You didn’t choose what you believe today. Something else did — one paused video at a time, and it’s still choosing, right now, as you read this.

A few weeks ago someone you know probably went down a hole. Maybe it started with one video about a “miracle” diet that made a bold claim about a food you eat every day. Maybe it was a horoscope reel that felt eerily specific. Maybe it was a movie’s first-look clip that seemed to be everywhere for a week, then vanished. Within days, their feed had quietly rearranged itself. It wasn’t a decision they made. It was a decision made for them, by a system that was watching how long they paused.

This is not a story about any political party, election, or ideology. It’s a story about a piece of infrastructure that now sits between nearly every person and the information they use to make decisions about their health, their money, their beliefs, and their vote — infrastructure that is fundamentally different from anything that came before it, and more dangerous than most of us have registered.

The old gatekeepers, for all their flaws, were visible

For most of the last century, information reached people through a small number of chokepoints. A newspaper editor decided what ran on the front page. A television network decided what aired at nine o’clock. Even the earliest viral communication tool most of us used daily — a WhatsApp forward — worked the same crude way: someone chose a message and sent it, identically, to everyone in the group.

These gatekeepers had real biases and real failures. But the filtering was a single, visible layer. You could name the editor. You could distrust the channel. You could ask “who sent this to me and why.” Two people reading the same newspaper, or sitting in the same WhatsApp group, saw the same version of events, argument, and error alike. Disagreement happened over a shared set of facts, even when the interpretation diverged.

What changed is not that platforms show us content. It’s that they show each of us a different reality, decided by a machine that has no idea what it’s showing

Modern recommendation systems — the ones powering Instagram Reels, TikTok’s For You page, YouTube Shorts — don’t have an editor, a bias, or an opinion. They have a scoreboard. The single metric that matters is whether you kept watching, whether you watched it twice, whether you sent it to someone privately. That’s it. The system has no concept of “this diet advice is dangerous” or “this astrology claim is unfalsifiable” or “this political claim is misleading.” It only knows what kept your thumb still.

A recent qualitative study of TikTok users who ended up trapped in eating-disorder content describes the mechanism with unsettling precision.1 Participants hadn’t searched for this material. They watched one video — sometimes recovery-oriented, sometimes just curious — and the system read that pause as interest. Within days they described “falling down the rabbit hole,” using the exact words “you watch one video… and then you just get more and more and more.” Many said the platform’s own “Not Interested” button stopped working once their behavior had already told the algorithm otherwise — the system trusted what they did over what they said they wanted.

The same mechanism, applied to a different topic, explains why one relative’s feed becomes wall-to-wall extreme diet content, another’s becomes an unbroken stream of horoscope reels, and a third’s becomes saturated with a single film’s promotional clips for a week straight. It’s not that the platform “believes” in astrology or thinks that film deserves the airtime. It’s that engagement compounds on itself, regardless of what the content is, and nobody is in the loop to ask whether it should.

Why this is a genuinely different kind of danger, not just a faster version of the old one

It’s invisible. There’s no editor to name, no channel to distrust. Most people don’t experience their feed as “curated” — they experience it as “what’s out there.” That makes the filtering far harder to notice, let alone question.

It’s personalized down to the individual, not the audience. A newspaper’s front page was the same for a million readers. A recommendation feed is different for every single person, shaped by their own past clicks. Two people can no longer assume they’re arguing from the same set of facts, because there is no longer a shared “the news” — there are millions of private, algorithmically-assembled versions of it.

It manufactures false confidence. This is the part that should worry people most. A controlled study on personalized information feeds found that people exposed to algorithmically narrowed content sampled far less information overall, developed measurably distorted understandings of the topic — and were more confident in their wrong conclusions, not less.2 A narrowed feed doesn’t feel narrow from the inside. It feels like “I’ve clearly seen enough of this to know.” That’s precisely the psychological signature behind a friend who’s suddenly certain about a fad diet, a family member convinced by a horoscope-driven decision, or anyone who’s stopped questioning a claim after weeks of a feed that only ever agreed with them.

