How Recommendation Algorithms Actually Work — and What That Means for Your Feed
Your feed isn't neutral, but it's also not the simple villain in most articles about it. Here's what the systems actually do, from someone who's read the papers and built smaller versions of them.
💡 Did You Know?
Most major recommendation systems predict not just clicks but a basket of engagement actions — watch-through, like, share, comment, follow — and combine them with learned weights into one ranking score. The weights are usually proprietary.
I once spent the better part of a week trying to figure out why my YouTube homepage had drifted into clips of people destroying expensive watches with hydraulic presses. I watched maybe two of them — one shows up, you're curious, you click. By day three the feed was eighty percent watch-pressing content. It was funny, and also a clean small-scale demo of how aggressively these systems lock onto a signal once they have one.
Most articles about "the algorithm" are written either by people who blame it for everything or by people who shrug and say it just shows you what you like. Both are too simple. As someone who has read the published papers behind these systems and built much smaller versions for previous work, I want to lay out what these systems actually do — and what they don't.
What a "feed algorithm" actually is
Strip away the metaphors and a modern feed is doing one specific job: given a candidate pool of millions of pieces of content, rank them in the order that maximizes some objective. That objective is almost always engagement — watch time, completed views, likes, shares, comments — because engagement is what platforms make money on.
To do that ranking, modern systems run roughly three stages:
1. Candidate generation. Of the millions of possible posts, pull a few thousand that you might plausibly care about. This stage uses fast, cheap signals: people you follow, people similar to you engaged with this recently, topics you've watched before, this post is going viral right now.
2. Scoring. Run a much more expensive machine learning model on each candidate to predict how likely you specifically are to do various engagement actions on it. Modern systems predict not just clicks but multiple actions (watch through, like, share, comment, follow) and combine them with weights into a single score.
3. Ranking and constraints. Order by score, then apply rules: don't show the same creator three times in a row, mix in fresh content, downrank flagged categories, respect ads inventory.
This is essentially what's described in the <a href="https://github.com/twitter/the-algorithm" rel="nofollow noreferrer" target="_blank">Twitter algorithm Twitter open-sourced in March 2023</a> — one of the few times a major recommendation system was made public — and the architecture is similar at TikTok, YouTube, Instagram, and elsewhere. The details differ, the shape doesn't.
Why it "knows" things you didn't tell it
This is the part that feels eerie and is actually mundane.
Modern recommendation models don't read minds. They learn embeddings — high-dimensional numerical fingerprints — for both users and content. Two users with similar engagement patterns end up close together in this space. Two pieces of content that engage similar audiences end up close together too. To predict what you'll like, the system finds content whose fingerprint is close to users whose fingerprints are close to yours.
You never tell the system "I'm interested in productivity hacks for software engineers." It just notices that people who watched the same three productivity videos you did also tended to watch a fourth one, and serves that.
The reason this can feel like the algorithm "knows" something private is that engagement is highly informative. Whether you finished a video, whether you scrolled past or hovered, when you tend to be active — these signals collectively carry more information than people realize, and modern models are extremely good at extracting it.
What recommendation systems actually do, and don't do
This is where I want to push back on a couple of confident claims you often see in articles about feeds.
They do strongly amplify whatever they pick up early. The watch-presses example. Once you give a feed a signal it can latch onto, it will aggressively exploit that signal. The same is true of mood-laden content: if you slow down on emotional posts, the feed will give you more of them.
They probably amplify outrage and emotionally extreme content. Not because anyone "wrote outrage into the algorithm," but because outrage drives the actions the system is optimizing for — comments, shares, time spent. Internal documents like the leaked <a href="https://www.wsj.com/articles/the-facebook-files-11631713039" rel="nofollow noreferrer" target="_blank">Facebook Files</a> showed Meta researchers themselves observing this pattern.
The "filter bubble" story is real but smaller than the framing suggests. A widely cited <a href="https://www.science.org/doi/10.1126/science.aaa1160" rel="nofollow noreferrer" target="_blank">2015 Science study by Bakshy and colleagues at Facebook</a> looked at how often users encountered politically cross-cutting content. The result was nuanced: algorithmic curation reduces exposure to opposing views some, but user choice (who you follow, what you click on) reduces it more. Subsequent work has roughly confirmed this. Algorithms don't lock you in a bubble against your will; they accelerate the bubble you've already started building by your own choices.
They don't have intent, but they do have alignment problems. "The algorithm is evil" is too strong. "The algorithm is optimizing engagement and engagement is not the same as your wellbeing" is closer to the honest version. A recommender doesn't want you to feel worse. It just doesn't know how you feel. The objective it can measure (clicks, time, shares) is a noisy proxy for the thing you actually care about (a worthwhile use of your evening).
What I actually do as a user
After spending time on the engineering side of this, here is how I treat my own feeds. I will not pretend any of this is heroic — it is just small habits, applied consistently, that have measurably helped.
Aggressively use the "not interested" and "show less of this" buttons. These are not vanity controls. They directly affect your embedding in the recommender. A handful of explicit signals is worth a lot more than passive scrolling.
Treat the home feed as junk food, the search box as the real meal. When I open YouTube intentionally — searching for a specific topic — the experience is qualitatively different from when I open it for background scrolling. The recommender is much more useful on a query than on an open question of "what should I watch."
Curate followed accounts twice a year. Most platforms now mix algorithmically suggested content into your follow feed. Pruning who you follow doesn't perfectly fix this, but it raises the floor of what the system is working with.
