You're Not Seeing Reality. You're Seeing What Spreads

For a long time I thought the main problem with social media was misinformation. People say false things, other people believe them, and the fix is better facts in more places.

I don't think that anymore. Misinformation is real, but treating it as the central problem assumes the system is basically a pipe: information goes in at one end and arrives, more or less intact, at the other. Fix the leaks and the pipe works.

A feed is not a pipe. It is a selection process, and the thing it selects for is not truth.

Once I started thinking about it that way I had to pull apart three things I had been quietly treating as one. There is what is actually happening. There is what people believe about what is happening. And there is what gets amplified enough for me to encounter it.

Those three are related, but they are not the same, and almost all of my confusion has lived in the gaps between them. The third one is the only one I have direct access to. I keep treating it as though it were the first.

The independence problem

Start with the part that works, because social proof gets dismissed too easily and the dismissal is wrong.

Using other people's observations as evidence is not a cognitive defect. It is one of the most efficient things a person can do. I cannot personally verify the structural safety of a bridge, the competence of a doctor, or the reliability of a piece of software. Nearly everything I know, I know because someone else checked and I had some reason to trust them.

Agreement genuinely carries information, too. If four people who have never met tell me a bridge is unsafe, that should move my belief a long way. Each of them had to look at the bridge and arrive at the same conclusion on their own. For all four to be wrong in the same direction, something unusual has to be going on.

The word doing the work in that sentence is independently.

Take the same four people and change one thing. One of them examines the bridge and the other three repeat what the first one said. The conclusion is identical. The number of people saying it is identical. But the evidence has collapsed from four observations to one observation and three acts of transmission.

Nothing on the surface of the claim tells you which of those two situations you are in.

The version I keep in my head is simpler. One person sees smoke coming from a building and posts that the building is on fire. Ten thousand people repost it.

You do not have ten thousand observations. You have one observation and ten thousand repetitions.

You cannot tell the difference by looking at the number. Ten thousand reposts of one person's mistake and ten thousand people who each walked past the building produce the same count, and from inside the feed they produce the same feeling of certainty.

Consensus is not the problem. Correlated consensus is the problem, and social media manufactures correlated consensus at industrial scale.

How repetition impersonates confirmation

The mechanism isn't mysterious. A claim gets posted. Some people repost it. Some quote-post it with a reaction. Someone writes a longer commentary. Someone summarizes the commentary for people who missed it. Someone disagrees. Someone reacts to the disagreement. Someone writes about the argument itself.

By the end of that chain a single underlying claim may have appeared in front of me twenty times, in a dozen formats, in the voices of a dozen different people, several of whom I have good reason to trust.

None of them checked. That is the part that matters. The claim has been transmitted twelve times and verified zero times, and transmission looks exactly like verification from where I am sitting. Every repetition arrives with the emotional texture of a fresh voice independently agreeing.

So I try to hold two sentences apart:

"I have seen this many times."

"I have seen many independent reasons to believe this."

The first is a fact about my feed. The second is a fact about the world. I used to let the first quietly stand in for the second, and I suspect that substitution is close to universal, because the two produce nearly the same sensation.

Your feed is a sample, and not a random one

Underneath the repetition sits a more basic problem, which is that I am not looking at a representative slice of anything.

At any moment there are millions of things people are saying. My feed shows me a few hundred. The interesting question is not how many I see but how those few hundred got chosen, because any selection rule changes the distribution of what reaches me, and I experience that changed distribution as though it were the world.

If you survey people leaving a football stadium about whether football is popular, you will get a striking result and a worthless one. The selection did the work before the question was asked. Feeds do the same thing continuously, and unlike the stadium, there is no visible moment where the sampling happened.

Stated plainly: how often I encounter an idea is mostly a fact about how well that idea survives selection. It is only loosely a fact about how many people hold it, and looser still about whether it is correct.

A mirror shows you what is in front of it. A filter shows you what got through.

I keep catching myself reading a filter as a mirror.

I want to be careful here, because this is the point where the argument usually turns conspiratorial. I am not claiming anyone designed the system to obscure reality. No intent is required. A selection process with any consistent bias produces a distorted picture whether or not anybody wanted it to, and "the ranking system optimizes for engagement" is a completely sufficient explanation. It is closer to a lens with a manufacturing defect than to a censor.

What the system actually optimizes for

Engagement metrics measure behavior. A like, a reply, a repost, watch time, a click. Each one records that a person did something after encountering a piece of content.

That is a real measurement. It is a measurement of the reader, not of the claim.

