Why Your Feed Makes You Angry: The Logic of Engagement Ranking

Nobody at any platform ever wrote a line of code that said make them angry. I want to be precise about this, because the conspiratorial version of the story is both more satisfying and less useful than the truth. What actually happened is duller and stranger: a room full of engineers decided the feed should be engaging, they handed that goal to a system capable of optimising for it relentlessly, and the system went looking for whatever produced the reaction. Engagement algorithms then discovered, on their own and at enormous scale, what every tabloid editor had known for a century — that outrage travels further than nuance. No villain was required. The machinery did it, and the machinery was working exactly as specified.

Most writing about this stops at the business model or at the behaviour it produces. I have written about both — the money side in the attention economy, and the two-in-the-morning thumb-scroll in doomscrolling. What I want to do here is open the box in the middle. How does a ranking system actually decide what you see first? Once you can picture the mechanism, the anger stops looking like a mystery or a plot, and starts looking like arithmetic.

What is actually happening between your thumb and the screen

Start with the scale of the problem the platform is solving. When you open an app, there may be tens of thousands of things it could plausibly show you — posts from people you follow, videos from people you don’t, replies, reshares, ads. It has to put them in one order, in a few hundred milliseconds, before your thumb gets impatient. Every recommender system I have seen described, in published papers and engineering blogs, solves this in roughly three moves.

First, candidate generation. Cheap filters cut the universe down to a few hundred or a few thousand plausible items. Recency, who you follow, what is popular in your area, what is similar to things you have watched. Nothing clever is happening yet; it is triage.

Second, prediction. This is the part that matters, and it is the part almost nobody describes accurately. For every surviving candidate, a machine-learned model is asked a set of questions, and each answer is a probability between zero and one:

• How likely is this person to tap on this?
• How likely are they to watch past three seconds? Past thirty?
• How likely are they to like it?
• How likely are they to comment?
• How likely are they to share it, or send it to someone in a message?
• How likely are they to hide it, or report it, or close the app afterwards?

The model learned to answer these by watching billions of past interactions — yours and everyone else’s. It does not know what the post says in any sense you would call understanding. It knows that items with these characteristics, shown to people with your history, at this hour, produced these behaviours at these rates.

Third, scoring and ordering. Those predicted probabilities are combined into a single number, usually as a weighted sum: multiply each predicted action by how much the platform values that action, add them up, sort the list, render the feed. The weights are the platform’s objective made numeric. If a comment is worth more than a like, comment-provoking posts rise. If a share is worth more than a watch, shareable posts rise. There are re-ranking passes on top — diversity rules, integrity demotions, ad insertion — but the spine of the thing is that sum.

The system is not asking whether a post is true, or fair, or good for you. It is asking one question, several million times a second: how likely is this to make you do something?

Notice what is missing from that list of questions. There is no term for is this accurate. No term for will this person be glad tomorrow that they saw it. No term for does this leave them better informed. Those things are hard to measure and slow to observe, and a ranking system needs a signal it can read in milliseconds from behaviour it can log. Clicks, watches, likes, comments and shares are available, abundant and cheap. So they become the definition of value — not because anyone believes they are the same as value, but because they are what the ledger can count.

Why outrage wins the arithmetic

Now run the mechanism forward and the rest follows almost mechanically.

Of all the behaviours in that list, the ones platforms tend to weight most heavily are the effortful, social ones: comments, reshares, replies. That is a defensible choice. A like is cheap and ambiguous; a comment means you cared enough to type. Weighting comments above likes sounds like a move away from mindless consumption and towards genuine connection, and I think it was often argued in exactly those terms internally.

But ask yourself what reliably makes an ordinary person stop and type. Not agreement — you rarely comment “yes, quite right” under something you already believe. Not subtlety; a careful, qualified argument leaves you with nothing to add. What makes people type is being wrong on the internet — somebody else being wrong, loudly, about something that touches your identity. Moral indignation is the single most reliable comment-generating emotion there is. It produces the reply, the quote-post, the screenshot sent to a friend with three words of commentary, the argument in the thread that keeps the post alive in ranking for another day.

So a model trained to predict comments and shares will, without being told anything about politics or emotion, learn to recognise the textures of content that provokes them. It learns that certain phrasings, certain framings, certain grievance-shaped structures correlate with people typing. It promotes them. Creators watch what gets distribution and make more of it — this is the part people miss, because the algorithm does not merely select from what exists, it teaches an entire professional class of publishers what to produce. Within a year or two of any ranking change, the supply of content has reshaped itself around the new weights.

