Recommendation Algorithms: How They Work and What They Want From You

A parent emailed me last year describing what she’d noticed: her 12-year-old had started watching one video about a specific topic — historical battles — and within two weeks had been served increasingly extreme content in the same genre, ending at graphic war footage the algorithm had surfaced entirely on its own. “I didn’t think it could escalate that fast,” she wrote. “There were no parental controls that caught it.”

The algorithm wasn’t malicious. It was doing exactly what it was designed to do: find content that kept her kid watching. The problem is that “keeps kids watching” and “is good for kids to watch” are not the same thing, and algorithms optimize for the first, not the second.

Key Takeaways

• Recommendation algorithms use collaborative filtering and content-based signals to predict what you’ll watch next
• What they optimize for — watch time, re-watch, shares — is not the same as what’s good for you
• The “rabbit hole” effect is real and documented: YouTube’s recommendation system has been shown to push content toward increasing extremity along engagement dimensions
• Understanding how the algorithm works gives kids a tool for using it deliberately rather than being used by it
• Parents can teach kids to override the algorithm — and doing so is a trainable skill, not just willpower

How Recommendation Algorithms Work

Most recommendation systems use some combination of two approaches: collaborative filtering and content-based filtering.

Collaborative Filtering

Collaborative filtering is the principle behind “people who liked X also liked Y.” The system looks at behavior patterns across millions of users. If user A and user B both watched videos C, D, and E — and then user A watched F — the system might recommend F to user B.

The key insight is that the algorithm doesn’t need to understand what the content is. It just needs to identify users with similar behavior patterns and extrapolate. Netflix’s early recommendation system, which won a $1 million prize competition in 2009, was built primarily on collaborative filtering with matrix factorization — finding patterns in a massive grid of user × content viewing histories (Bell & Koren, 2007).

Content-Based Filtering

Content-based filtering recommends items with similar characteristics to things you’ve engaged with before. For YouTube, this involves analyzing metadata (title, description, tags), audio features, visual features, and increasingly, the content itself via computer vision and transcription.

What Modern Systems Actually Use

Current systems at the scale of TikTok and YouTube use neither approach in isolation — they use hybrid deep learning models that process hundreds of signals simultaneously. A 2016 paper from Google’s YouTube team described their two-stage recommendation architecture: a candidate generation network (filtering billions of videos to hundreds of candidates) followed by a ranking network that scores each candidate using dozens of features including watch time history, search history, demographics, and content signals (Covington et al., 2016).

TikTok’s system is less publicly documented, but research suggests it weights your moment-to-moment behavior — how far you scroll, whether you replay a video, how quickly you scroll past something — extremely heavily. A 2021 investigation by the Wall Street Journal created test accounts and found that TikTok’s algorithm identified a user’s primary interest and began serving it almost exclusively within 40 minutes (Hagey & Horwitz, 2021).

What the Algorithm Is Actually Optimizing For

This is the part that’s hardest for kids to understand because it’s counterintuitive: platforms optimize for what they can measure, not what’s actually valuable.

Watch time is measurable. So is re-watch, share count, like rate, and comment volume. What’s not easily measurable: whether a video made someone feel better or worse, whether it was accurate, whether it taught something useful, whether it led the person toward a more diverse information diet.

Algorithms optimize for measurable signals. And the content that tends to generate the most watch time and engagement is content that provokes strong emotion — outrage, fear, curiosity, awe, humor. This is why recommendation feeds can drift toward emotional extremes even when the individual user didn’t choose that direction.

Research from MIT confirmed this mathematically: across six major platforms, content that generated moral outrage spread faster and received more engagement than neutral content, independent of its factual accuracy (Brady et al., 2020). The algorithm didn’t “want” to spread outrage — it was just very good at finding what spread, and outrage happened to spread best.

Platform	Primary Optimization Signal	What They Claim to Optimize	Known Rabbit Hole Patterns
YouTube	Watch time + click-through rate	”Viewer satisfaction”	Increasingly extreme content along engagement axes
TikTok	Video completion rate + replay	”Content you’ll enjoy”	High-velocity interest lock-in within ~40 minutes
Instagram Reels	Reach (shares/sends)	“Inspiring and relevant”	Body image content, engagement-bait chains
Netflix	Completion rate + re-engagement	”Content you’ll love”	Autoplay drives binge-watching
Spotify	Listen completion + playlist adds	”Music you enjoy”	Genre tunnel vision, reduced discovery over time

The Rabbit Hole: How It Actually Works

The “rabbit hole” isn’t an accident. It’s a structural feature of systems that optimize for engagement.

Here’s the mechanism. The algorithm has to pick the next video (or post, or song) from a vast space of candidates. The “safest” pick — the one least likely to cause the user to stop — tends to be content that’s similar to what they just engaged with, but more intense in the engagement dimension. Not quite the same (boring) but not too far afield (uncertain). Incrementally more extreme along the axis of whatever emotion the previous content triggered.

For kids watching gaming content, the drift might be toward increasingly high-emotion competitive moments. For kids watching fitness content, toward increasingly unrealistic body standards. For kids watching political content, toward increasingly partisan framing. The specific direction depends on what the algorithm found most engaging in their history — but the directional drift toward intensity is systematic.

A 2019 study by Ribeiro et al. tracked over 300,000 YouTube videos and found a consistent pathway from mainstream political content to more extreme alternatives through recommendation chains (Ribeiro et al., 2019). The debate about the magnitude of this effect continues, but the directional finding is robust.

What Kids Should Know: The Algorithm Is Trainable

Here’s the most useful thing to teach a kid: the algorithm is not fixed. It responds to your behavior. You can train it.

