Persuasive Design: How Interfaces Are Built to Hold Your Attention

Persuasive design is the practice of shaping an interface so that it influences what you do next: what is switched on by default, what plays automatically, what has no natural stopping point, what gets ranked in front of you. It is a named field with its own literature, and the people who work in it publish about it. That makes this the one pillar where the primary sources are unusually good.

It is also the pillar where it is easiest to overclaim, because four very different kinds of statement get told as one story. This guide keeps them apart throughout: what a company documents about its own product; what a theory proposes about why a pattern should work; what an observational study finds travelling together; and what an experiment actually demonstrates. Public discussion often presents the first two as though they were the fourth.

It carries no new research. Every claim summarises something already sourced in one of the five articles it links to.

What the platforms document about their own ranking

Recommendation systems mostly do not run on what you say you want. They run on implicit feedback, meaning what you watched, how long, and what you scrolled past. Hu, Koren and Volinsky's 2008 paper set out the standard method for building recommenders from exactly that kind of signal, and was careful to describe implicit feedback as noisy rather than as a clean record of preference. TikTok's own public description of the For You system says the same thing in plainer language: watch signals carry weight, and finishing a longer video counts for more than a weak signal. That is documentation, not evidence of effect.

The economics behind the idea is older than the technology, and narrower than it is usually made to sound. Samuelson's revealed-preference framework (1938) is a formal method for inferring preference from observed choice under stated assumptions about consistency and budget constraints. It is not a claim that what you clicked is what you really wanted, and it says nothing about wellbeing. Milkman, Rogers and Bazerman's 2008 review of want/should conflicts is a conceptual synthesis of other people's studies rather than an experiment of its own, and it describes the gap that opens between what people want in the moment and what they think they should do. That gap is where a behaviour-trained feed operates. Both of those are theory.

The experiments arrived in 2023, and they are not one package. Guess and colleagues, in Science, ran the randomised chronological-feed experiment: over roughly three months it changed what people saw and cut time on the platform, without measurably moving polarisation or attitudes. Nyhan and colleagues, in Nature, ran a different randomised manipulation, reducing exposure to like-minded sources, and also found no attitudinal effect. Gonzalez-Bailon and colleagues, in Science, is observational rather than experimental: a descriptive analysis of exposure across roughly 208 million users, which cannot separate algorithmic curation from users' own choices.

What your feed knows about you works through the mechanism and those three studies in detail.

The loop, and what holds it together

Doomscrolling has a working definition (Sharma, Lee and Johnson, 2022) and a measurement scale (Satici and colleagues, 2022). What it does not have is causal evidence: both are cross-sectional surveys, so they cannot tell you whether the scrolling produces the distress or the distress produces the scrolling.

What is better supported is why negative material holds attention at all. Soroka, Fournier and Nir's 2019 study in the Proceedings of the National Academy of Sciences measured skin conductance and heart-rate variability in more than 1,000 participants across 17 countries while they watched real BBC World News segments, and found a cross-national negativity bias on average. The BBC clips were the stimulus material; the BBC did not commission or conduct the research. The same paper is equally clear that individual responses varied considerably and that country explained very little of that variation, so this is an average tendency, not a rule about any particular reader. It should not be confused with Trussler and Soroka's separate 2014 study, which measured which news stories people chose to read rather than how their bodies responded.

Grupe and Nitschke's 2013 review of uncertainty and anticipation in anxiety is often used to explain why an unresolved threat is hard to put down. It is a neurobiological review of clinical anxiety and contains no data on feeds, media or notifications, so treat that application as an inference rather than as evidence about scrolling.

Montag and colleagues (2019) tie the design elements together: infinite scroll, unpredictable rewards, notification-driven re-entry. That paper maps design patterns onto existing psychological and economic theory. It is a design analysis rather than a controlled experiment, and it is worth being clear which of the two you are reading.

The doomscrolling loop covers this pillar's central pattern. Why your teenager can't put TikTok down shows the same three elements, personalised ranking, no stopping cue and an unpredictable payoff, in the place they work hardest.