Nothing in that process evaluates whether the thing is true, and not because anyone forgot. Truth is not the kind of property a distribution system can cheaply measure. Verifying a claim can take an afternoon, a laboratory, or a decade. Measuring whether someone tapped a button takes a millisecond. The system optimizes what it can see.

Once you optimize for reaction, certain kinds of content are structurally advantaged. Confidence travels further than care, because certainty is easier to consume than uncertainty. Novelty beats accuracy, because a familiar true thing produces no reaction at all. Conflict beats agreement, because disagreement generates replies while agreement generates a like at best. A story beats a statistic, because the story arrives whole and the statistic requires work. Simple explanations beat complicated ones even when the thing being explained is genuinely complicated. And a correction almost never travels as far as the claim it corrects, because learning that something you already believed is slightly more nuanced than you thought is one of the least shareable experiences available.

None of that means viral things are false. This is the distinction I most want to keep hold of:

Visibility and truth are separate dimensions.

Something can be true and spread widely. Something can be false and spread widely. Something can be carefully correct and reach almost nobody. The problem is not that popular things are wrong. The problem is that I am handed a distorted sample and I read the distortion as a measurement.

The feed doesn't only shape what you see

The part I underestimated for years is that the selection process does not stop at content. It works on people.

Post something and watch what happens. Post something else and watch what happens. Do that for a year and you will have absorbed a detailed model of what produces a reaction, whether or not you ever set out to learn it. This is ordinary conditioning running on a reward schedule that pays out within minutes.

First you learn what your audience responds to. Then you start writing with an anticipated response already in mind. Then, eventually, you begin having ideas that arrive pre-shaped by the format that will carry them.

That last step is the one worth noticing, because it happens upstream of anything I would recognize as a decision. I am not choosing to write for the algorithm. The thought simply shows up already fitted to a medium that rewards a particular shape of thinking.

Which means the loop is larger than it first appears:

Platform → what you learn → what you write → how people react → what the platform distributes

The user is not standing outside the selection mechanism receiving its output. The user is a component inside it. Whatever I add to the corpus has already been filtered by my own model of what works, before any ranking system sees it.

I notice this most clearly when I publish something I think is genuinely my best work, it goes nowhere, and I feel the small unmistakable pull to write the other kind of thing next time. That pull is the training signal, and it does not care what I think about it.

When the conversation becomes the object

Something happens. Someone observes part of it. They make a claim about what they observed. Other people react to the claim. Someone characterizes the reactions. People argue about that characterization. Someone writes a considered piece about the argument.

That is seven layers, and only the first one is the world.

Each layer compresses the one before it, and each is written by someone responding mainly to the previous layer rather than to the original event. By the fifth layer the thing generating attention is the discourse itself. People are no longer disagreeing about what happened. They are disagreeing about what one group said about what another group said about what happened.

You can tell you are deep in the stack when the volume of discussion stops tracking anything you could go and check. There is an enormous amount being said, it feels consequential, and there is no observation anywhere near the middle of it.

What makes this hard to see from inside is that the discourse is genuinely real. The arguments are real, the people are real, the stakes to the participants are real. It is a real phenomenon and worth studying. It is simply not the thing everyone believes they are discussing.

Popularity is still information

I do not want to land in the opposite error, where popularity means nothing and obscurity is treated as evidence of depth. That is its own kind of laziness.

Popularity is real information. It is information about a different subject than people assume.

If something reaches a million people, I have learned something. I have learned that many people found it emotionally interesting, or that a framing was unusually compelling, or that a community cares about this more than I realized, or that a claim was surprising enough to be worth passing along, or that people are confused about something and this filled the gap, or that a fight is under way and this made a good weapon in it.

Those are useful facts, and some of them are unavailable any other way. If I want to understand what a group of people care about, how an idea is being framed, or where a live disagreement sits, engagement data is close to the ideal instrument for the job.

Popularity is a reliable measurement of attention and a very weak measurement of reality.

The error is not using it. The error is reading it off the wrong axis.

What a stronger signal looks like

If distribution is not much evidence, it is worth asking what is. The things that have actually changed my mind tend to share a few properties, and none of them have anything to do with reach.

The strongest is proximity. A primary source rather than a description of one. The paper rather than the post about the paper. The dataset rather than a screenshot of a chart someone built from it. Every layer of intermediation is a place where something gets dropped, sharpened, or reframed, usually without anyone lying at any point.

