Over the past few months, I found myself wondering a question I never expected to ask.

Why does it feel like the world suddenly dislikes Indians?

It wasn’t something I had experienced personally. It wasn’t based on conversations with colleagues or friends across different countries. It was simply — what my social media feed seemed to suggest. Video after video, comment after comment, a pattern that felt undeniable because it kept showing up.

Then a second question occurred to me, one that turned out to be far more interesting than the first.

What if my feed wasn’t showing me reality at all?

That question is the reason I’m writing this.

A Question Everyone with a Smartphone Has Asked

How many people have reached strong conclusions about the world based on a few hundred videos chosen by an algorithm they’ve never seen and don’t understand?

I don’t ask that rhetorically. I ask it because I think most of us have done exactly this — about politics, about crime, about a rival, about a whole country — without ever noticing that the sample of “the world” we were reacting to was never the world at all. It was a selection. And selections are made by something, for a reason.

What the Algorithm Is Actually Optimizing For

Most people assume recommendation algorithms are designed to show us what’s important.

They’re not.

They’re designed to predict what will keep us engaged for a few more seconds. That’s the entire objective function. Not truth, not balance, not representativeness — engagement. Everything downstream of that single design choice follows logically, even when the outcome feels irrational.

Strip away the mystery and the pipeline is fairly simple:

Millions of posts
        ↓
Candidate Generation
        ↓
Ranking Model
        ↓
Top 20 appear in your feed
        ↓
You engage
        ↓
Model updates
        ↓
Repeat, thousands of times

Out of millions of possible posts, a candidate generator narrows the field to a manageable set. A ranking model scores that set based on what’s predicted to hold your attention. The top handful reach your screen. You engage — even a pause counts as a signal. The model updates. And the cycle repeats, thousands of times a day, quietly training itself on you.

None of this requires a conspiracy. It requires only that the system does exactly what it was built to do.

The Insight This Article Is Really About

Here’s the sentence I’d want you to remember after you close this tab:

AI doesn’t change reality. It changes the sample of reality that reaches you.

That distinction is subtle, but it’s the whole story.

Human beings don’t perceive the world directly — we perceive it through the information available to us, and we build our sense of “what’s normal” and “what’s happening” from that sample. This is how perception has always worked, long before algorithms existed. What’s new is that the sample itself is now actively, continuously, invisibly curated by a system optimizing for something other than accuracy.

If the sample is skewed, our conclusions become skewed — even when we’re reasoning perfectly logically from what we’ve seen. The failure isn’t in our thinking. It’s upstream of our thinking.

How the Loop Reinforces Itself

The mechanism that makes this so powerful is a feedback loop:

Interest
   ↓
Engagement
   ↓
Recommendation
   ↓
More engagement
   ↓
Stronger recommendation
   ↓
Perceived prevalence
   ↓
Stronger belief

You linger on something once — maybe out of curiosity, maybe out of irritation — and the system reads that as interest. It shows you more. You engage again. The system strengthens its confidence. Soon what began as a handful of videos looks, from inside your feed, like an overwhelming pattern. Not because it is one, but because the loop has manufactured the appearance of one, tailored specifically to you.

Nothing in this loop requires malicious intent.

Every component is working exactly as designed.

Yet the emergent behavior can dramatically distort our perception of reality. This is, in a strange way, an engineering lesson as much as a social one: complex systems can produce outcomes nobody intended without any single component ever malfunctioning.

It Was Never Only About One Topic

Today it might be discussions about Indians.

Tomorrow it could be politics. Climate change. Crime. Immigration. AI replacing jobs. Financial markets. Whichever group, belief, or fear happens to generate the most engagement in a given season.

The mechanism is the same regardless of the subject. Once the algorithm learns what captures your attention, it quietly reshapes your perception of how common that thing really is — how many people believe it, how urgent it is, how representative it is of “everyone.” The topic is interchangeable. The mechanism is not.

That’s what makes this worth understanding, even after the specific example that prompted it has faded from the news cycle.

A Shared World, Quietly Splitting

Here’s the part that unsettles me more than any single distorted belief.

If my feed is curating a sample of reality for me, then everyone else’s feed is curating a different sample for them — shaped by their own pauses, their own clicks, their own three-second hesitations. We’re not all looking at slightly different windows onto the same world anymore. We’re each being shown a world custom-built to hold our attention, and none of those worlds are required to overlap.

Two people can sit across from each other, both certain they’re simply “paying attention to what’s going on,” and be reacting to almost entirely different realities — each internally consistent, each backed by hundreds of confirming data points, each invisible to the other. That’s not a disagreement about values. It’s a disagreement about facts, produced by systems that were never trying to produce facts in the first place.

A shared world is the quiet infrastructure underneath every other kind of agreement — political, social, even personal. It’s not something we’ve historically had to build; it was just the water we shared, by default, because we were all reading from the same limited set of sources. That default no longer holds. And a society that can no longer agree on what’s happening has a much harder time deciding what to do about it.

The Question Worth Sitting With

Perhaps the most important AI system in our lives isn’t the one generating answers.

It’s the one deciding which questions we ask in the first place.

If an algorithm controls the sample of reality we consume, then understanding recommendation systems isn’t just a technical skill anymore. It’s becoming a prerequisite for thinking clearly — for anyone, not just engineers.

I didn’t start with an opinion about my feed. I started with a feeling that something didn’t add up, and it turned out the feeling was pointing at the right thing — just not in the direction I expected. Not at the world. At the window I was using to look at it.