Cap the apps that compound the problem. Apps that combine an infinite feed, autoplay, and a high engagement objective (TikTok, Reels, Shorts) are designed for the longest possible session. Setting an OS-level daily limit isn't a moral choice; it's an engineering response to a system optimizing against your attention.
Reset when the feed has clearly gone off the rails. YouTube and TikTok both let you clear watch history. After my hydraulic-press week I cleared YouTube history once and the feed reset much faster than I expected — within a day or two it was back to a roughly reasonable mix.
What's coming
The shift to watch is the move from "predict what you'll engage with" to "let an LLM generate a custom rollup of content for you." TikTok and YouTube are already experimenting with AI-summary layers on top of their recommenders. This will probably make the systems more useful (LLM-generated short context can replace some of the doomscrolling) and more powerful (better at understanding what you actually want). It will also make the objectives even less visible to the user.
For most people the practical question stays the same. You can't engineer the platform's objective function. You can adjust how much of your attention you give to a system whose goals are partly aligned with yours and partly not.
The honest summary
Your feed is a machine-learning model running on the engagement signal you and millions of others provide. It is excellent at finding more of what you have shown it you'll engage with. It is not a neutral mirror; it is a directional amplifier. The most useful response is neither panic about "the algorithm" nor surrender to whatever it serves up without understanding how these systems actually work. It is to be a little more aware of what you are training it with, and to use the controls — explicit signals, history resets, time limits, intentional search — that the platforms have, sometimes grudgingly, provided.
If you have built or studied these systems and disagree with anything here, I want to know. Email <a href="mailto:hello@curiospark.org">hello@curiospark.org</a>.
Most Surprising Fact
When Twitter open-sourced its ranking algorithm in March 2023, much of it looked roughly like what academic papers had described for years — multi-stage candidate generation, embedding-based similarity, and a final scoring model. The interesting part was the specific weights and signals, not the architecture.
A recommender doesn't want you to feel worse. It just doesn't know how you feel. The objective it can measure isn't the same as the thing you actually care about.
— CurioSpark
💬 Perfect for sharing on social media
📚Sources & Further Reading
- •Twitter's the-algorithm source code— Twitter, Inc. (open-sourced March 2023)
- •
- •The Facebook Files: A Wall Street Journal investigation— The Wall Street Journal
- •AI Index Report 2024 — Chapter on Recommender Systems and Social Media— Stanford Institute for Human-Centered AI
All facts on CurioSpark are verified by our editorial team using peer-reviewed research and authoritative sources. Learn about our fact-checking process
🔗 Related Facts You'll Love
What 'AI in Daily Life' Actually Means — From a Developer Who Ships It
AI doesn't look like robots in 2026. It looks like autocorrect that's gotten weirdly good, search that quietly reads your context, and a thousand small product decisions you barely notice. Here's what's actually under the hood, from someone who's shipped these features.
AI Safety for Everyone Else: The Practical Version, Not the Existential One
Most articles about 'AI safety' are about long-term risks researchers think about at conferences. This one is about the small, near-term mistakes regular people are actually making — and how to avoid them.
Honey Never Spoils—Archaeologists Found 3,000-Year-Old Honey
Honey found sealed in 3,000-year-old Egyptian tombs has been chemically intact and theoretically edible. The reason has nothing to do with magic — it's three boring chemistry properties stacked on top of each other in a way no microbe can deal with.
Octopuses Have Three Hearts and Blue Blood
Octopuses have three hearts and blue, copper-based blood — and the main heart actually stops beating when they swim. Here is why their biology took such a different path from ours.
Did this blow your mind? Share it with someone who needs to know!
Software developer turned writer. Said covers technology, psychology, and human behavior — focusing on what the research actually shows rather than what headlines suggest. Every article is read line-by-line and fact-checked against primary sources before publishing.
Frequently Asked Questions
Closer to: the algorithm is designed to maximize engagement, and emotionally intense content (including outrage) drives more engagement actions than calm content. So the system amplifies outrage as a side effect of its objective, not because outrage was explicitly programmed in. Internal documents from Meta and others show their own researchers observing this pattern. The mechanism is real; the framing 'they want you angry' overstates intent.
Related Articles
What 'AI in Daily Life' Actually Means — From a Developer Who Ships It
AI doesn't look like robots in 2026. It looks like autocorrect that's gotten weirdly good, search that quietly reads your context, and a thousand small product decisions you barely notice. Here's what's actually under the hood, from someone who's shipped these features.
AI Safety for Everyone Else: The Practical Version, Not the Existential One
Most articles about 'AI safety' are about long-term risks researchers think about at conferences. This one is about the small, near-term mistakes regular people are actually making — and how to avoid them.
Honey Never Spoils—Archaeologists Found 3,000-Year-Old Honey
Honey found sealed in 3,000-year-old Egyptian tombs has been chemically intact and theoretically edible. The reason has nothing to do with magic — it's three boring chemistry properties stacked on top of each other in a way no microbe can deal with.
Read Next
The Placebo Effect Works Even When You Know It's a Placebo
It is one of the strangest findings in modern medicine: a sugar pill, openly labeled 'placebo,' still helps real patients. Here is what the studies actually show — and what they do not.
Your Brain Keeps Rewiring Itself — But Not in the Way Self-Help Books Suggest
Neuroplasticity is real, well-studied, and quietly misrepresented by half the books written about it. Here's what the actual research shows — and where the optimistic version oversells.
Honey Never Spoils—Archaeologists Found 3,000-Year-Old Honey
Honey found sealed in 3,000-year-old Egyptian tombs has been chemically intact and theoretically edible. The reason has nothing to do with magic — it's three boring chemistry properties stacked on top of each other in a way no microbe can deal with.