How the Algorithm Actually Decides What You See Next
"The algorithm" gets talked about like a mysterious, almost sentient force with opinions about you - it "knows" you, it's "watching" you, it "wants" you to keep scrolling. The actual mechanism is less mystical than that framing suggests, and understanding the plain version of it is more useful than the mysterious one, because the plain version explains exactly why the feed behaves the way it does.
It Doesn't Need to Understand Anything
A recommendation system doesn't need to know what a video is "about" in any meaningful sense to recommend it well. What it needs is behavior data: how long you watched, whether you watched to the end, whether you rewatched it, how fast you swiped away, whether you liked or shared it. It compares your specific pattern of these signals against the patterns of millions of other users, finds people whose behavior looks statistically similar to yours, and surfaces what those similar users engaged with that you haven't seen yet. This is the same basic principle - collaborative filtering - that powers a "customers who bought this also bought" recommendation, just running continuously, on video, with far more granular behavioral signal than a single purchase.
The One Number That Explains Almost Everything
The single most consequential design decision in any recommendation system isn't the machine-learning technique - it's the choice of optimization target: the one number the system is built to maximize. For most short-form platforms, that number has historically been some version of engagement - watch time, session length, return-visit frequency - rather than content accuracy, user wellbeing, or anything resembling "quality" in a human sense. This choice matters enormously, because a system built to maximize a specific number will find and exploit whatever reliably moves that number, with total indifference to whether the content doing the moving is good for the person watching it.
Engagement became the default target for practical reasons more than ideological ones: it's directly measurable at massive scale, it correlates cleanly with advertising revenue, and it doesn't require any human judgment call about what counts as "good" content - a judgment that's genuinely hard to encode and easy to dispute. Watch time is just a number. Numbers are easy to optimize.
Nobody Programmed It to Be Upsetting
It's worth being precise about the causality here, because it's less a conspiracy than an emergent statistical fact. Nobody wrote a rule that says "show more content that makes people anxious." What happened, across enough billions of data points, is that the optimization process discovered - empirically, the same way any curve-fitting process discovers patterns - that certain categories of content reliably produce longer watch times: surprising content, emotionally intense content, content that invites social comparison. The system doesn't have an opinion that this content is good. It has a correlation between this content and the metric it's built to maximize, and it acts on that correlation relentlessly, without any capacity to weigh the correlation against anything else.
Why It Gets More Effective the Longer You Use It
Because the system is optimizing against your individual behavioral signal, not a generic population average, its model of what keeps you specifically watching improves with every session. Your first hour on a platform gets a reasonably generic feed. Your five-hundredth hour gets a feed shaped by five hundred hours of evidence about exactly what makes you personally pause, rewatch, or linger - which means the system isn't just getting better at recommending in general, it's getting more precisely tuned to your particular triggers over time. That's not a bug in the system or a sign it's trying harder. It's what curve-fitting looks like with a continuously growing dataset that happens to be your own attention.
Why "Just Pick Better Content" Doesn't Work Around It
A common piece of advice is to actively curate the feed - follow better accounts, engage deliberately with higher-quality content, train the algorithm toward what you'd actually choose. This works only partially, and it's useful to understand why. The system isn't optimizing toward what you'd deliberately choose if asked; it's optimizing toward whatever measurably holds your attention longest, and those two things overlap less than people assume. You can follow entirely thoughtful, high-quality accounts and still find the system, over time, surfacing more of whatever specific sub-category within that content reliably produces the longest watch times - the most dramatic clip from an otherwise measured creator, the most emotionally charged post from an otherwise balanced one. Deliberate curation shapes the topic. It has much less power over the emotional register the system keeps gravitating toward within that topic, because the register, not the topic, is what the optimization target actually tracks.
What Would Actually Have to Change
The only lever that meaningfully changes this pattern at the source is the optimization target itself - what number the system is built to maximize. A platform that optimized for, say, self-reported satisfaction after a session, rather than time spent during one, would very plausibly produce a different feed entirely, favoring content people feel good about having seen over content that's merely hard to look away from. Some platforms have experimented with signals like this at the margins. None have made it the primary target, for a fairly direct commercial reason: watch time correlates with ad revenue in a way that "did this feel worthwhile afterward" doesn't, at least not as cleanly or as measurably. Understanding this doesn't hand an individual user the power to change the target. It does explain precisely why individual-level curation, however well-intentioned, is working against a much larger structural pull rather than redirecting it.
Why "Personalized" Doesn't Mean "Chosen"
The word "personalized" carries a flattering implication - that the feed reflects your interests, your taste, something like a considered choice made on your behalf. The mechanism described here suggests a more accurate word would be "conditioned": the feed reflects what has reliably kept you watching in the past, which overlaps with genuine interest but isn't the same thing as it. A person can be genuinely interested in cooking and still find their feed drifting toward the most dramatic, most conflict-laden cooking content available, not because that's what they're actually most interested in, but because that specific register within the topic tested best against their engagement pattern. "Personalized" describes the precision of the targeting. It says nothing about whether the target is something you'd have picked yourself, given an honest choice outside the feedback loop.
Where That Leaves an Individual User
Arguing with the algorithm's "judgment" is a category error - there's no judgment to argue with, only a statistical process responding to whatever data you feed it. The more workable lever is reducing how much data the loop gets to refine itself against, specifically inside the feeds built most aggressively around this exact mechanism. Dam It does that narrowly - it doesn't try to out-argue the recommendation engine inside YouTube Shorts, Reels or TikTok, it just limits how many additional cycles of "watch, learn, refine" the system gets to run against you in a single sitting.