against the Rohingya and that the company owed reparations. By December 2021, Rohingya refugees had filed a $150 billion lawsuit against Facebook [5] in U.S. and UK courts, alleging that the platform’s algorithmic amplification of hate speech had enabled the genocide. The case remained active as of 2024.

In September 2021, Facebook’s own researchers privately concluded that their platform was contributing to teenage mental health problems, specifically body image issues among girls, at a scale that warranted action. Facebook’s public response was to dispute the framing, CEO Mark Zuckerberg wrote to a senator [3] claiming the research showed only that “some teenagers feel worse” about their bodies, not that Instagram caused this. The internal documents showed researchers using stronger language. Instagram’s product team continued recommending the same categories of content.

Independent research has documented YouTube’s recommendation system steering users toward increasingly extreme content over time, the so-called “rabbit hole” effect [4]. Multiple studies, including work by Knight First Amendment Institute, find that algorithmic recommendation accounts for the majority of content consumption on the platform [4]. The recommendation system is not a minor feature, it is the primary way most people experience the platform, and internal documents showed it nudging users toward more emotionally extreme content.

The Stakes Now

It is 2026. Haugen’s Senate testimony is now five years old. Reporters have analyzed the Facebook Papers, memorialized them in books and documentaries, and largely watched the news cycle absorb them without producing structural change. Meta’s stock has more than recovered from its 2022 decline. Zuckerberg testified before the Senate Judiciary Committee in January 2024, apologizing to families of children harmed by social media, and the platforms have continued operating under the same algorithmic incentives.

The Section 230 debate, whether platforms should lose legal immunity for algorithmically amplified content, has stalled in Congress. The EU’s Digital Services Act requires large platforms to assess and mitigate systemic risks from their recommendation systems, with enforcement handled by national regulators. Early cases are underway, but no court has yet forced any platform to fundamentally redesign its engagement architecture.

Meta has made adjustments since the Haugen revelations: Instagram added labels to heavily-edited images in some markets, introduced a chronological feed option, and created a “Supervised” account mode for users under 16 in the United States. Facebook modified the angry emoji weighting modestly. These are changes to the surface. The underlying system, which ranks content primarily by engagement signals and treats angry engagement as high-value, has not been structurally altered. Meta’s 2024 advertising revenue of $164.5 billion [6] suggests the fundamentals have not changed.

The Science study from November 2025 offered the clearest evidence yet that the choice is real and available: when researchers experimentally reduced algorithmic amplification of partisan and outrage content, users felt better, reported lower levels of anger, were less polarized, and did not quit the platform. The technology for a less harmful version of social media demonstrably exists. It would generate less revenue. The platforms have made their choice, documented it in their own internal research, and continued making it every quarter.

New York, California, and twelve other states have advanced legislation seeking to hold platforms liable for algorithmic harms to minors. The outcomes of these cases and bills will determine whether the legal framework finally catches up to what internal Facebook researchers were documenting in 2018 [1]. Fourteen states had filed cases or legislation as of early 2026. The platforms’ legal teams are better funded than any state attorney general’s office. Without federal legislation, any structural resolution stretches across years, not months.

The user data that fuels the recommendation algorithm comes from a parallel commercial pipeline: data brokers selling profiles on every American consumer.

The One Thing That Matters

The November 2018 memo asked whether Facebook rewards outrage. The answer was yes. The company knew it in 2018 [1]. The company confirmed it in a hundred subsequent internal documents. Frances Haugen walked out with those documents in 2021 [3] because she understood what it meant when a company conducts its own research, finds that its product is making the world measurably worse, and files the research away. It means they’ve decided the revenue is worth more than the harm. Everything else, the Senate hearings, the congressional letters, the incremental product changes, the earnest public commitments to user well-being, fills the space between that decision and the moment when its cost becomes unavoidable. That moment has not yet arrived. The machine runs at $164 billion per year [6]. It will keep running until it doesn’t.