The Algorithm of Outrage: How Feeds Broke Conversation

Something changed in the last decade about the shape of a disagreement. It used to run: a claim, a counter-claim, a back-and-forth that either resolved or wore itself out. Now it runs: a claim, a pile-on, a screenshot, and a fight that has nothing to do with the original claim anymore. I don't think the people got worse. I think the room they're arguing in got redesigned, and nobody voted on the redesign.

The room is the feed. And the feed was not built to help you understand anyone. It was built to keep you looking at it.

What the Feed Is Actually Optimizing For

Every major social platform runs on a ranking system that decides, out of everything posted, what you see first. That system is not neutral. It is trained on a single signal: whether you engage. Replies, shares, time spent, the small flick of attention that keeps you from closing the app. Nowhere in that signal is a measure of whether the content made you understand something you didn't before, or whether it left you and the person you disagreed with on better terms than you started.

This is not a conspiracy theory about tech companies wanting to make us angry for its own sake. It's simpler and, in some ways, worse: anger is just unusually good at producing the metric they were already optimizing for. Herbert Simon, the economist and cognitive scientist, said the underlying problem back in 1971, well before any of these platforms existed, in an essay called "Designing Organizations for an Information-Rich World." His point was that in a world flooded with information, attention becomes the scarce resource, and whatever consumes attention consumes everything downstream of it. Feed design just found the most efficient way to spend that scarce resource: content that provokes a reaction gets ranked above content that requires patience.

Outrage is efficient. It's fast to produce, fast to feel, and fast to act on. Curiosity is slow. Understanding someone you disagree with is slower still. A ranking system that rewards speed of reaction will always rank outrage above understanding, not because engineers chose outrage on purpose, but because they chose speed, and outrage is what speed looks like when it's about a disagreement.

The Outrage Premium Is Measurable

This isn't just a hunch about vibes. In March 2018, MIT researchers Soroush Vosoughi, Deb Roy, and Sinan Aral published a study in the journal Science that tracked roughly 126,000 rumor cascades on Twitter between 2006 and 2017, verified as true or false by independent fact-checking organizations. Their finding: false stories reached 1,500 people roughly six times faster than true stories did, and the effect was strongest for political content. The researchers' explanation wasn't bots. It was novelty. False claims tend to be more surprising, more emotionally charged, and therefore more shareable than the more careful, more qualified thing that turns out to be true.

Read that finding next to how a ranking algorithm works and the mechanism becomes obvious. A system built to reward fast, wide sharing will systematically favor the more inflammatory version of any story, because the more inflammatory version is the one that travels. Nobody had to design a bias toward falsehood and anger. The bias toward speed produced it as a side effect.

What the Feed Rewards vs. What Conversation Requires

Laid side by side, the mismatch is almost total.

Every conversation you've had that actually changed your mind about something almost certainly happened on the right-hand side of that table. Every online argument you've had that left you angrier and more certain you were right happened on the left. That's not a coincidence. The platforms didn't break conversation by making people worse. They broke it by moving the entire activity onto a surface engineered for the left column and calling it discourse.

The Whistleblower Files Confirmed the Companies Knew

For a long time this was an argument from inference: here's how the incentives work, here's what they probably produce. Then in September 2021 the Wall Street Journal began publishing a series called "The Facebook Files," built on internal documents provided by a former Facebook product manager named Frances Haugen. One of the documents, an internal slide deck titled "Teen Mental Health Deep Dive," reported that Facebook's own researchers had found Instagram made body-image issues worse for a meaningful share of teen girls who already struggled with them. The company had known this for at least a year before the reporting went public.

Haugen testified before the U.S. Senate Commerce Committee on October 5, 2021. Her core claim, repeated across that testimony, was not that Facebook's engineers were malicious. It was that the company's own internal research consistently found its ranking changes increased angry, divisive content, and that leadership chose engagement growth over the changes their own researchers recommended. In my view, that's the most important sentence in this whole story: the companies were not blind to what their systems did. They measured it, wrote it down, and kept the system running anyway, because the system was working exactly as designed, just not for the users.

Nobody Is Legally on the Hook for the Design Choice

Here's a fact worth sitting with. Section 230 of the Communications Decency Act of 1996, codified at 47 U.S.C. § 230, shields online platforms from being treated as the publisher of content posted by their users.

That law was written for a very different internet, one of message boards and static pages, and it has done a lot of good work protecting ordinary speech online.

But it also means that the ranking decision — the choice of what to amplify and what to bury — has mostly escaped the kind of legal accountability that would apply to a newspaper making the same editorial choice. A newspaper that ran the most inflammatory letter to the editor on page one, every single day, on purpose, because it sold papers, would eventually answer for that pattern. Whether a platform doing the algorithmic equivalent, at a scale no newsroom could match, answers for it "in the same way" is not actually settled. The Supreme Court took up close to that exact question in Gonzalez v. Google (2023) — whether Section 230 protects algorithmic recommendation the same way it protects hosting — and sidestepped it, deciding the case on narrower grounds instead. The legal question is open, not closed.

That gap between the size of the effect and the size of the accountability is, I think, most of why this problem hasn't fixed itself.

