Why TikTok Is So Addictive

TikTok ranks videos by predicting which one you will engage with next, reading signals you produce by watching rather than choosing: how long you stayed, whether you finished, whether you skipped. TikTok publishes the signal list and some qualitative comparisons between signals, but no numeric weights and no stable ranking.

You opened the app to check one thing and an hour went missing. There is a mechanical explanation for that, and it does not require you to be diagnosed with anything.

Many popular explanations go beyond what TikTok and the retrieved research establish. They state weights and formulas that TikTok has not published, and they borrow an explanation from gambling research that has not been tested on a short-form video feed in the retrieved literature. What follows separates three things: what TikTok documents, what the interface does, and what the research does and does not support.

TikTok publishes a list of signals, and no numbers

TikTok’s Transparency Center hosts a page called Introduction to the TikTok recommendation system, and it is the closest thing to a primary source that exists (TikTok Transparency Center).

It names three groups of inputs. The first is your predicted interactions: whether you like, share or comment on a video, whether you mark it “Not Interested”, whether you follow the account, whether you “finish” or “skip” it, whether you favourite it, your “time spent viewing”, and whether you tap the soundtrack. The second is video signals: post time, region, language, soundtrack, length and hashtags. The third is contextual signals: your region, your device language and your operating system.

Two of those deserve attention because they are the ones you generate without deciding to. Finishing a video is an input. Skipping a video is an input. You are not choosing to send a signal when you swipe; you are choosing whether to keep watching, and the system reads that choice.

The page carries no exposed last-updated date. The most authoritative public description of the system does not tell you when it was last true.

TikTok says the signals are weighted, and that the weights change

The second primary source is a TikTok Newsroom post, How TikTok recommends videos #ForYou, dated 18 June 2020 (TikTok Newsroom 2020). It is six years old and it is still the document most articles quote.

It says signals are “weighted based on their value to a user”, and it gives qualitative examples of what that means. Finishing a longer video counts as a stronger indicator than matching a video to a viewer in the same country, which the post describes as a weak one. Device and account settings receive lower weight. And it describes the system as a “continuous process” in which TikTok will “reassess the factors and weights”.

So the boundary is this. TikTok has published qualitative comparisons between some signals, and those comparisons are genuinely informative: a finished long video means more than a shared country, and your device settings matter less than what you watch. It has not published numeric weights, and it has not published a complete stable ranking of signals. Any article handing you a percentage, or a ranked list of every signal in order of importance, is going past what the platform has disclosed. Third parties do publish research and estimates about how the system behaves in practice, and some of that work is worth reading. What none of it can be is an official weight, because no official weight has been published.

The replay question, precisely

The 18 June 2020 post mentions replays once, in its new-user bootstrap passage, where it says: “Your first set of likes, comments, and replays will initiate an early round of recommendations as the system begins to learn more about your content tastes.”

Note where that sits. It is the passage about getting a first read on someone with no history. Replays do not appear in that post’s own list of ranking factors, and they do not appear in the current Transparency Center description of the recommendation system, which omits them.

So the accurate statement is narrow. A six-year-old post named replays among the interactions that start a new user’s recommendations. No weight for replays has been published, and neither has a ranking order between replays and finishing a video. Nor does TikTok rank on completion and replay as a fixed formula: its own text rules that out when it says it will reassess the factors and weights.

The slot machine comparison is an analogy, and it has not been tested here

The standard explanation for the missing hour is variable-ratio reinforcement: the schedule, familiar from gambling research, in which a reward arrives after an unpredictable number of attempts and the unpredictability itself sustains the behaviour. It is a genuine phenomenon with a real literature behind it. The problem is the transfer.

Clark and Zack’s 2023 paper in Addictive Behaviors, “Engineered highs: Reward variability and frequency as potential prerequisites of behavioural addiction”, is a theoretical paper proposing a framework (Clark and Zack 2023). It proposes; it does not test. The documented retrieval for this dossier, dated 4 September 2026, did not locate an empirical test of a variable-ratio schedule against a short-form video feed.

The two descriptions are also not obviously the same machine. A variable-ratio schedule delivers reward after an unpredictable number of responses and is indifferent to who is responding. TikTok’s documented mechanism is a ranking system that predicts which video you are most likely to engage with and serves that. Whether a personalised ranking system can nonetheless produce, in practice, the pattern of unpredictable reward the schedule describes is an open question, and nothing retrieved for this dossier settles it either way.

