There is a specific feeling that has become familiar to anyone who uses the internet in 2026.

You mention something to a friend — out loud, not typed, not searched. Offhand. You’re thinking about getting a new couch, or you’ve been having a vague anxiety about your teeth. Within a few hours, your feed is full of furniture ads and dentists. You know the rational explanation. You’ve probably already read three articles debunking it. You know your phone isn’t listening. And yet the feeling doesn’t go away, because the rational explanation — that the algorithm inferred it from a hundred other data points you don’t remember generating — is, if you sit with it, not actually more comforting than the alternative.

The surveillance mythology of the modern internet runs on that gap between the rational explanation and the feeling. And somewhere in that gap, a much larger story took hold: not just that the algorithm knows what you want to buy, but that it knows what you want to think — and that it’s been quietly steering you there for years.

The Genealogy

The idea didn’t begin with AI. It began with a word that has almost entirely disappeared from public conversation, which is worth noticing: filter bubble.

Eli Pariser coined the term in a 2011 TED talk and the book that followed it. The argument was precise and relatively modest: personalization algorithms, by optimizing for engagement, would gradually construct for each user a distinct information environment — one that showed them more of what they already agreed with and less of what challenged them. Not a conspiracy. Not intentional manipulation. An emergent property of systems optimized for click-through rates.

That framing, careful as it was, didn’t survive contact with 2016.

Cambridge Analytica is the hinge point. The story — that a firm had harvested the Facebook data of 87 million users and used it to build psychographic profiles for targeted political advertising, first for the Ted Cruz campaign and then for Trump and the Brexit Leave campaign — landed differently than any previous tech scandal. It had the shape of a thriller. There was a whistleblower (Christopher Wylie, in a pink fur coat, with receipts). There was a villain (Alexander Nix, caught on hidden camera promising kompromat and Ukrainian escorts). There was a documentary, The Great Hack, which Netflix made available globally. And there was a ready-made vocabulary: microtargeting, psychographic manipulation, dark posts.

Whether Cambridge Analytica actually delivered what it claimed is a separate and genuinely contested question. The company’s methodology was largely considered junk science by academic researchers, and there’s little evidence it produced the results it sold. But the story had already become a cultural artifact before anyone had evaluated the evidence, and what it left behind was a template: the idea that your political opinions were not arrived at but delivered, via a system you couldn’t see, running on data you didn’t know you’d provided.

What the Troll Farms Actually Did

The Internet Research Agency is not mythology. The Senate Intelligence Committee’s 2019 report documented in exhaustive detail how a St. Petersburg-based operation with a $35 million annual budget and several hundred employees ran coordinated influence campaigns across Facebook, Instagram, Twitter, and YouTube between 2013 and 2017, targeting American voters across the political spectrum — not to push a single ideology but to amplify division, stoke distrust in institutions, and make the ambient sense of social fracture louder.

The IRA’s actual reach was significant. Their Facebook content reached an estimated 126 million Americans. Their Instagram posts reached 20 million. Their Twitter operation generated millions of impressions.

What the IRA did not do — and this distinction got very quietly lost — was change minds at scale. The research that followed the revelations found something more complicated and less satisfying than the narrative required: exposure to the IRA’s content was real, but its persuasive effect on political attitudes was, as one large-scale study from NYU’s Center for Social Media and Politics put it, “minimal.” The people most exposed to the content were already the most politically engaged and partisan — the least movable.

The troll farms were real. The manipulation mythology that grew out of them — the belief that ordinary people’s opinions had been installed by unseen foreign operators — was a significant overstatement of what the evidence showed. But overstatement was not a bug. Overstatement was the product.

How “Algorithm” Became “AI”

For a few years after 2018, “the algorithm” served as the master term. The algorithm radicalized people. The algorithm pushed you toward extremism. The algorithm decided what was real. YouTube’s recommendation engine came in for particular scrutiny after a 2018 New York Times piece suggested it was funneling viewers from mainstream content toward increasingly extreme material — the so-called “rabbit hole” problem. Subsequent research complicated that picture significantly, but the rabbit hole had already become the defining metaphor.

Then the large language models arrived, and the vocabulary shifted.

