The Algorithm Told Me I'd Love This. It Was Wrong. Again.

There's a specific kind of frustration that has become a defining feature of modern digital life. You open Netflix after a long day, genuinely wanting to find something good to watch. Forty-five minutes later, you've scrolled through three rows of content you've already seen, seven thumbnails of the same actor's face, and a mysterious "Top 10 in the U.S. Today" list that seems to have been assembled by someone who has never met a human being. You give up and rewatch The Office. Again.

This is not a personal failure. This is a systems failure — and it's happening across every major platform simultaneously.

The Promise Was Enormous

Cast your mind back to the early days of algorithmic recommendation. Netflix's famous $1 million prize for improving their recommendation engine. Spotify's Discover Weekly, which felt — when it launched in 2015 — like someone had finally cracked the code on musical taste. The early TikTok "For You Page" that seemed to read your mind with an almost unsettling accuracy.

The pitch was seductive: give us your data, your time, your attention, and we will learn you. We will surface the exact thing you didn't know you needed. We will eliminate the paralysis of infinite choice by making the right choice obvious.

For a brief, shining window, it sort of worked. And then the platforms got bigger, the incentives shifted, and the whole thing quietly broke.

What Happened to Discover Weekly

Spotify's Discover Weekly is a useful case study because its decline has been so well-documented by users. In the early years, the playlist felt genuinely revelatory — obscure artists, unexpected connections between genres, music that matched your taste without just repeating it. People talked about it the way they used to talk about a friend with great taste.

Now? Most users report a playlist that circles the same twenty artists they already know, occasionally interrupted by a song they've explicitly disliked before. The discovery engine has become a familiarity engine.

The reason is structural. As Spotify's catalog expanded and its advertising business matured, the recommendation system started serving multiple masters. Labels pay for promotional placement. Spotify's own podcast content gets surfaced aggressively. The algorithm that was supposedly learning your taste was also quietly learning how to sell you things — and those two goals are not the same.

TikTok's For You Page Is Eating Itself

TikTok built its entire identity on the For You Page — the idea that even a brand-new account with zero followers would immediately see content tailored to its interests. It was genuinely impressive. The FYP felt alive in a way that Instagram's feed, bloated with ads and influencer partnerships, never did.

But TikTok's advertising model has metastasized into the recommendation layer in ways that are increasingly hard to ignore. The platform now carries so much sponsored content, creator fund pressure, and brand partnership material that the signal-to-noise ratio has collapsed. More importantly, TikTok's algorithm optimizes for completion rate and re-watch rate — metrics that measure engagement but not satisfaction. You can watch a video three times because it's genuinely delightful, or because you're confused, or because it made you angry. The algorithm doesn't know the difference. It just sees the number.

This is the core dysfunction at the heart of every major recommendation system: engagement and satisfaction are not the same thing, and platforms have systematically chosen to optimize for the former while claiming to deliver the latter.

Netflix and the Thumbnail Industrial Complex

Netflix has a different problem, though it stems from the same root. Their recommendation engine is famously sophisticated — and famously bad at helping people find things they actually want to watch.

Part of this is scale. Netflix's catalog is enormous, and the platform's algorithm has to serve hundreds of millions of users across wildly different taste profiles. But a bigger issue is that Netflix's recommendation system has become entangled with its content strategy in ways that distort the results.

Netflix heavily promotes its own original content, regardless of whether it matches your taste. The platform has also become notorious for its A/B testing of thumbnails — showing you different artwork for the same title depending on your viewing history, trying to find the image most likely to make you click. This sounds helpful until you realize it means the platform is optimizing for the click, not for whether you'll enjoy what comes after it.

The result is a homepage that feels personalized but is actually a carefully constructed sales floor. The "recommended for you" row is doing promotional work as much as curatorial work, and users have started to sense the difference.

Choice Overload Was Supposed to Be the Enemy

The original promise of algorithmic curation was that it would solve the paradox of choice — the well-documented psychological phenomenon where too many options leads to worse decisions and lower satisfaction. Give people a curated selection, the theory went, and they'll be happier than if you gave them everything.

Instead, we have platforms with thousands of titles, millions of tracks, and billions of videos — and recommendation engines that have made the experience of navigating them more stressful, not less. The algorithms haven't reduced the cognitive load of choosing. They've just added a layer of noise on top of the abundance.

There's also a deeper issue: these systems are trained on past behavior, which means they're structurally incapable of introducing genuine novelty. They can find you more of what you've already liked. They cannot easily show you something you've never encountered before and help you understand why you'd love it. Discovery — real discovery — requires a kind of intuitive leap that data models struggle to replicate.

The Human Curation Comeback

Interestingly, the failure of algorithmic recommendation has sparked a quiet revival of human curation. Newsletter-based recommendation services are booming. Letterboxd — a social film logging app with a distinctly human, community-driven feel — has exploded in popularity, particularly among younger users who are exhausted by Netflix's homepage. Music blogs, once declared dead, are finding new audiences. Substack writers who do nothing but recommend books or movies have built substantial followings.

People are, in other words, going around the algorithms. Finding other humans to trust with the task that the machines promised to handle.

This isn't a full solution. Human curation doesn't scale the way algorithmic systems do, and it carries its own biases. But the trend reveals something important: what users actually want from a recommendation isn't just accuracy. It's taste. They want to feel like something was chosen for them by someone who understood what they were looking for, not just someone who tracked what they clicked.

The Fix Nobody Wants to Build

The frustrating reality is that better recommendation is technically achievable. Systems that ask users direct questions about mood, context, and intent. Algorithms that weight satisfaction signals — like finishing something and immediately seeking more like it — over raw engagement. Curation models that separate "what the platform wants you to watch" from "what you actually want to watch."

None of the major platforms are building these things at scale, because their business models depend on keeping the two categories blurry. An algorithm that perfectly serves user satisfaction might recommend less sponsored content, fewer originals, and more obscure titles that don't generate licensing revenue. That's not a great quarter.

So the algorithm keeps getting it wrong. And you keep rewatching The Office.

At least that one never lets you down.