A simple test
Open TikTok. Look at your "For You" feed. Now try to describe what those videos have in common. Not the surface topic — that is easy — but the emotional tone, the type of conflict that shows up, the level of irony, the pace of the editing, the kind of creator that hooks you.
You probably cannot fully articulate it. But TikTok can. It has 500 variables about your preferences, built from 10,000 micro-actions you took without thinking. And it is more accurate at describing your psychology than you are.
What data they actually collect
When we think about the "data" that social networks collect, we usually think of the obvious: what we post, who we follow, what we search for. That is 5% of the picture.
The other 95% is implicit behavioral data, which reveals far more about who you are than what you explicitly declare:
Micro-attention spans
TikTok measures exactly how many milliseconds you spent at each point in each video. Not just whether you watched it all: whether you paused at second 3.7, whether you rewound between seconds 12 and 18, whether your scroll pace slowed down at that moment.
These micro-timings reveal what captures you emotionally before you consciously process the decision. Your involuntary attention is more honest than your likes.
Session patterns
What time you open the app. How long it takes until you get hooked. When you start scrolling faster (a sign that the content is not resonating). When you slow down the most (the kind of content that anchors you). How each session ends and in what mood.
The algorithm knows whether you open TikTok when you are bored vs. when you are anxious vs. when you are lonely — because each emotional state produces a different usage pattern.
Passive biometrics
This is the most unsettling one. The phone's accelerometers register the muscle tension in how you hold the device. Blink rate (captured through the front camera on some devices) changes with attention level. Scroll pace correlates with the state of emotional arousal.
Meta holds a patent (2021) to infer emotional states from these sensors. It does not require explicit camera access: motion sensors are enough.
The psychological model: what they build with that data
With these signals, recommender systems build what engineers at these companies call "user embeddings": high-dimensional vector representations of your psychology.
What they can infer with documented accuracy (>80% in internal validation studies at these companies, leaked by researchers):
Big Five personality traits
The five major personality traits in psychology (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) are inferable from social media behavior with an accuracy that rivals 100-question tests administered directly.
Youyou et al. (2015) in PNAS showed that with 300 Facebook likes, an ML model predicts a user's personality better than their own spouse. With more data from TikTok, the accuracy is even higher.
Real-time emotional state
Algorithms detect changes in the user's emotional state during the session and adjust the content to keep it at the optimal arousal level for engagement. Not too calm (they leave). Not too disturbed (they also leave). The goal is the zone of maximum engagement: slightly emotionally activated, with sustained curiosity, slightly unsatisfied so they keep searching.
Vulnerabilities and risk states
In 2017, an internal Facebook document (leaked to The Australian) described how the algorithm could detect when a teenager felt "insecure," "worthless" or "in need of confidence" — and how that information could be used to serve advertising in those moments of vulnerability.
Facebook denied using this information for advertising. But building that model requires detecting those states. The model exists.
How the algorithm uses that model to maximize engagement
Once the psychological profile is built, the algorithm uses it to solve an optimization problem: maximize total session time and return frequency. The main levers:
Calibration of variable reinforcement
The most powerful addiction model known in psychology is variable reinforcement: unpredictable rewards on a variable ratio. Slot machines use it. So does scrolling.
The algorithm does not just show you relevant content. It calibrates the ratio of highly relevant vs. mediocre content to keep the "reward-seeking" feeling active. If everything were perfect, scrolling would lose its "maybe the next one will be better" component. The optimal ratio produces the longest session time.
Exploitation of each person's specific vulnerabilities
If the model detects that you are susceptible to social comparison (signal: lingering longer on content that shows aspirational lifestyles), it will serve you more content of that type. If you are susceptible to moral outrage, your feed will have more conflict and outrage.
It is not that the algorithm is malevolent. It is optimizing for a metric (engagement), and the exploitation of psychological vulnerabilities is the most efficient solution it has found.
The intensification loop
The algorithm learns with every session. If a type of content produces more engagement, it serves more of the same, with slight escalation. Users who started with moderate political content report feeds dominated by extremes after months of use, without having actively sought that content.
This phenomenon — algorithmic radicalization — is documented in independent research on YouTube (Ribeiro et al., 2020) and in internal Facebook communications leaked in the Facebook Papers (2021).
What you cannot consciously control
The fundamental problem is one of information asymmetry. You do not know what model the algorithm has of you. It knows exactly what model it has of you.
When you think you are consciously choosing what to watch, you are actually executing conditioned responses that the algorithm has calibrated over months. The choice is real — no one forces you to stay — but the menu of options and the order of presentation are designed to make one option substantially more likely than the others.
It is the difference between freely choosing in a supermarket designed by consumer behavior psychologists. Technically you choose. But choice patterns are predictable with high precision.
What can you do about this?
This is not about paranoia or deleting every app. It is about understanding the real nature of the interaction so you can make more informed decisions.
1. Actively curate your feed
The algorithm learns from your behavior. If you actively stop interacting with certain types of content (no likes, no comments, even reporting "not interested"), the model adjusts. The feed can be retrained, but it takes sustained intentional effort over weeks.
2. Use "Non-personalized" mode
TikTok, YouTube and other platforms have options to view non-personalized content or in guest mode. You lose relevance, but also the exploitation of specific vulnerabilities. For informational use, it may be preferable.
3. Awareness of usage patterns
Keeping a simple log of when and why you open apps creates a mirror between your actual behavior and your self-image. Most users significantly underestimate their usage time before measuring it objectively.
4. Technical restriction tools
Screen time limits, grayscale modes, and app restrictions are deliberately hackable (platforms push you to disable them). But adding extra technical friction for access does significantly reduce impulsive use.
The underlying lesson: know yourself before the algorithm does
The deepest irony of algorithmic modeling is this: these platforms know what emotionally activates you, what makes you feel inferior, what kind of conflict hooks you, when you are most vulnerable. And many users do not know these things about themselves with that precision.
The antidote is not just to reduce use. It is to invest that time in the kind of self-knowledge that makes the algorithmic profile less exploitable: knowing when you are vulnerable, what emotions move you, what makes you make decisions you regret. Not for the algorithm. For you.