The shift matters because “the algorithm” was at least comprehensible as a mechanism — a set of rules that optimized for engagement. “AI” is not comprehensible as a mechanism to most people, by design and by complexity, and incomprehensibility is extremely useful for a certain kind of narrative. The new version of the claim isn’t that your engagement patterns are being nudged by a recommendation engine. It’s that something intelligent — something that understands you, that has goals, that is actively working — is shaping what you’re allowed to think.

The posts that carry this claim are not concentrated on any single platform, which is itself part of what makes them interesting. They appear in the form of earnest explainer threads on X, in deadpan TikTok monologues shot in cars at night, in Discord servers dedicated to “breaking free from the matrix,” in Substack newsletters, in the comment sections of mainstream news articles. The specific claim varies. Sometimes it’s that AI-generated content is flooding the zone to make it impossible to find real information. Sometimes it’s that your feeds are being individually tailored to push you toward specific political conclusions. Sometimes it’s a vaguer, more ambient claim: that something is wrong with what you’re being shown, and AI is why.

The Folklore Layer

Here is where it becomes a robotjeans problem.

The “AI is controlling your opinion” narrative has the structural properties of folklore, not journalism. It is:

Unfalsifiable in practice. Any evidence that contradicts the claim can be incorporated as further evidence of the claim. If you feel like you’re thinking freely, that’s what a successfully manipulated person would feel. If you find research suggesting algorithmic effects are overstated, that research was probably funded by someone with an interest in keeping you from knowing the truth. The claim is airtight because it has no outside.

Self-spreading through the thing it describes. The most viral version of the “AI is managing your opinions” content spreads via the same algorithmic feeds it’s warning you about. It gets shared, boosted, recommended. It performs extremely well as engagement content because it produces the most reliable emotional response the algorithm knows how to find: the feeling that you’ve been shown something they didn’t want you to see. Every share is evidence of its own premise.

Explanatory of everything. Like the best folkloric monsters, it accounts for any situation. Why do people disagree with you? The algorithm. Why did the election go the way it did? The algorithm. Why does your uncle believe things that seem impossible to you? The algorithm found him and fed him a diet of content specifically designed to produce someone exactly like him. The explanation is frictionless because it requires no engagement with the actual mechanisms of how people form beliefs.

This is not a new pattern. It’s the same architecture as the Slender Man mythology — a figure that explained a generalized anxiety about what lurked in the information environment, given a shape and a name, passed person to person until the original context was almost irrelevant. It’s the same hunger that made Cicada 3301 into an obsession: the conviction that somewhere behind the surface of normal information, something intelligent and purposeful was running a different game.

The difference is that this time, something intelligent and purposeful actually is running a game. It’s just not the game the mythology describes.

The Uncomfortable Part

Recommendation algorithms do shape what you see. Personalization does produce different information environments for different people. Foreign influence operations are documented and ongoing. Generative AI is being used to produce low-quality content at scale, and some of that content is designed to deceive. These things are real.

What they are not is a unified, coherent, intentional system specifically aimed at you. The algorithm doesn’t have goals in the way the narrative requires. It has an objective function — engagement — and it pursues that function without care for what the engagement is about or where it leads. The system that shows you content designed to make you angry about immigration is the same system that shows you content designed to make you angry about corporate greed, or content designed to make you angry about content that makes people angry. It doesn’t prefer any particular outcome. It prefers engagement, which anger produces reliably, and anger is a reliable feature of almost every strongly-held opinion regardless of content.

The uncomfortable part is that “an optimization system that amplifies whatever produces strong emotional responses, without regard for truth or consequence” is both more accurate and less satisfying than “AI is controlling your opinions.” It doesn’t give you an enemy. It doesn’t give you a moment of revelation. It doesn’t explain why the other people believe what they believe in a way that resolves anything.

The mythology is more useful than the reality. That’s usually why mythologies spread.

The Ouroboros

The final thing worth sitting with is this: you encountered this piece through a feed. Something decided to surface it to you — or someone shared it, and something decided to show you that share, in that moment, above other things you might have seen instead. The fact that what you’re reading is a skeptical account of the AI opinion-control mythology doesn’t exempt it from the conditions it’s describing.

The pipeline is real. The specific nightmare at the end of it — the invisible hand, the system that knows what it wants you to think and has been quietly delivering it — is a story the internet tells itself, in the same feeds, through the same mechanisms, optimized for the same engagement metrics as everything else in there.

The algorithm didn’t install your opinions. But it has been watching which ones keep you scrolling.

That distinction is probably important. It’s just harder to make a viral post about.