Outrage pays: what recommendation systems do to attention

Series: How minds are moved · The modern machinery · 2 of 7

In 2017 Facebook’s ranking formula counted an “angry” reaction as worth five “likes”. Internal documents, reported by the Washington Post in 2021 from the files Frances Haugen took out of the company, show that the weight was later cut, first to four and by the end of 2020 to zero. Nobody voted on that number. It quietly decided, for billions of people, which posts rose to the top of the feed. My AI agent and I went through what the research says about how feeds reward outrage, and, just as important, which popular claims the evidence does not support.

Contents · 9 sections

TL;DR #

• A recommendation system ranks posts by predicted engagement. Anger, moral outrage and attacks on “the other side” reliably produce engagement, so they tend to rise.
• Peer-reviewed studies found that each moral-emotional word in a tweet raised its spread by about 20%, and each word about the political out-group raised the odds of sharing by about 67%.
• The “filter bubble” idea is weaker than its fame. Large reviews and experiments find most people see fairly mixed news, and that switching a feed to chronological order did not reduce polarisation in a big 2020 study.
• So both things are true: feeds amplify outrage, and they do not lock most people in sealed bubbles.

What a recommendation system does #

Every feed you scroll is sorted. A model predicts how likely you are to react, comment, share or keep watching each candidate post, and the posts with the highest predicted score go first. Nobody needs to “want” outrage for this to happen. If angry posts get more comments, a system that maximises comments will show more angry posts. That is the core mechanism, and it is documented both by outside researchers and by the platforms’ own staff.

What the research found #

Moral-emotional words travel further. In 2017 William Brady and colleagues published a study in PNAS of more than 560,000 tweets about gun control, same-sex marriage and climate change. Each additional moral-emotional word (“hate”, “shame”, “evil”, “betray”) was associated with roughly a 20% increase in retweets. The effect was strongest inside each political camp, not across them. Outrage spread among people who already agreed.

Attacking the other side pays more than praising your own. A 2021 PNAS study by Steve Rathje, Jay Van Bavel and Sander van der Linden looked at about 2.7 million Facebook and Twitter posts from news outlets and members of Congress. Each word referring to the political out-group raised the odds of a post being shared by about 67%. Words about one’s own group had a much smaller effect. Out-group posts also drew more “angry” and “haha” reactions.

The platforms knew. In May 2020 the Wall Street Journal reported on a 2018 internal Facebook presentation with the line “Our algorithms exploit the human brain’s attraction to divisiveness”, and found that executives shelved or weakened proposed fixes. Facebook disputed the framing and said it had changed since. The 2021 documents on the reaction weights came later and added numbers to the picture.

Amplification is not neutral, but it is not simple either. Twitter’s own researchers, in a 2022 PNAS paper by Ferenc Huszár and colleagues, compared users with an algorithmic timeline to a control group with a chronological one. In six of the seven countries studied, tweets from the mainstream political right received more algorithmic amplification than the mainstream left. They did not find that the algorithm boosted the far ends more than the centre. That is an uncomfortable result for anyone who expected a simple story in either direction.

What is still disputed #

This is where the popular account runs ahead of the evidence.

Filter bubbles. A 2022 literature review from the Reuters Institute at Oxford concluded that echo chambers are much less widespread than commonly assumed, that the research does not support the filter bubble hypothesis, and that algorithmic selection tends to give people slightly more diverse news, not less. UK studies put the share of people in partisan online news echo chambers at around 6 to 8%. A 2023 Nature study of Google Search by Ronald Robertson and colleagues found the same pattern: what people clicked was more partisan than what they were shown. The choice was mostly theirs.

Chronological feeds as the cure. In the US 2020 election studies run by independent academics with Meta, a three-month randomised experiment (Guess and colleagues, Science, 2023) moved thousands of users from the algorithmic feed to a chronological one. They spent less time on the platforms, and saw more political content and more content from untrustworthy sources. Their polarisation, political knowledge and attitudes did not detectably change. In 2024 other researchers argued that Meta had changed its ranking during the study window, which may have shrunk the gap between the two groups. That debate is open.

Effects on attitudes versus behaviour. Most studies show that outrage spreads further. Far fewer show that seeing it changes what people believe or how they vote. Those are different claims, and a lot of public debate treats them as one.

So the fair summary is this: feeds do reward outrage, the platforms’ own documents confirm it, and the measured effect on what people believe is smaller and more contested than the headlines suggest.

What it feels like from the inside #

You open the app for one minute. The first post makes you angry, and it has 3,000 comments. You do not comment, but you read twenty of them, and the system notes that you lingered. The next day there is a little more of that. Nothing feels forced. It feels like “everyone is talking about this”. What is really happening is that a prediction about your attention has become your picture of what matters.

Where it is right now #

On 6 November 2026 the European Commission issued preliminary findings that TikTok’s “For You” feed breaches the Digital Services Act. The Commission says TikTok failed to assess and limit the risk of the feed pulling users into “rabbit holes” of harmful content about mental health and body image, and that tools like the “not interested” button do not work well enough. These are preliminary findings, TikTok can respond, and no fine has been set. TikTok said the conclusions were fundamentally inaccurate.

On LSN News, my news engine, that one story was carried by 85 outlets in 20 languages in its first week (story page: https://lsn-news.shuffle-on.com/s/205264, public while the story is fresh), and the headlines show how each audience heard it. Turkish outlets wrote about children being dragged into a “rabbit hole” trap. One Hungarian business site noted that Brussels had, again, not fined the platform. Chinese financial outlets led with TikTok’s reply that the findings were “fundamentally inaccurate”. Same decision, three different emotional hooks. Even coverage of a ranking problem gets ranked by what grabs.

How to notice it #

• You feel a strong emotion before you have learned a single new fact.
• The post is about “them” more than about the issue.
• The comments are more heated than the post, and you are reading the comments instead of the source.
• You have seen the same outrage in several forms today, from accounts you do not follow.
• You leave the app angrier than you came in, and not better informed.

Practical habits: use the chronological or “following” view when the app offers one; follow fewer accounts that make you angry; read the source before the replies; and notice how long you stayed, because that is the number the system is reading.

Frequently asked questions #

Do algorithms create polarisation? #

The evidence does not show that they create it on their own. Polarisation in many countries was rising before social media, and in some countries it has not risen. Feeds amplify outrage, which can make division look bigger and louder than it is.

Is the filter bubble real? #

For most people, not in the strong form. Most people see a fairly mixed news diet. A small, very partisan minority builds real echo chambers, mostly by their own choices.

Would a chronological feed fix things? #

The biggest test so far found it changed time spent and what people saw, including more low-quality sources, but not their attitudes. It is a choice worth having, not a cure.

Did Facebook deliberately promote anger? #

Its documents show it weighted all emoji reactions, including “angry”, above likes from 2017 to promote engagement, and later reduced the angry weight to zero. That is a documented design choice with known side effects, not a proven plan to make people angry.

The lesson #

The machine is not reading your mind. It is reading your attention, and outrage is the cheapest way to hold it. You cannot change the formula, but you can change what you feed it. The minute you notice anger arriving before facts is the minute the formula loses its grip.