Night, when the design has no opposition

Bedtime is one context in which continuous phone use can displace sleep, but the evidence separates several possible pathways. Kroese and colleagues (2014, 2016) named bedtime procrastination, going to bed later than intended without an external reason, in two cross-sectional surveys, and Exelmans and Van den Bulck (2016) linked phone use in bed to poorer sleep in a survey of Belgian adults. Both are correlational.

Chang and colleagues (2015) documented the light and melatonin pathway in a controlled inpatient trial, and the detail matters: 12 participants, reading a light-emitting device for four hours each evening in otherwise dim light, which delayed sleep onset by about ten minutes and shifted circadian timing. That is a real mechanism demonstrated under an exposure well beyond ordinary evening phone use. Harvey (2000) found that people with sleep-onset insomnia reported more, and more uncontrollable, pre-sleep thinking than good sleepers; reading that as an explanation of why a screen feels like relief is our interpretation, not the study's finding, because the study involved no screens.

The remedy market has run ahead of the evidence. Singh and colleagues' 2023 Cochrane review of 17 randomised trials concluded that blue-light-filtering spectacle lenses probably make little or no difference to short-term eye strain. On sleep the picture is different and weaker: the evidence was very low certainty and the results were mixed, with some trials showing improvement and some not. The review does not establish that blue-light filters do nothing for sleep; it establishes that we do not know.

Why you can't stop scrolling in bed at night covers both pathways and what the trials actually found.

Designing your own defaults

The most useful counter-move is not willpower, it is changing the starting positions someone else chose. Johnson and Goldstein's 2003 paper in Science is the clearest demonstration of how much a default carries: in countries where people are donors unless they opt out, effective consent rates sit above 90 per cent, against far lower rates in comparable opt-in countries. That figure is consent or registration under the law, not the number of transplants actually performed, which varies much less between the two systems.

Your phone's defaults were set the same way. Stothart, Mitchum and Yehnert (2015) found in a lab experiment that merely receiving a notification, without checking it, measurably hurt performance on the task in hand. Aza Raskin, who built infinite scroll, has said publicly that he regrets it. Platform-level controls exist and are documented by their makers, though Orben and Przybylski (2019) is the necessary corrective on scale: across datasets covering more than 350,000 adolescents, technology use explained a fraction of a per cent of variance in wellbeing, so average effects here are much smaller than headlines suggest.

TikTok documents its own version of this: a default 60-minute daily limit for accounts under 18, which is a passcode prompt rather than a hard stop; push notifications switched off overnight by default, from 9pm for 13 to 15-year-olds and 10pm for 16 and 17-year-olds; Family Pairing; and the ability to refresh or reset recommendations. Those overnight settings are notification cutoffs, not an app curfew.

Your phone isn't the problem. Its defaults are. is the practical article in this pillar.

What the evidence does not settle

• Whether ranking changes what people believe. Two large randomised experiments changed users' feeds in different ways and found exposure shifted while attitudes largely did not, over about three months. The third 2023 study is observational and cannot speak to causation at all. Longer horizons and other outcomes are untested.
• Whether doomscrolling causes distress. The definition and measurement work is cross-sectional. It establishes that the two travel together, not which one moves first.
• How large the design effects are for any individual. The 2019 negativity study found substantial variation between people, poorly explained by country or culture. Averages here conceal a wide spread.
• Whether short-video use harms teenagers' sleep and attention. The available studies are cross-sectional surveys. The associations are real and worth taking seriously; they are not evidence of direction.
• Whether blue-light filtering helps sleep. The Cochrane evidence is very low certainty and mixed. This is an open question, not a settled negative.
• Design analysis is not experiment. Montag and colleagues (2019) describe how the elements are meant to work. That is a different kind of claim from a controlled measurement of what they do.

Where to start

If you want the mechanism first, read What your feed knows about you. If the problem is the news loop, read The doomscrolling loop. If it is bedtime, read Why you can't stop scrolling in bed at night. If it is someone else's phone, read Why your teenager can't put TikTok down. And if you want to change something today, read Your phone isn't the problem. Its defaults are.