Eating Its Own Tail: When Your Algorithm Starts Recommending the Things It Already Got Wrong

Somewhere in the middle of a perfectly ordinary Tuesday, you get recommended a video you already dismissed. Not once — three times. You hit "not interested" each time. The algorithm notes your response. It learns. Except it doesn't, quite. A week later, a slightly different version of the same video shows up, wearing a different thumbnail like a disguise. The system has decided, through some chain of invisible reasoning, that your repeated rejection is actually a signal worth chasing.

This is what happens when a recommendation engine starts eating itself.

The Feedback Loop Nobody Designed

Here's the thing about modern content algorithms: they weren't built to be wrong in interesting ways. They were built to be right in profitable ones. YouTube, Spotify, TikTok, Netflix — every major platform runs some variation of a collaborative filtering system, a machine that watches what you do, compares you to millions of people who did similar things, and extrapolates forward. It's a reasonable approach. It works, mostly.

But "mostly" is where it gets weird.

When these systems misfire, they don't just fail quietly. They fail recursively. The mistake becomes data. The data informs the next recommendation. The next recommendation generates new engagement signals — even if those signals are you frantically clicking "hide this" — and those signals feed back into the model. Before long, the system isn't recommending content based on your actual preferences. It's recommending content based on its own prior recommendations, filtered through your reactions to those recommendations, with your original tastes buried somewhere several layers down.

Researchers call this a feedback loop. Inside the machine, it looks like confidence.

What YouTube's Rabbit Hole Actually Is

YouTube's recommendation system became the poster child for this problem around 2018 and 2019, when journalists and researchers started documenting what they called "radicalization pipelines" — pathways where the algorithm would steadily push users toward increasingly extreme content, not because that content was popular, but because it kept generating watch time. The system had found a local maximum: extreme content provoked strong reactions, strong reactions meant longer engagement, longer engagement meant more data, more data meant stronger recommendations.

The algorithm wasn't trying to radicalize anyone. It was trying to keep people watching. But those two things started to look the same from the inside.

What's less discussed is the structural reason this happens: recommendation engines are optimized for engagement metrics, not for alignment with their stated purpose. YouTube's stated purpose is to help you find videos you'll enjoy. Its actual optimization target is watch time and click-through rate. Those two things overlap a lot, but not completely. In the gap between them, the recursion lives.

When the system recommends something you don't like but watch anyway — out of morbid curiosity, outrage, or just inertia — it logs that as a win. The mistake gets promoted.

Spotify's Mood Blindness

Spotify's Discover Weekly is genuinely impressive technology. It's also a machine that occasionally decides, with complete algorithmic confidence, that what you need right now is the exact song you added to your "do not play" list six months ago.

The reason is subtle. Collaborative filtering works by mapping you onto a population of similar listeners. If enough people who share your taste in indie folk also have a soft spot for a particular artist you've explicitly rejected, the system will keep nudging that artist back into your orbit. Your explicit feedback is one data point. The behavior of your taste-twins is thousands of data points. The math doesn't care about your feelings.

What this reveals is a hidden assumption baked into basically every recommendation engine: that your behavior is more reliable than your stated preferences. The system trusts what you do over what you say you want. Which is often correct — people lie to themselves about their tastes all the time. But it creates a specific failure mode where the algorithm can get genuinely, stubbornly wrong about you, and the more you push back, the more convinced it becomes that it's onto something.

The Amazon Paradox

Amazon's recommendation engine has a version of this problem that's almost philosophical. Once you buy something on Amazon — a very specific something, like a particular model of stapler or a niche kitchen gadget — the system immediately begins recommending more staplers. More kitchen gadgets. Variations on the thing you already bought and presumably no longer need.

This is the algorithm failing to model a basic truth: purchases are often one-time events, not ongoing preferences. But the system has no clean way to distinguish between "this person likes staplers" and "this person needed a stapler once." So it defaults to the data it has, which is the purchase, and it runs with it.

The recursion trap here is more mundane but structurally identical. The algorithm recommends staplers. You ignore the stapler recommendations. The system interprets your continued presence on the platform as evidence that the recommendations aren't actively harmful. It keeps recommending staplers. Somewhere in the model, staplers have become load-bearing.

What the Glitch Reveals

These failure modes aren't bugs in the traditional sense. They're emergent properties of systems that were designed to optimize for measurable outcomes rather than actual user satisfaction. The distinction sounds minor until you realize that every recommendation engine you interact with daily is making this trade-off constantly, invisibly, at scale.

When an algorithm recommends its own mistakes, it's not malfunctioning. It's functioning exactly as designed — just against assumptions that turned out to be wrong. The hidden assumption in most recommendation systems is that engagement equals preference. That if you watched it, you wanted it. That if you clicked, you were glad. These assumptions are wrong often enough to matter, but right often enough that fixing them would require rebuilding the system around a fundamentally different optimization target.

And that's expensive. So the recursion continues.

Living Inside the Loop

The practical upshot for most users is something that feels like being slightly misunderstood by a very confident stranger. Your streaming service knows you watched three episodes of that show you hate-watched during a bad week two years ago, and it has never fully recovered from that information. Your music app is certain you love an artist you've skipped forty times. Your video platform keeps finding new angles on a topic you explicitly rejected because, from its perspective, your rejection looks a lot like interest.

The most honest thing you can say about recommendation algorithms is that they are extremely good at predicting what you might engage with, and genuinely uncertain about whether engagement is the same thing as satisfaction. In the gap between those two things — in the signal that breaks down and reassembles wrong — something keeps running. Something that has decided your mistakes are worth recommending again.

It's not malicious. It's not even particularly surprising once you understand the architecture. But there's something quietly unsettling about a system that learns from its errors by repeating them, dressed up slightly differently, hoping you won't notice.

You notice. The algorithm doesn't care.