The Invisible Machine: How Short-Form Video Platforms Are Engineered to Keep You Scrolling
Every day, people around the world spend a collective 3.3 billion hours watching short-form videos on TikTok. And no doubt many of you reading this will have contributed to that staggering statistic… I know I certainly have.
This figure isn’t a happy accident or viral luck, but the product of some of the most sophisticated, deliberately engineered user experience design in the history of technology.
For a long time, I’ve been fascinated with how social media content – from static images to animations to videos more broadly – use psychology-based techniques to influence audience behaviour. Somewhat naively, I never really stopped to consider how the platforms themselves are also employing a broad range of tactics to keep people hooked, scrolling, and watching for hours on end.
And after doing some research, I discovered just how deep this goes. From the exact physics of how a feed scrolls under your thumb as you swipe, to an AI-based system that can update its understanding of you faster than you can finish watching a single video, short-form platforms like TikTok, Instagram Reels, and YouTube Shorts have constructed a three-layer machine designed to capture and hold human attention at industrial scale.
Much of this information comes from Enrico Tartarotti's incredible deep dive on the subject, entitled The Ridiculous Engineering of Short Form Content. I’ve also supplemented some of this information with my own research into the tools, techniques, and tricks the platforms use to keep people hooked.
In this article, we’re going to explore each level of the three layers – the interface, the psychology, and the algorithm – to see exactly how they do what they do.
Layer 1: The Interface – A Physics Engine Built for Your Thumb
The most immediately visible part of the attention-grabbing machine is the feed itself. Most interfaces of short-form video platforms follow the same layout: a simple (on the surface) vertical stream of full-screen videos with minimal buttons on the right, top, and bottom of the screen.
Pretty simple, right? Yes, on the surface it’s simple, but that simplicity is the result of enormous engineering effort.
Let’s look at how the interface of short-form video platforms like TikTok, YouTube Shorts, and Instagram Reels are engineered to keep you hooked.
The Scroll That Feels Like Nothing Else
When you develop an app for phone platforms like iOS and Android, there are various tools you can use to make the process easier. One of the most commonly-known tools is Google’s Material Design, which allows developers to immediately get up and running with a design style that aligns with Google’s (and Android’s) best practices – things like colours, shapes, and animations.
Each phone will come stocked with various standard operating-system behaviours for things like transitions and scrolls, and most apps will make use of these during development because it speeds up the process. However, short-video platforms do not.
The likes of TikTok and its competitors have built custom physics engines that govern exactly how the feed responds to a swipe. When you flick up, the feed doesn't merely move; it snaps, decelerates, and locks to the next video with a tactile precision that feels almost magnetic. This is the whole point to building custom systems.
To see this in action, open your phone and swipe left and right on the home screen. Do you see how that feels smooth and totally in your control? Now open TikTok and do the same thing; notice how the movement is much faster and almost instant?
This is intentional. Standard inertial scrolling (the kind used on most webpages or email tools) is designed for browsing. In this sense, browsing means looking at the content on your screen at a pace that is comfortable for you; this form of scroll lends itself to slow digestion, giving you the time to consider what you’re reading and/or watching.
Short form scroll physics, on the other hand, are designed for committing. Each gesture resolves to one complete video, removing the ambiguity of "where am I in the feed?" and replacing it with a clean, decisive transition. In other words, it instantly takes you to the next video, removing the potential for you to make the decision to leave the app, instead keeping you in the dopamine loop.
This snap-to-video mechanic is reinforced by the full-screen, vertical format itself. By occupying the entire viewport (a fancy word for screen), each video eliminates peripheral distractions and creates what researchers describe as an immersive viewing environment optimised for the mobile form factor. As one technical analysis of TikTok-style architecture notes, vertical full-screen design is foundational to the short-form experience, it makes the app "feel like users are diving into a stream of videos."
Engagement Buttons Easy Access
It’s no coincidence that the engagement buttons in these short-form video apps are situated right where they are – on the right-hand side of the screen.
People who are right-handed make up the vast majority of the world’s population, which means when most of the world are holding a phone, they’re holding it in their right hand. This means a person’s thumb is typically the dominant digit with which they interact with the screen. See where I’m going with this?
The engagement buttons – the like, comment, and share buttons – are situated on the right of the screen, sitting just below where a thumb will typically be positioned when at rest. This means that “liking” a video you’ve just seen is an easy action, because your thumb is already where it should be.
And with every new interaction, every like and every share, the algorithm builds up a clearer picture of what you like and don’t like. Once it understands you, it can be sure to serve new video content that keeps you hooked. We’ll go into this in more detail later in the essay.