The result is a system that selects for likely-to-react rather than true or useful, and that trains the people feeding it to specialise in reaction. Nobody chose it. Everybody optimised, and this is what optimisation found.

What the internal research reportedly showed

The reason we can say any of this with confidence about specific companies, rather than reasoning from first principles, is that some of it leaked.

The disclosures brought forward by Frances Haugen in 2021, and reported on extensively by several newsrooms working from the same document set, described ranking in terms recognisably like the ones above. Reporting on those documents indicated that Facebook had retuned its feed around a metric it publicly called “meaningful social interactions”, which weighted comments, replies and reshares more heavily than passive likes; that reaction emoji, including the angry one, were for a period weighted considerably more heavily in that score than a plain like; and that internal researchers subsequently found content drawing angry reactions was disproportionately likely to be low-quality, toxic or misinformation. The angry reaction’s weight was, according to that reporting, later cut.

I am hedging deliberately, and I would ask you to hold these claims at the same level I do. What is well established is the shape of the thing: that the ranking was a weighted combination of predicted interactions, that the weights were chosen by people with reasonable-sounding intentions, that internal research identified harms flowing from those weights, and that the fixes were contested internally before some of them shipped. The exact numbers are secondhand, and I have no interest in quoting internal documents I have not read.

The important part of that episode is not the scandal. It is the confirmation that the harm was legible from inside, in the company’s own data, and that the gap between knowing and changing was a matter of institutional will rather than technical possibility. That gap is the subject of the broader pattern I have traced in how technology gets captured: the tool is not evil, the incentive structure around it bends it, and the bending is visible to the people doing it long before it is visible to us.

In fairness to engagement

Here is where I want to resist the easy version of my own argument. Engagement is not a stupid proxy. Most of the time it is a decent one.

If you watch a twenty-minute video to the end, that is real evidence you found it worth watching. If you send a post to a friend, you probably thought it was worth their time. If you never click anything from a particular source, the system stops showing it to you, and you are better off. A feed ranked purely by recency — the thing people nostalgically demand — is not neutral either. It is ranked by whoever posts most often, which in practice means the most prolific accounts and anyone willing to automate. Chronological feeds were abandoned partly because people found them worse, not only because they earned less.

Engagement is a good proxy for value in the ordinary case and a catastrophic one in the tail — and the tail is where a billion-user system spends most of its consequences.

Every alternative ranking has its own failure mode. Optimise for stated satisfaction and you are trusting survey answers from a small, unrepresentative slice of users who bother to respond. Optimise for “quality” and someone has to define quality, which hands a private company an even more explicitly editorial role than it already has. Let users choose their own ranking and most will accept the default, as people always do. Down-weight anger and you will sometimes down-weight the justified anger of people with real grievances and no other megaphone — which is not a hypothetical risk in a country with our history of who gets heard.

What could actually change

None of that is an argument for leaving it alone. It is an argument for being specific. Four changes seem to me both technically available and worth arguing for.

1. Optimise for stated value, not just revealed reaction. Platforms already run surveys asking whether a post was worth your time. Those answers can be used as a training target, so the model predicts will they say this was worthwhile alongside will they tap. It is more expensive and noisier. It is also closer to what anyone actually wants from a feed.
2. Explicitly down-weight predicted anger and predicted regret. If a model can predict the probability of a comment, it can predict the probability of an angry reaction or of a session that ends with the app closed in irritation. Predicting a signal and then subtracting it is a normal engineering move, not an exotic one.
3. Make ranking a real choice. Chronological, follows-only, and topic-scoped feeds should be first-class options that persist when you close the app, rather than settings that quietly revert. A default is powerful precisely because most people keep it — which is why the option needs to be honest rather than decorative.
4. Auditability. Not publishing the source code, which would help nobody and be gamed within a week, but giving independent researchers structured access to what is being amplified and to what effect. We should not have to wait for the next set of leaked documents to learn what a system serving billions of people is optimising for.

What I keep coming back to is how ordinary the failure is. There is no moment in this story where someone chose cruelty. There is a moment where somebody chose a metric, because a metric was needed and that one was measurable, and then a very powerful optimiser did precisely what it was asked. The feed that makes you angry is not evidence of malice. It is evidence that we let a proxy stand in for a value, at a scale where the difference between them becomes the whole story.

Which is why “stop being so online” is such an inadequate response, and why I do not think the fix is mainly a matter of your self-discipline. You are arguing with an optimiser. It has more data than you, it never gets tired, and it was never asked whether any of this was good for you — only whether you would react. The right question is not how to resist it better. It is who gets to choose the objective, and what they would have to answer for if they chose badly.