The Manipulation Infrastructure

In Part 1, I described the surveillance camera: the system that records your behavior every time manipulated content crosses your screen.

In Part 2, I showed you what’s on the footage. Your emotional triggers. Your attention decay rate. Your skepticism threshold. Your correction resistance. All of it stored, modeled, and ready.

Given all of that, you might now be wondering, “Who’s behind this?”

It’s the right question but it’s also the wrong frame. And the reason it’s the wrong frame is the most important thing to understand in this entire series.

Before I go further, I want to be precise about what we actually know and where it comes from, because this matters for everything that follows.

Platforms don’t publish their ranking logic. They don’t disclose what their algorithms optimize for, how behavioral profiles are built, or what the feedback loops look like internally. What we know comes from three sources: what independent researchers can measure from the outside, what former employees have described, and what companies have disclosed — usually under legal pressure or in leaked documents they never intended to make public.

Where I’m drawing on documented evidence, I’ll say so. Where I’m making an inference from observable outputs — from what we can see the system doing — I’ll say that too. The distinction matters. But I’d also argue that the opacity itself is part of the story. The fact that we have to piece this together from the outside is not a gap in the evidence. It’s evidence.

The Most Tempting Question Is a Dead End

Here’s how the search for orchestration usually goes.

Something manipulative goes viral. People notice. Investigators start pulling threads — who posted first, which accounts amplified it, whether the money traces back somewhere interesting.

Sometimes they find it. Sometimes they don’t. When they do, a specific bad actor gets the blame. When they don’t, the conclusion is often: this wasn’t coordinated, so maybe it wasn’t so bad.

Both conclusions miss the point.

The question “who coordinated this?” assumes that coordination is what produces manipulation. That somewhere there’s a room full of people deciding what you should believe. Pull back the curtain, expose the room, and the problem is solved.

There is no room. There doesn’t need to be.

The manipulation happens automatically. Not because a conspiracy is well-hidden, but because the infrastructure produces it as a default output — the same way a factory produces waste. You don’t need anyone to plan the pollution. You just need the factory to keep running.

What Algorithms Actually Optimize For

An algorithm’s job is straightforward: keep you on the platform. That’s it. Every feature — the feed, the recommendations, the notifications — exists in service of that goal. More time on platform means more ads. More ads means more revenue.

To keep you on the platform, the algorithm learns what holds your attention. And it turns out — because this has been measured, replicated, and published — that the content types that hold attention longest are not the most accurate, or the most informative, or the most important.

They’re the most emotionally activating.

Content that makes you angry holds your attention longer than content that makes you calm. Content that confirms your existing beliefs gets shared faster than content that challenges them. Content that triggers fear, outrage, or moral indignation performs better — by every metric the platform cares about — than content that requires careful thinking.

This isn’t speculation. In 2018, researchers at MIT published the largest study of its kind — analyzing 126,000 stories shared on Twitter over eleven years by roughly three million people. False news spread farther, faster, deeper, and more broadly than true news across every category they measured. The effect was most pronounced for political content. And critically: it wasn’t bots driving the spread. It was people. The researchers found that false stories were more novel and triggered stronger emotional reactions — fear, disgust, surprise — than true ones. The system didn’t create that preference. It just learned to feed it.

Think back to the Coldplay video. It had visible AI artifacts — fused fingers, warped elbows, unnatural skin tones. The information needed to question it was in the clip itself. It didn’t matter. The emotional charge overrode everything else. The algorithm didn’t cause that response. But it was designed to select for exactly that kind of response, over and over, because that response keeps you scrolling.

Optimize for engagement long enough, and you’ve optimized for emotional manipulation. They’re not two different things. They’re the same thing with different labels.

The algorithm doesn’t know or care about manipulation. It only knows what you clicked, how long you stayed, and whether you came back. But the result of a decade’s worth of that optimization is a system that has become extraordinarily good at pushing your emotional buttons — not because anyone designed it to, but because that’s what the data rewarded.

The Factory Doesn’t Need a Manager

Here’s the analogy I want you to consider.

Imagine a factory built to maximize output. No one at the factory is trying to pollute the river downstream. The workers show up, do their jobs, and go home. The executives track margins. Nobody in a meeting says, “Let’s poison the water.”

But the factory pollutes the river anyway. Because that’s what happens when you build that kind of factory and run it at maximum capacity without controls. The pollution isn’t a decision. It’s a consequence of the design.

Social media platforms are that factory. The river is your information environment.

The engineers who built the engagement algorithms weren’t trying to create a manipulation machine. They were solving an attention problem. The executives who set engagement targets weren’t trying to radicalize anyone. They were trying to hit quarterly numbers. Nobody in those rooms said, “Let’s optimize for outrage and make it impossible to tell real from fake.”

But that’s what the factory produces. Reliably. At scale. Whether or not anyone intends it.

We don’t have to guess about this dynamic. In 2021, former Facebook product manager Frances Haugen leaked tens of thousands of internal documents to the SEC and the Wall Street Journal — documents that became known as the Facebook Papers. Among the things those documents revealed: Facebook’s own internal research had identified that engagement-based ranking was amplifying emotionally provocative content, and that the company had repeatedly weighed that finding against its growth targets and chosen growth. Haugen said it plainly in an interview with 60 Minutes: “No one at Facebook is malevolent, but the incentives are misaligned. Facebook makes more money when you consume more content. People enjoy engaging with things that elicit an emotional reaction. And the more anger that they get exposed to, the more they interact.” That’s not an outside critic’s assessment. That’s a description of the factory’s design from someone who worked on the inside.