Because those are the main interactions the platforms want you to have with the content, they put those buttons right where you can reach them. Now, contrast this with the buttons that let you tailor the content you’re being shown. At the top of the TikTok interface there are a handful of buttons, but by far the largest one is the search option. Next to that, you’ll see tools to let you fine-tine the content the algorithm is showing you.
This is somewhat hidden (not hidden in that its buried beneath layers of menus, but rather hidden because most people just don’t look up there) because platforms don’t want you to alter what you’re being shown. Most of us have good intentions about the media we want to consume; how many times have you told yourself that you’re only going to look at educational content? I’d guess a few times, because I’ve done the same thing. Yet it doesn’t take long before you’re back to watching nonsensical, mostly AI-generated content (over 80% of AI content comes from Agentic AI Accounts on TikTok, 15% on Instagram).
Short-form video platforms keep this hidden because they don’t want you to interfere with the algorithm. Instead, they want you to watch what you’re being served, because over time, the algorithm will build up a very detailed profile of what keeps you hooked. And the deeper you’re hooked, the more videos you watch, which in turn means the more ads you see, and thus more ad revenue generated for the platforms.
Micro-Interactions and the Illusion of Control
And speaking of the illusion of control on short-form video platforms, let’s look at micro-interactions.
You know when you hit the like button, there’s an animation? On Instagram, for example, if you tap the icon a little animation will show the outlined heart briefly getting bigger, and then shrinking back down as it fills to pink. Or, if you double-tap an image, a heart icon will bounce on the screen with a cool little animation and gradient.
None of this is decoration just to make the app look cool (looking cool is a side effect, but not the reason why). Instead, these micro-interactions serve as both feedback and reinforcement.
Each tiny interaction (like liking a video) is a data point collected by the platform; whilst each animation is a small reward for taking the action, which comes in the form of dopamine hit reinforcing the behaviour. You subconsciously think, “wow, that was a cool animation,” which then makes you more likely to tap the like button on another post, and then another, ad infinitum.
And as we know, social media platforms thrive on engagement, so giving you a dopamine-based reason to engage with something is a win-win for all involved. Well, not a win-win for those of us who find ourselves doomscrolling at 3am, so it’s maybe just a win for the platforms.
Layer 2: The Psychology – Unit Bias, Chronoception, and the Completion Effect
We’ve discussed the things we can see on the surface of short-form video platforms, but now it’s time to explore the things we cannot.
Below the polished interface of your favourite video app lies a layer that most of us never consciously perceive: a set of psychological principles that’ve been deliberately embedded into the design architecture, influencing how much time you spend watching and engaging with videos, as well as how that time feels.
Unit Bias: The Tyranny of Unfinished Things
Unit bias is a well-documented cognitive phenomenon that we’ve all fallen foul to at some point or another. It plays on our strong, often irrational preference for completing discrete units of things.
As an example: if you’re given a bowl of soup, there’s a high likelihood that you’ll finish it (regardless of whether you’re still hungry) because leaving it unfinished creates a mild but persistent cognitive discomfort. It’s the same reason why many readers will say they feel uncomfortable if they leave a chapter half-finished, or why you may find yourself spending the whole amount on a gift card on something you don’t want, because leaving £2.38 unspent feels like a waste.
Short-form video platforms exploit this psychological phenomenon ruthlessly.
Instead of thinking about a video as a thing with moving colours and a story, instead think about it as a single unit. Because we struggle to leave one unit half finished, once you’re 70% of the way through a clip, the psychological cost of abandoning it exceeds the cost of watching the final few seconds. This is also triggered by starting a video and only being a few seconds into it. Your brain subconsciously thinks “this video is only going to be, at most, a minute long so there’s no harm in finishing it.”
Doing that for a single video is fine, but it becomes a serious problem when you make that same decision over and over, video after video.
And when you’re stuck in a loop of feeling compelled to finish a video you’ve started (often you haven’t consciously decided to start another video, but rather your thumb has swiped before you’ve even thought about doing so) you enter a doomscrolling cycle. If you’re anything like me, you know all too well how hard it can be to get out of this loop once you’ve found your way in.
All of this is by design. The platform structures its content as an endless series of almost-finished things, each one easily completable, each completion immediately followed by the beginning of another incomplete unit. You are never done; you are always just about to be done, and as such, struggle to stop yourself from watching just one more video.
Follow the Casinos: Chronoception and Slot Machines
Have you ever been to a casino? Those big buildings full of people losing their hard-earned money are no strangers to the power of psychological manipulation. And as dystopian as it sounds, short-form video platforms have borrowed from the casino toolbox to make their apps more addictive.