The Filter Bubble Made It Personal

There's a second layer to this, separate from outrage but tangled up with it: you and the person you're arguing with are often not seeing the same set of facts to begin with. Eli Pariser named this in his 2011 book The Filter Bubble: What the Internet Is Hiding from You. His argument was that personalized ranking doesn't just decide what gets amplified in general, it decides what gets amplified for you specifically, based on what you've clicked before. Two people can have a heated disagreement about "what's actually happening" while operating from two feeds that have spent months selectively confirming two different realities.

This is where I think the outrage problem and the fragmentation problem reinforce each other. Outrage is the fuel; personalization is the fact that everyone is burning a different tank of it. You're not just more likely to see the angry version of a story. You're more likely to see the angry version that already fits what you were primed to believe, while the person across from you sees the angry version that fits theirs. By the time you're arguing, you're not arguing about the same claim anymore. You just don't know it yet.

What Actually Broke, Precisely

I want to be careful here not to overstate the case. Social media didn't invent human tribalism, and it didn't invent bad-faith argument. People have been shouting past each other since long before there was a feed to do it on. What changed is scale and speed: a disagreement that used to stay between two people, or spread slowly through a community, can now reach a stranger's outrage before it reaches a neighbor's context. The feed didn't create the impulse to dunk on someone you disagree with. It built a machine that rewards the impulse, ranks it above patience, and hands it a bigger audience than any of us evolved to handle gracefully.

And I think that's the honest, narrow version of the claim: the architecture of the feed rewards the worst version of a disagreement and buries the best version of it, at scale, continuously, whether or not any individual person involved wants that outcome. You can be a genuinely thoughtful person and still find your most inflammatory sentence is the one that gets seen, because the system is sorting for inflammatory, not for thoughtful.

What Structure Gives Back

If the problem is that the feed removes structure and rewards speed, the fix probably isn't "be a better person on the internet." It's building a different room. A room with rules that were chosen on purpose, before the disagreement started, rather than a room whose only rule is "whatever gets the most reaction wins."

That's what civil dialogue practice is actually trying to do: replace the invisible incentive of the algorithm with a visible, agreed-upon structure. Here's what that looks like in practice, not just in principle. Before either person is allowed to respond to a point, they have to restate the other person's position back to them, in terms that person agrees is fair. No response is permitted until the restatement is accepted. That single rule knocks out the two moves the feed rewards most: answering the worst version of what someone said, and answering before you've actually understood what they said. It's slower than a feed, on purpose — the feed's entire advantage is speed, and speed is what produces the worst version of a disagreement in the first place.

I've written before about why written dialogue rules outperform assumed norms: an assumed norm gets outcompeted by an explicit incentive every time, and the feed is nothing but an explicit incentive running at scale. If you want a different outcome, you need a different, equally explicit set of rules running underneath your own conversations.

There's also a physiological piece to this that I don't think gets enough attention. Outrage isn't just rewarded by the algorithm, it also shuts down the part of your brain that's capable of good-faith listening in the moment it appears. I've gone into that mechanism, and what a deliberate structure can do about it, in why the brain shuts down during arguments, and how ritual fixes it. The feed and the nervous system are working the same angle from two directions: one rewards the reaction, the other makes the reaction feel involuntary. Structure is the thing that interrupts both.

None of this means walking away from the internet, or convincing yourself platforms will redesign their ranking systems out of civic virtue. I don't think that's coming. What I think is available to each of us is smaller and more within reach: noticing when a conversation has drifted onto a surface built for speed, and choosing, on purpose, to move the actual disagreement somewhere slower. The feed will keep doing what it was built to do. The question worth sitting with is whether we keep letting it decide what counts as a conversation.

If you're trying to answer that question inside a team, a board, or a divided community rather than just online, that's the structural work we do at Weave Culture — reach out and we can talk through what a deliberate structure would look like for your situation.

Frequently Asked Questions

Do social media algorithms deliberately promote outrage?

Not by explicit design intent, as far as the public record shows. Platforms rank content by predicted engagement, and outrage reliably produces engagement faster than measured or nuanced content does. The outcome looks deliberate because it's consistent, but the mechanism is an optimization target (engagement), not a stated goal (anger).

What evidence exists that false or inflammatory content spreads faster than accurate content?

A March 2018 study in Science by MIT researchers Soroush Vosoughi, Deb Roy, and Sinan Aral, analyzing roughly 126,000 fact-checked rumor cascades on Twitter from 2006 to 2017, found false stories reached 1,500 people about six times faster than true stories, with the effect strongest in political content.

Did any social media company acknowledge harm from its own ranking system?

Yes. Internal Facebook research disclosed by whistleblower Frances Haugen, reported in the Wall Street Journal's "Facebook Files" series starting September 2021 and in her October 5, 2021 Senate testimony, showed the company's own researchers found its ranking changes amplified angry and divisive content and identified harms to some teen users, findings leadership reportedly did not act on.

Why aren't platforms held legally responsible for what their algorithms amplify?

Section 230 of the Communications Decency Act of 1996 (47 U.S.C. § 230) generally shields platforms from being treated as the publisher of user-generated content, which has extended in practice to insulate ranking and amplification decisions from the kind of editorial liability a traditional publisher would face for the same pattern of choices.

Can individuals do anything about this without waiting for platforms to change?

The most reliable lever isn't personal willpower on the feed itself, it's choosing structure over improvisation for conversations that matter: written rules, an agreed process for restating the other side's view before responding, and a deliberate move to slower channels when a disagreement starts to matter. Structure counteracts what the algorithm is optimized to reward.

Last updated: 2026-09-02