If you watch something to the end, you’re telling the algorithm: “more of this.” If you scroll past something quickly, you’re saying: “less of this.” If you deliberately seek out content outside your usual patterns, you’re telling the algorithm: “I have broader interests than you think.”

Kids who understand this can actively shape their recommendations rather than passively receive them. They can:

• Use YouTube’s “Not interested” and “Don’t recommend this channel” buttons deliberately
• Clear their watch history periodically to reset recommendations
• Use search proactively rather than relying on the feed
• Deliberately watch content on topics they want to explore, not just what they’re already interested in
• Recognize when they’ve been in a rabbit hole and consciously change direction

This is a learnable skill, not just willpower. And learning it requires understanding the underlying mechanism — which is why teaching kids how the algorithm works is not just tech literacy but practical self-determination.

For more on how AI systems influence what kids see and learn, see our broader guide on how AI works for kids and parents.

How to Teach Your Kid About Recommendation Algorithms

Ages 5–8: The “What Did the Computer Think You’d Like?” Game

After watching a few videos together on YouTube Kids or a similar platform, pause before the next one loads. Ask: “What do you think will come up next? Why?” Then check. Over several rounds, kids start to notice the pattern: it usually recommends something similar to what they just watched. You can introduce the word “algorithm” without math — just “a set of rules the computer uses to decide what to show next.”

Ages 9–12: Feed Archaeology

On a platform your child uses, scroll through their recommendation feed together. Ask: “Why do you think this is being recommended?” Trace it back — it was probably because of something they watched or searched recently. Then ask: “Is this what you actually want to spend time on, or is it just what got suggested?” The goal is to make the gap between “recommended” and “chosen” visible.

Ages 13+: Deliberately Reshape the Algorithm

Have your teen pick a topic they’ve never explored on their primary platform and deliberately watch 5–10 videos about it, skipping anything unrelated. Check the next day: has the feed shifted? Document the change. This makes the feedback loop concrete: the algorithm is responsive to your deliberate choices, not just your passive consumption. Discuss: if you can train it toward any topic — what topic do you actually want it to learn?

The question to ask: “If the algorithm shows you more of what you already like — how do you ever find something new that you’d like even more?”

What to Watch For Over the Next 3 Months

Month 1: Observe, don’t restrict. For two weeks, pay attention to what content your child’s feeds are surfacing. Notice patterns. Are the recommendations narrowing into a specific type of content? Getting more emotionally intense? This baseline matters.

Month 2: Introduce the vocabulary. “Rabbit hole,” “algorithm,” “recommendation feed.” Just naming the mechanism changes kids’ relationship with it. Research on media literacy education consistently shows that naming the persuasive technique reduces its effectiveness.

Month 3: Have a deliberate “algorithm reset” experiment. Clear watch history on one platform together. Observe how the feed changes. Use it as a launch pad for discussing what content your child actually wants to be recommended, versus what the algorithm had learned to show them.

Frequently Asked Questions

Why does TikTok feel so much more addictive than other platforms?

TikTok’s algorithm is specifically optimized for short-form video completion rate and processes your moment-to-moment behavior (replay, scroll speed, profile visits) faster and with more granularity than longer-form platforms. The short video length also reduces the commitment cost of each viewing decision, which increases the rate at which the algorithm can iterate on your preferences.

Can I turn off recommendation algorithms?

Partially. YouTube offers a “Don’t recommend channel” option and allows you to watch without recommendations by using direct search. TikTok allows you to clear your interest history in settings. Netflix allows you to remove items from your viewing history (which affects recommendations). None of these fully remove algorithmic influence, but they all reduce it.

Is the algorithm actually bad for kids?

It depends on what it’s recommending. The algorithm itself is neutral — it reflects your behavior back at you, amplified. If your kid’s feed is full of craft tutorials and science explainers, the rabbit hole leads to more of those. The concern is when the feedback loop amplifies emotionally intense or harmful content. Monitoring the feed is more useful than avoiding the platform entirely.

At what age can kids understand how recommendation algorithms work?

Research on media literacy suggests children as young as 8–9 can grasp the basic concept that “the app shows you things it thinks you’ll watch, based on what you watched before.” Full understanding of optimization and feedback loops develops in early adolescence. But naming the phenomenon — “that’s the algorithm” — is useful even for younger kids.

---

About the author Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Covington, P., Adams, J., & Sargin, E. (2016). “Deep Neural Networks for YouTube Recommendations.” ACM RecSys 2016. https://doi.org/10.1145/2959100.2959190
2. Brady, W. J., Wills, J. A., Jost, J. T., et al. (2020). “Emotion Shapes the Diffusion of Moralized Content in Social Networks.” PNAS, 114(28), pp. 7313–7318. https://doi.org/10.1073/pnas.1618923114
3. Ribeiro, M. H., Ottoni, R., West, R., et al. (2019). “Auditing Radicalization Pathways on YouTube.” arXiv. https://arxiv.org/abs/1908.08313
4. Bell, R. M., & Koren, Y. (2007). “Lessons from the Netflix Prize Challenge.” ACM SIGKDD Explorations, 9(2), pp. 75–79. https://doi.org/10.1145/1345448.1345465
5. Hagey, K., & Horwitz, J. (2021, July 21). “Facebook Tried to Make Its Platform a Healthier Place. It Got Angrier Instead.” Wall Street Journal. https://www.wsj.com/articles/facebook-algorithm-change-zuckerberg-11631654215
6. Pew Research Center. (2023). “Teens and Social Media.” https://www.pewresearch.org/internet/2023/12/11/teens-and-social-media/