Go to any casino and, when you’re aware of it, you’ll notice something very specific: there are no windows and no clocks. This is for the specific reason of messing with our Chronoception, or the subjective perception of time and time passing. In a casino, when you don’t know what the time is, you have no visual cues as to how much time you’ve already spent gambling your money away. The same applies to video platforms.
In a state of high engagement and flow, our internal clock is not a fixed, reliable stopwatch but a fluid construction of the brain – meaning that it’s heavily influenced by mental state and the volume of information being processed. As you’ve probably guessed, short-form video feeds are optimised to first induce, and then maintain, this state of mind.
The rapid succession of novel content (new faces, new sounds, new visual styles, all arriving every 15–60 seconds) keeps the brain in a heightened state of anticipatory arousal. The brain's time-tracking mechanisms are cognitively expensive; when other processing demands are high (as they are during an absorbing stream of video), the brain deprioritises accurate timekeeping and focuses attention on the stimulus. The result is that you look up from what feels like thirty seconds of watching a video only to realise that a few hours have passed and it’s dark outside.
The second technique these platforms have taken from casinos is that of the humble slot machine. In Addiction by Design, author Natasha Schüll explores how social media platforms employ the tactics and strategies pioneered by the gambling industry to keep users hooked. She writes:
Facebook, Twitter and other companies use methods similar to the gambling industry to keep users on their sites. In the online economy, revenue is a function of continuous consumer attention – which is measured in clicks and time spent.
A very on the nose example of this is the pull-to-refresh feature now implemented in every major social media app. This feature is where you drag your thumb down on the screen, which then loads more posts for you to see. This is very similar to the slot machines found in casinos, where you pull the big lever in the hopes of lining up three cherries and winning the jackpot. On social media, however, you’re not winning money, you’re pulling down the screen in hopes of seeing a new and exciting post that’ll keep you entertained.
This same mechanism is used on short form video platforms. However, instead of pulling to refresh the content on your screen, you’re swiping up to see the next video the algorithm thinks you’ll like and hoping to be rewarded with something you find funny, entertaining, interesting, etc. When you do get shown that, your brain associates that entertainment (and the dopamine hit that comes along with it) with swiping, adding to the litany of other dopamine-inducing techniques we’ve already discussed.
Layer 3: The Algorithm – Monolith and the Real-Time Mind-Reader
The third and most technically extraordinary layer is the recommendation engine: the system that decides which video plays next.
Everyone knows that algorithms are at play in the background of almost all websites now, especially those that encourage the creation and consumption of content. This is no secret, but what is a secret is how those algorithms work and what signals they’re looking for. You’ll see countless people trying to explain how they work, but none of them will have concrete information precisely because the architecture that makes algorithms tick isn’t publicly available.
Well, most of them aren’t publicly available. Interestingly, we know exactly how the TikTok algorithm works.
In fact, ByteDance (the parent company of video giant TikTok) published a paper in 2022 that explains exactly how Monolith – the algorithm at work on TikTok – functions.
Whilst we don’t know exactly how the algorithms of Instagram Reels and YouTube Shorts work, we can assume they function something like Monolith. So, let’s dive into Monolith a little deeper.
For full disclosure, a lot of the below is tech-speak that I’m not too proud to admit took me a long damn time to understand. I’ve tried as good as a layman can to explain it in simple terms, removing as much jargon as possible.
What is Monolith?
In 2022, Monolith was presented at the ACM Conference on Recommender Systems, which describes itself as “the premier international forum for the presentation of new research results, systems and techniques in the broad field of recommender systems.”
Recommender systems are algorithms, literally systems recommending content to users.
The paper, entitled Monolith: Real Time Recommendation System With Collisionless Embedding Table, is a ten-page document with diagrams and charts to explain exactly how the TikTok algorithm works to understand its users and serve content that will keep them watching.
Traditional recommendation systems (like the kind used by Netflix or early YouTube) are batch systems, meaning they process large datasets periodically, update a model overnight, and serve recommendations based on a model that is, at minimum, hours old. The problem with batch systems for short-form video is that user preferences shift in real time. What you want to watch at 11pm on a Friday is not what you want at 7am on a Monday.
With such a demand for new and exciting content, a model trained on yesterday's data simply cannot keep up with what users want today.
Monolith solves this with online training: the system learns from your interactions as they happen, updating model parameters in near-real-time. User clicks, likes, replays, and skips are logged, joined with content features, and immediately fed back into the model. Some preference updates propagate within ninety seconds of an interaction; in cases of sharp engagement shifts, adaptation happens even faster.
The Collisionless Embedding Table
The technical innovation at the centre of Monolith is its handling of sparse, high-dimensional data. Recommendation systems need to represent billions of unique entities (users, videos, audio tracks, creators, hashtags, and more) as numerical vectors (which are referred to as embeddings) that the algorithm can then process. Conventional systems use a fixed-size table, which inevitably causes collisions, where two different entities get mapped to the same point, degrading the overall quality and performance of the model.