Then independence, which is the earlier problem wearing different clothes. Two sources that agree are only worth more than one if they did not get it from the same place. So the useful question about a second source is not whether it agrees but where it got its information.

Then firsthand evidence. Someone who built the thing, ran the experiment, worked in the industry for six years, or simply tried it and reported what happened, including the parts that did not work.

Mechanism matters more than assertion. If someone explains why something should be true, I can evaluate the explanation on its own terms. If they can only assert that it is true, all I can do is decide whether to trust them. An explanation gives me something to check. A claim gives me something to believe.

Falsifiability is the one I find most diagnostic in practice. Asking what would change someone's mind separates people who hold a model from people who hold a position. Someone with a model usually answers immediately. Someone with a position often cannot answer at all, and the shape of that failure tells you a lot.

Closely related is what happens when I deliberately go looking for the strongest case against something I already believe. Not the weak version that is easy to dismiss, the strongest one. If I have never done that, I do not really know what I think. I know what I encountered first.

And track record, which is slow to build but cheap to keep. Not whether someone is clever or confident, but whether the specific things they said would happen actually happened. Almost nobody online is scored this way, which is exactly why it is so informative.

Very little of this correlates with distribution. Some of it is negatively correlated. A careful, mechanism-heavy, appropriately uncertain explanation is precisely the kind of thing that performs badly.

Questions, not beliefs

The practical change that has helped me most is small. I try to let the feed produce questions rather than conclusions.

A viral claim used to leave me with a belief. Now the good outcome is that it leaves me with "is that actually true, and how would I find out?" A technical post becomes "how does that work?" A surprising number becomes "where did that come from, and what exactly was counted?" A disagreement becomes "what are these people actually disagreeing about, and is it a fact or a definition?" Someone showing an experiment becomes "could I reproduce that?"

It sounds like a minor reframe and it changes the whole activity. A conclusion ends an inquiry. A question starts one. And since most of these questions are answerable with twenty minutes of effort somewhere other than the feed, the habit naturally pushes me out of the feed, which is most of the benefit.

The failure mode of the opposite habit is that you accumulate a very large number of confident beliefs about subjects you have never investigated, every one of which feels like your own.

Reading a noisy sensor

The metaphor I keep returning to is instrumentation, because it gets the attitude right.

A noisy sensor is not a broken sensor. Every real instrument has noise in it. Nobody throws away a thermometer because the reading fluctuates. You learn the instrument instead: how it fails, which direction it drifts, what conditions make it unreliable, how much weight a single reading deserves.

The practices are well understood. Take multiple measurements rather than one. Check genuinely independent instruments rather than reading the same one twice. Treat an extreme reading as a question rather than a fact, because outliers are frequently the instrument and not the world. Carry an estimate of uncertainty alongside the number instead of pretending the number is exact. And above all, keep the measurement and the measured thing in separate compartments.

That is the discipline I want with a feed. It is a real instrument pointed at a real system: what people are saying, what they react to, what a ranking system predicts they will engage with. That system is genuinely worth measuring.

But it is an instrument with known and severe biases, and the readings are not the system. The failure is not trusting it too much or too little. The failure is forgetting you are reading an instrument at all.

The bottleneck has moved

Information used to be expensive and hard to reach. That problem is solved to a degree that is difficult to fully appreciate from inside it.

Models generate text on demand. Anyone can publish. Platforms distribute at no marginal cost. Retrieval is instant. The supply is effectively unlimited and still increasing.

So the scarce resource has moved. It is no longer access. It is judgment about what deserves attention.

That is a different skill from the one the old environment rewarded. When information was scarce, the valuable ability was finding it. When information is abundant, the valuable ability is deciding what to ignore, noticing the difference between something merely interesting and something actually informative, and being able to update on evidence rather than on volume.

I do not mean this nostalgically. I would not go back. But I notice that nobody taught me the second skill, and I need it constantly.

How I want to use it

This is roughly where I have landed, and it is a posture more than a system.

When I open a feed I am not trying to find out what everyone thinks. That question has no good answer, and the answer I would get is an artifact of the instrument anyway.

I am looking for a much smaller set of things. People doing real work, so I can follow them instead of the topic. Claims interesting enough to be worth an hour of checking. Primary sources I did not know existed. Ideas that do not fit my current model of something, which are rare and valuable precisely because my model resists them.

The measure of a good session is not that I found out what happened. It is that I left with better questions than I arrived with, and a short list of things to go and look at somewhere else.

The feed can tell me where to look. It shouldn't decide what I believe.

You're not seeing reality. You're seeing what spreads.