Sophistication is not protection. A 2025 study on algorithmic awareness found that people with high professional or occupational education were, surprisingly, among the least aware of how personalization actually shapes what they see.3 Being sharp in your career does not automatically make you literate about why your feed looks the way it does. Nobody gets to assume “this can’t happen to me” on the basis of intelligence alone.

The speed and scale have no historical precedent. A newspaper story took a news cycle to spread. A single Reel can be algorithmically pushed to millions of screens within 48 hours of posting, entirely without any human editorial decision approving that reach at any point in the chain.

Where it’s already doing damage

This shows up far beyond any one domain. Health and wellness content is one of the clearest cases — genuine anxiety about food, appearance, or illness produces exactly the rewatch-and-share behavior that trains a feed to escalate alarm-framed claims, regardless of whether they’re medically sound. Astrology and horoscope content spreads unusually fast because it’s built for identification — “this is about me” — which drives the sharing behavior the algorithm rewards most. Film and entertainment promotion rides the same wave: a hype clip that gets early traction gets algorithmically amplified into the feeds of anyone with adjacent interest, manufacturing a sense of inescapable buzz that isn’t really consensus, just compounding. And yes — political communication across the spectrum, everywhere in the world, is adapting to the same incentive. Academic research analyzing political parties’ most-viewed short-form video content has found a consistent pattern: platforms reward “sensationalist spectacle and affective, emotionally stimulating content” over substantive communication, regardless of which party or which country is posting it.4 This isn’t a story about one side gaming the system better than another. It’s a story about every actor, in every domain, being pulled toward the same shallow, high-arousal content by the same underlying incentive.

This is not a call to panic. It’s a call to notice — and act, at three levels

What you can do personally, starting today. The best-replicated finding in this entire field is almost anticlimactic: simply pausing to ask “is this actually accurate” before you engage measurably changes what gets amplified to you.5 Researchers found this wasn’t really about ideology — most people share low-quality content out of inattention, not conviction, and a single moment of reflection was enough to raise the quality of what got shared, even in live, real-world tests. Related research on deliberate “friction” — closing an app after specific content instead of letting it autoplay, consciously searching out a source instead of accepting whatever’s served, repeatedly telling the platform “show me less of this” — shows the same effect: slowing down interrupts the loop that speed is designed to protect.6

What technology could be built differently. The deepest fix isn’t more persuasive content competing for the same attention — that just adds a second combatant to the same arms race and tends to make things worse. Researchers working on “bridging-based ranking” have proposed and, in limited cases, built systems that reward something different: not “what maximizes engagement” but “what resonates with people who normally disagree.”7 X’s Community Notes is a live example — a note only earns visibility if people across different viewpoints rate it helpful. It’s proof that a fundamentally different value model is buildable, not just theoretical.

What policy can require. The EU’s Digital Services Act is the first real regulatory attempt to legislate against the mechanism itself. For the largest platforms, it mandates that recommendation logic be explained in plain terms, and — more importantly — that a genuine non-personalized, non-algorithmic feed option must exist and be reachable, not buried in settings.8 No equivalent rule exists in most of the world yet, but it’s proof that this is a solvable regulatory problem, not an unavoidable cost of being online.

The bottom line

Print media, television, and even WhatsApp forwards could mislead people — but they misled everyone roughly the same way, in the open, by a filter that had a name. What we’re living through now is different in kind: a personalized, invisible, engagement-optimized machine that doesn’t share our reality with anyone else, quietly manufactures confidence in whatever it feeds us, and does all of it faster and at a scale no prior communication technology ever approached. Recognizing that difference is the first and most necessary step. Everything else — the pause before sharing, the deliberate friction, the push for better-designed platforms and real regulation — only works once enough people actually see the wall that’s been built around them.

— Elango Raghupathy
Filmmaker · Founder, Karkei & ERA Foundation · elangoraghupathy.com

Sources