This is why proving intent is a dead end. Even if every person who ever worked on these algorithms had only good intentions — and some clearly did — the output would be the same. The factory runs regardless.

Every Post Is a Test

Part 2 established that behavioral data gets harvested with every viral moment. Here’s the mechanical loop that makes it so difficult to escape.

When content goes viral, whether real or synthetic, the platform doesn’t just count how many people saw it. It records the entire behavioral signature. Who shared it first. Which networks it spread through. How long different people spent on it. Who fact-checked it. Who got angry. Who forgot about it in 48 hours.

That data goes back into the model.

The model updates. Not because a person reviewed it — because that’s what the system does automatically. It processes behavioral data and adjusts predictions. What will this specific user engage with next? What emotional frame? What messenger? What time of day?

The next piece of content is deployed with that updated model informing everything: who sees it, how it gets amplified, where it shows up in the recommendation stack.

Then the behavioral signature of that post feeds back into the model.

Repeat.

Every post is simultaneously a message and an experiment. Your behavior automatically adjusts the next experiment. You’re not a passive recipient of content. You’re an unwitting participant in a continuous trial.

I want to be clear about what I’m describing here. The specific mechanics of how these feedback loops work — the exact weighting, the precise triggers, the granularity of individual behavioral profiles — are not public. Platforms treat this as proprietary. What we can observe is the output: what gets amplified, what disappears, what content consistently outperforms accurate content across years of documented behavior. The inference from outputs to mechanism isn’t a leap. It’s the same logic epidemiologists use when they trace a disease back to a source they can’t directly observe. You don’t need to see the gears to know the clock is running.

And no one has to be watching. The system learns, adjusts, and deploys the next iteration without human intervention. The manipulation becomes more sophisticated over time not because anyone is working to improve it, but because that’s what optimization loops do.

Someone Is Exploiting This. That’s Not The Point.

A reasonable person reading this will say: okay, maybe the platform isn’t running a conspiracy. But surely someone is taking deliberate advantage of this?

Yes. Obviously.

State actors use these systems. Political operatives use them. Bad-faith influencers use them. Disinformation campaigns actively exploit these dynamics.

But here’s why that doesn’t change the analysis.

The deliberate exploitation works because the infrastructure exists. The manipulator doesn’t build the manipulation machine — they plug into it. They benefit from the vulnerability maps the algorithm has already drawn. They use the emotional response data the platform has already collected. They deploy content through the virality pathways the system has already optimized.

Stopping the deliberate exploiters — even if that were possible — doesn’t stop the infrastructure from producing the same vulnerability maps tomorrow. Automatically. Without any human making a decision.

When you focus on the bad actor, you create the impression that removing the bad actor solves the problem. It doesn’t. The factory doesn’t need that particular manager. It runs itself.

“No Conspiracy” Is Not Good News

I want to head off the most common misunderstanding of this argument.

When I say the infrastructure produces manipulation without coordinated human direction, I’m not offering comfort. People hear “no conspiracy” and their shoulders drop. Oh, it’s just how the system works. No one’s targeting me specifically. I can relax.

That not how this works.

The absence of conspiracy is not reassurance. It’s actually the most alarming version of this story.

A conspiracy has edges. There are co-conspirators to identify, operations to disrupt, human decisions to reverse. A conspiracy is bounded.

An automated optimization system has no edges. There’s no decision to reverse. There’s no operation to disrupt, because there’s no operation — just a system running the way it was built to run. Every attempt to regulate it runs into the issue that the behavior emerges from the interaction of millions of individual decisions, no single one of which is obviously the problem.

You can’t arrest an algorithm. You can’t prosecute an engagement metric. You can’t prove that “maximize time on platform” was criminal intent, because it wasn’t intent at all. It was a design choice with catastrophic emergent effects.

This is much harder to fix than a conspiracy. And the people who benefit from inaction know that “no evidence of coordination” sounds like “no problem.”

It isn’t.

The Architecture Is the Agency

For the past decade, debates about online manipulation have centered on agency. Who did this? Who decided? Who benefits? These are important questions in some contexts. But as the primary frame for understanding the manipulation problem, they’ve been a distraction.

The right frame is architecture.

The design of these platforms — engagement optimization, behavioral profiling, automated amplification — produces manipulation as a structural output. Not as a side effect. Not as an accident. As the predictable consequence of how the system was designed and what it was designed to do.

Your vulnerability map exists in the data whether anyone plans to use it or not. Your skepticism threshold has been measured whether anyone intends to exploit it or not. The feedback loop is running whether anyone is watching the results or not.

You’re in the data. Not hypothetically. And the profile built from your behavior gets more accurate every single day you use these platforms — every viral moment you engage with, every story you share before checking, every outrage cycle you ride and then forget.

The infrastructure doesn’t need a conspiracy. It is, by itself, all the evidence we need. And here’s what that means for you specifically: there’s a model somewhere that knows more about how you form beliefs than you do. Not what you think but how you come to think it. That model didn’t require anyone’s malicious intent to build. It only required your attention, which you’ve already given.

Which raises the question: what do we do about a machine that manipulates by default?

In Part 4 we’ll look at what this means for democracy — and what can actually be done about infrastructure-level manipulation.

For clear, evidence-based analysis of how AI and media are reshaping politics and reality, subscribe here.