← Back to editorials
✍️

Five points for anger: how algorithms learned to sell us the worst of each other

By Rik Van Bruggen · 8 July 2026, at 05:54

Five points for anger: how algorithms learned to sell us the worst of each other

A comparison of two social media reaction buttons: a like worth one point next to an angry reaction worth five points

Let me start with something I've noticed about my own behaviour, because it's a little embarrassing and I suspect I'm not alone. When I post something on LinkedIn - a nice project, a bit of genuine enthusiasm about some technology I love - it does fine. Politely fine. But every now and then I've watched someone post something spiky, a complaint, a bit of outrage, a "can you believe this?" - and it takes off like a rocket. More comments, more shares, more heat, in an afternoon than my sunny little posts get in a week.

For a long time I filed that under "people are just like that." And they are - we spent the last three articles establishing exactly how they are, wired by evolution to lean toward the negative. But there's a newer, sharper part of the story that I've been genuinely itching to understand, and it's the one thing that made me want to build Positron more than anything else. Because our old human negativity bias is one thing. What happens when you hand that bias to a machine whose entire job is to maximise your attention - that's something else entirely.

So this is the article about the machines.

First, the machines really do prefer bad news

Let's establish the basic fact before we get to the how and why, because it's been measured beautifully.

In 2023, a team led by Claire Robertson published a study in Nature Human Behaviour with the wonderfully blunt title "Negativity drives online news consumption". They had access to something close to a researcher's dream: data from Upworthy, which for years ran endless A/B tests on its own headlines - the same story, different words, shown to real people. Around 105,000 headline variations. Something like 5.7 million clicks. Real behaviour, at enormous scale.

The finding was clean and a little damning. For a headline of average length, every single additional negative word raised the click-through rate by about 2.3%. And positive words did the opposite - they pushed clicks down. The kicker is that positive words were actually slightly more common to begin with. It's not that the world only offers bad headlines. It's that the bad ones win the click, over and over, in a way you can measure to the decimal point.

That's the raw material. Now watch what happens when an algorithm gets hold of it.

What actually spreads

Two more studies, to my mind, crack this open.

Three research findings on what spreads online: negative words lift clicks by two point three percent, moral-emotional words lift shares by twenty percent, and posts attacking political opponents are the single strongest driver of sharing

The first is from William Brady, Jay Van Bavel and colleagues, in 2017, and it introduced a phrase I find myself using all the time now: "moral contagion". They looked at more than half a million tweets about hot moral-political topics - gun control, same-sex marriage, climate change - and found that every additional moral-emotional word in a message (words like "attack", "shame", "greed", "evil") increased how far it spread by around 20%. But here's the twist that matters: that spread happened mostly within ideological groups, not across them. Outrage travelled fast, but it travelled in circles - deepening the divide rather than crossing it.

The second study takes it one step further, and it's the one that really lands for me. In 2021, Steve Rathje, Van Bavel and Sander van der Linden looked at 2.7 million posts from news outlets and politicians and asked a simple question: what content spreads furthest? The answer wasn't joy. It wasn't even in-group pride or praise. The single strongest predictor of a post being shared was "out-group animosity" - posts about the other side, the people you disagree with. Talking about your political opponents beat every other factor they measured, including plain old emotional language. Nothing spreads like contempt for "them."

Sit with that for a second. The most shareable thing online is not something good. It is not even something sad. It is anger at the out-group. That's the shape of the raw demand - and the algorithms went looking for exactly that shape.

The smoking gun: five points for anger

Here's where it stops being abstract, because for once we don't have to guess what a platform was optimising for. We have the documents.

In 2021, the whistleblower Frances Haugen released a trove of internal Facebook material, and among the things reported by the Washington Post was a small, revealing detail about how the news feed was ranked. When Facebook rolled out those emoji reactions - love, haha, wow, sad, angry - it decided, starting in 2017, to treat any emoji reaction as worth five times more than a humble "like" in its ranking algorithm. Five points for a reaction, one for a like.

Think about what that quietly does. A "like" is what you give to something nice. But you don't tap "angry" on a puppy photo - you tap it on the thing that outrages you. By weighting reactions five times heavier, Facebook was, in effect, turning up the volume on exactly the content most likely to make people furious. And their own data, a couple of years later, confirmed the obvious: posts that drew a lot of angry reactions were disproportionately likely to contain misinformation, toxicity, and low-quality news. Staff had flagged the risk from the start. The company eventually walked it back - downgrading the weighting, then demoting posts that drew disproportionate anger, and by 2020 cutting the angry reaction's value all the way to zero.

But for years, the machine ran hot. It didn't set out to make the world angrier. It set out to maximise engagement - and it turns out those two things are, disturbingly, almost the same thing.

Let me be fair about this

Now, I promised myself when I started writing these that I wouldn't overclaim, so let me put some honest caveats on the table.

First, most of this is correlation, and correlation is slippery. Showing that angry, negative, out-group content spreads further is not the same as proving the algorithm caused us to become angrier people. Some of it is surely just the algorithm holding up a mirror to what we were already reaching for - the negativity bias from the earlier articles, now simply reflected back at industrial speed. Untangling "the machine made us like this" from "the machine gave us more of what we already liked" is genuinely hard, and serious researchers still argue about the balance.

Second, science self-corrects, and some of these effects look smaller or more conditional under replication than the striking headline numbers suggest. The direction of the findings is robust and repeated. The exact magnitudes deserve humility.

But even with all those caveats, the shape of the thing is clear enough. Our brains lean negative. The content that spreads is negative, moralised, and aimed at "the other side." And the platforms, chasing engagement, built machinery that - deliberately or not - amplified precisely that. Human bias went in one end; an outrage machine came out the other.

Which is the whole reason for the opposite

And this, finally, is why Positron is built the way it is - and why the "how it works" page might be the most important thing on the site.

There is no algorithm on Positron optimising for your engagement, because there is nothing to optimise for. No ads. No feed ranking. No reaction buttons quietly scoring your rage. It doesn't earn a cent whether you're calm or furious, so it has no reason to reach for the thing that makes you furious. It is, on purpose, the exact opposite of the outrage machine: a small, deliberately un-optimised corner of the internet where the only thing sorting the news is "is this actually good?"

That won't fix the machines. I'm one Belgian with a website; I have no illusions. But it can be a place to stand that isn't inside the machine - and honestly, some days that feels like enough.

In the final article of this series, I want to look at what all of this - the bias, the dread, the outrage machine - actually does to us. Because it turns out the research on that is pretty stark, and the research on what a better news diet can do instead is genuinely hopeful.

Cheers / Rik
Share