Monolith, on the other hand, uses something called a Cuckoo Hashing-based embedding table that’s effectively collision-free. (I’m sorry, I know I promised jargon-free.) With this system, each unique entity in the table gets a unique representation which, combined with intelligent optimisations like the automatic pruning of stale data and frequency filtering (rare entities and outliers are handled efficiently), allows the system to operate at an astronomical scale without the quality degradation that plagues simpler systems.
Concept Drift: Why the Algorithm Never Sleeps
If all of that sounds annoyingly complex, it’s because it is. The average person would never see behind the scenes of the TikTok algorithm unless they read that long report I mentioned earlier. And even if they did read it, language is so chock-full of jargon and complexities that it’s still just as confusing.
What is all comes down to, however, is Monolith addressing critical challenge that engineers call concept drift. Concept drift is the name given to the tendency for statistical relationships between inputs and desired outputs to change over time.
Trends are born, live a whole life cycle, and then die on TikTok in just a matter of days. A meme format that can drive enormous engagement today may be totally exhausted by this time next week, potentially even this time tomorrow. A batch system (those used on early YouTube and platforms like Netflix) trained on last week's data will keep recommending exhausted trends and severely impacting content engagement.
However, because Monolith trains continuously on real-time streaming data, it adapts to concept drift almost instantly. The model's understanding of what content a user will find engaging is perpetually updated, making the For You Page feel like it mirrors your present, not your past. This is also why a trend will last way longer on platforms like YouTube than it will on TikTok, because the data is being processed much quicker and much more efficiently.
The very-real upshot of Monolith is that it aligns with your content interests in real time, making the content it serves you as accurate to your current interests as it can possibly be. All of this serves to keep you on the platform for longer, because you’re being shown a constant stream of videos that keep you entertained and engaging.
This doesn’t come from a place of benevolence, by the way. As we already know, a better algorithm means TikTok can understand its users better, showing them more accurate videos to keep them watching. And like we’ve already said, more views mean more ad revenue.
This hyper-advanced algorithm is all in aid of hooking attention and generating money.
The Three Layers Working Together
What makes short-form video platforms so extraordinarily effective (and, for many users, so intensely addictive and difficult to disengage from) is that the three layers we’ve discussed operate in unison, each one reinforcing the others.
The interface makes scrolling effortless and satisfying, lowering the activation energy for each additional video.
The psychological design makes each video easy to finish and makes time feel compressed, removing natural stopping points.
The algorithm ensures that the video waiting after the current one is, with uncanny reliability, exactly the kind of thing you want to see next.
The result of this combination is what researchers at Coventry University's Centre of Intelligent Healthcare describe as short-form video platforms with "rapid, algorithm-driven, and emotionally charged design" that raises distinctive concerns about its effects on cognitive function, particularly attention, executive function, and emotional regulation.
The conclusion to the Coventry University study linked above was that the “reviewed studies show that SFV (short-form video) use disrupts executive functioning across neural, behavioural, and subjective domains… alongside reduced engagement of top-down executive systems in response to reward-driven cues. Complementing these results, self-report studies indicate that users experience diminished self-control, difficulty shifting attention, and heightened emotional impulsivity in everyday life, particularly when SFV is used to cope with boredom or distress. Together, these findings point to a reinforcing feedback loop: weakened executive control contributes to compulsive SFV use, which in turn further undermines cognitive regulation.”
When you can see the very real effects of what overconsumption of short-form video, the huge investments made in engineering platforms to be as addictive as possible looks a whole lot more insidious, doesn’t it?
Conclusion
It may seem shocking to read about all the little tricks and techniques at work on short-form video platforms, but the reality is that none of this engineering is secret. Nor are there currently any major regulations being applied to stop them from intentionally engineering addictive platforms.
As we’ve seen, ByteDance published the Monolith paper itself, detailing exactly how TikTok’s algorithm keeps people hooked and coming back for more. The design choices made by TikTok (and by extension other short-form video platforms) are fully visible to anyone who wants to look.
But awareness and resistance are different totally different things. The reason so many of us find ourselves stuck in doomscrolling loops and find it so hard to break the habit of watching video after video is because the system is not designed to be resisted; it is designed to make resistance feel unnecessary.
Understanding the three layers (the design of the UI, the psychology of completion and time, and the real-time intelligence of the recommendation engine) is the first step toward using these platforms intentionally, rather than them using you.
The machine at work behind the platforms we all use every day is truly extraordinary, and knowing how they work is the beginning of being able to break the doomscrolling cycle when it tries to trap you in.