Instagram's CEO Just Admitted Something. Let's Talk About What It Actually Means.

Adam Mosseri published something last week that's worth more attention than it got.

It was framed as a feature announcement. Instagram is rolling out something called Your Algorithm, which lets you see the topics the system thinks you're interested in, and adjust them. Fine. New feature. Moving on.

Except there was a paragraph in the middle that stopped me. And I keep going back to it.

"For years ranking models have been built with technologies that aren't legible to people — no human can read a neural net and explain why an algorithm thought you might be interested in a given video."

He said it like it was obvious. Like everyone already knew. But I don't think everyone knows what that actually means. Because if you unpack it... it's a pretty significant thing to admit.

So let's do that. Let's actually unpack it.

First, what's a ranking model

You open Instagram. You follow maybe 300, 400 accounts. There are also millions of accounts you don't follow, churning out content every minute. Way more than you could ever scroll through.

Something has to decide what you see. That's the ranking model. It's the system that looks at the entire queue of possible content and sorts it, in real time, specifically for you.

And it's not sorting alphabetically. It's not sorting by time posted. It's sorting by what it predicts you'll engage with. Based on everything you've done before. Every video you watched all the way through. Every post you scrolled past without stopping. Every reel you sent to a friend at 11pm. The system is watching all of it, all the time, building a picture of you.

That part most people vaguely understand. The part that's less clear is how the system actually holds that picture internally. Because this is where it gets strange.

The map nobody told you about

Here's something called embeddings. Bear with me for a second because this is actually fascinating once it clicks.

When the ranking model looks at a piece of content, it doesn't think about it the way you do. It doesn't go "oh, this is a menswear video." It converts that content into a long string of numbers. Thousands of numbers. A coordinate in a mathematical space so large you couldn't visualize it if you tried.

Every piece of content gets its own coordinate. You get one too, built from your behavior. And the model's whole job is to find content coordinates that are close to your coordinate. Close in this space means "this person will probably engage with this."

Mosseri actually describes this in his post. A photo of someone's outfit sits near a video from their stylist, which sits near a cluster the system has quietly labeled "men's fashion." But here's the thing: nobody wrote that label. Nobody sat down and said "put fashion content here." The system grouped those things together because users who engaged with one tended to engage with the other. The category emerged from the pattern. The label came after.

And none of this is visible to you. There's no column that says "this user likes fashion." It's all encoded in the relationships between numbers inside a neural network that doesn't have an instruction manual. You can look at the math. You can't read it.

Why "we can't tell you" was always the honest answer

This is what makes neural networks different from older recommendation systems.

The older ones were more like rules. "User watched three videos tagged X, so show them more videos tagged X." You could trace the logic. You could explain it in a sentence.

Neural networks don't work like that. The reason the system recommended something to you isn't stored anywhere legible. It's distributed across millions of parameters, none of which individually mean anything. It's a bit like asking why a crowd moved in a particular direction. No single person decided. It emerged from all of them interacting at once.

So when you've ever wondered why Instagram showed you something strange... the honest answer was always that even the people who built it couldn't fully tell you. Not because they were hiding it. Because the system's reasoning doesn't exist in a form humans can read.

Mosseri puts it well: "the conversation with the system became one-sided." It was learning from you constantly. You had no real way to tell it what you actually wanted. You could only behave and hope it drew the right conclusions.

That's not a small thing. That's billions of people, for years, interacting with a system that was shaping their experience in real time, with no legible way to push back.

What actually changed

So why can Instagram do this now, when it couldn't before?

Large language models. Specifically, their ability to look at those clusters of numbers and describe them in plain language.

That coordinate that used to just be a point in a space no human could read... an LLM can now examine it and say: "based on what's in this cluster, this looks like men's fashion content." The number becomes a label. The label becomes something you can show a user. And suddenly the user can say: yes, more of that. Or: no, remove it, I don't know why you think I care about that.

The underlying ranking model hasn't changed. The neural network is still doing what neural networks do. What's new is the translation layer. An LLM sitting between the system's internal language and yours, interpreting.

It's genuinely useful. It's also worth being precise about what it is and what it isn't.

But is this actually control?

Here's where I want to slow down. Because not everyone is convinced this is the meaningful shift Mosseri is framing it as.

Social Media Today put it bluntly: the feature "is also likely to be a PR win, in that it addresses users' desire to manage their in-app experience." And then, almost in the same breath: "People won't use it, of course, but having it as an option seems to quell a lot of chatter on this front."

That's a pretty sharp observation. The bet Meta is making, and they know their users well, is that most people will never actually open their algorithm settings. They'll feel better knowing the option exists. But the feed will keep running on autopilot, because convenience always wins.

There's also a deeper critique from users themselves. In the comments on Mosseri's original post, someone asked a question that stuck with me: "If the Reels tab and the main feed are designed to show us topics we're interested in, when do we actually interact with the people we follow?"

That's the part the feature doesn't address. You're not getting control over the ranking model. You're getting a window into what it thinks about you, and some limited tools for nudging it by topic. The system is still deciding who surfaces and who doesn't. The people you chose to follow are still competing against millions of algorithmically recommended accounts, and the algorithm still holds most of the cards.

Research on this dynamic is worth noting. A 2023 study published in the journal New Media & Society found that users who felt they had more control over their social media feeds reported higher satisfaction, but this effect was largely driven by the perception of control rather than actual behavioral change in how they used the platform. The feeling of agency and the reality of agency are not the same thing.

The bigger question Mosseri raised and didn't answer

The most interesting part of Mosseri's post isn't the feature. It's the paragraph near the end.

He describes a future where AI could generate entirely different app experiences for each person. Not just different content, but different structures, different purposes, different interfaces. A version of Instagram built in real time, just for you.

And he flags the tension himself: if everything is personalized, what happens to shared experience? What happens to the cultural references that connect people? "If AI can generate entire apps and experiences that each of us wants, you and I might no longer share any sense of space."

Researchers studying algorithmic polarization have been raising this concern for years. Eli Pariser's concept of the "filter bubble," introduced in his 2011 book The Filter Bubble: What the Internet Is Hiding from You, argued that personalization algorithms create invisible barriers around users, narrowing the range of perspectives they encounter. A decade later, the evidence is more complicated than Pariser's original framing suggested, but the core concern hasn't gone away. As personalization gets more precise, the question of what we share with each other gets harder to answer.

Mosseri's optimistic response is that people genuinely want shared experience, and that pull is durable. "We have access to nearly infinite music, and a lot of us still listen to the same songs." It's a reasonable point. But it also sidesteps the structural question: if the platform is actively optimizing for individual engagement above everything else, shared experience has to fight for space within a system not designed to protect it.

What's worth holding onto

The black box is starting to open. Not all the way. Not in ways that give you anything like real control over what shapes your feed. But enough that the conversation is shifting from "trust the algorithm" to "here's what the algorithm thinks about you, and here's how you can push back a little."

That's different from where we were five years ago. And it matters, even if the feature itself turns out to be something most people never use.

What I'd want to see next isn't more topic toggles. It's transparency about the thing Mosseri acknowledged but didn't resolve: that the ranking model's primary optimization target is still engagement, not wellbeing, not connection, not the quality of the time you spend there. Until that changes, every tool for user agency is operating within a system whose incentives don't fully align with yours.

Knowing what the algorithm thinks you want is a start. Knowing what it's actually optimizing for is the more important question.