So hold variable reward as what it is here: an explanatory analogy imported from gambling research, not an empirically established account of short-form feeds in the retrieved literature.

What the interface removes

This section is design analysis. It describes what the product does, not a measured effect on behaviour. A traditional media object has an edge. An article ends. An episode runs out. A page has a bottom. Each of those is a stopping cue, a moment where continuing requires a decision and the decision is visible to you.

TikTok’s feed has no edge. Videos play automatically, one swipe replaces the current unit with the next, and the feed does not end. Because the units are short, the decision point that would normally arrive at the end of a piece of content arrives constantly and passes in under a second, which is not the same thing as arriving in a form you can act on.

Whether removing stopping cues measurably extends viewing sessions is a separate empirical question, and it is not answered in the retrieved literature.

The research most articles cite does not say what they say it says

Two sources dominate the citation trail on this query, and both are weaker than they appear.

Montag, Yang and Elhai (2021), in Frontiers in Public Health, is cited constantly as evidence about TikTok users. It is a narrative review (Montag et al. 2021). It collected no original data and has no sample size, despite the phrase “empirical findings” in its title. It does not analyse variable-ratio reinforcement or operant conditioning, which does not stop it being cited in support of both.

The most recent attempt at a synthesis is Arouch et al. (2025), a systematic review of short-form video use and cognitive and mental health outcomes (Arouch et al. 2025). It screened 519,101 records and found only 17 studies that met inclusion criteria, which is itself the finding: the literature is thin. The evidence it collects is predominantly cross-sectional self-report, longitudinal work is minimal, and the causal direction is unresolved. Its authors state plainly that the evidence cannot distinguish whether short-form video use impairs cognition or whether people already prone to attentional difficulty gravitate toward it. That review is a preprint on medRxiv and has not been peer reviewed.

So: short-form video use is associated with attentional disruption, on a small body of mostly cross-sectional self-report evidence, synthesised in a preprint whose authors decline to claim a direction. That sentence is as far as the retrieved evidence goes.

There is no diagnosis at the end of this

TikTok addiction is not a diagnosis. Neither is social media addiction. Neither is recognised in DSM-5-TR or ICD-11.

It is worth being precise about what those two classifications do contain, because the gaming case is regularly borrowed to argue about feeds. DSM-5-TR includes Internet Gaming Disorder in its section of conditions recommended for further study. That is a research proposal, not a formal diagnosis, and the American Psychiatric Association draws the boundary itself: “This proposed condition is limited to gaming and does not include problems with general use of the internet, online gambling, or use of social media or smartphones” (American Psychiatric Association).

ICD-11 goes further for gaming and no further for anything else. Gaming disorder, code 6C51, is a formal diagnosis, alongside 6C50 gambling disorder, and the World Health Organization’s terminology page lists those two and no others under disorders due to addictive behaviours (World Health Organization).

What exists for short-form video instead is scale-development and cross-sectional survey work measuring self-reported problematic use. Galanis et al. (2024) is representative: a cross-sectional validation of a TikTok addiction scale with 429 participants in Greece (Galanis et al. 2024). Studies like it can tell you how people answer questions about their own use. A scale named for a condition that no classification recognises cannot establish that the condition exists.

That matters for how you read your own lost hour. The hour is real. It does not come with a condition attached, and nothing in this article should be read as an assessment of anyone.

What is actually going on

Strip out what cannot be supported and a specific picture remains, most of it from TikTok’s own documentation.

A ranking system predicts which video you will engage with next, reading inputs you produce by watching rather than deciding: how long you stayed, whether you finished, whether you skipped, whether you tapped the sound. The signals are weighted, TikTok says the factors and weights are reassessed over time, and the numbers have never been published. The interface those predictions arrive through has no endpoint.

That is the mechanism as far as the documentation carries it. It is not a diagnosis, and it has not been shown in the retrieved literature to work like a slot machine. A continuous personalised feed with no endpoint is a plausible design contribution to an hour that goes missing. Its independent causal effect on how long anyone watches is not established here, and this article does not claim it.

If you are reading this because of someone else’s hour rather than your own, the version written for that conversation is why your teenager can’t put TikTok down. What the ranking system infers about you, rather than what it does to you, is the subject of what your feed knows about you.

How these design patterns work across platforms rather than inside one app is set out in the persuasive design guide.

We read the platform documentation and the papers, and publish where the claims stop. The Weekly System has not launched yet. You can join the launch list.