Substance + claimed effects

Addictive app design is the deliberate application of behavioral-psychology techniques — variable-ratio reinforcement, intermittent rewards, social-approval feedback, push-notification triggers, infinite scroll, autoplay, streaks — by social and mobile applications to maximize user engagement (session count, session length, daily active users). The blueprint is explicit in industry literature: Fogg's behavior model B = MAT (motivation × ability × trigger) Fogg 2003 and Eyal's Hook model (trigger → action → variable reward → investment) Eyal 2014. Claimed downstream effects span attention (focus erosion, attention residue, cognitive cost even when notifications are ignored), sleep (delayed onset, fragmented architecture, suppressed melatonin from pre-sleep use), mood (depressive symptoms, anxiety, social-comparison hits), time use (US adults average 4–5 h/day on smartphones; adolescents 7–9 h on screens), relationships (phubbing, reduced conversational closeness), and self-perceived control (felt compulsion, automatic phone-checking, the gap between intention and behaviour).

Evidence by addressing question

mechanism

Variable-ratio reinforcement is the most resistance-to-extinction schedule in operant conditioning — the schedule that runs slot machines. App features map onto it directly: pull-to-refresh delivers an unpredictable inbox, like-counts arrive asynchronously, the For You feed reshuffles on every open. Neural mechanism: phasic dopamine neurons fire to prediction error, not to reward per se — unpredictable rewards generate larger transient signals than predictable ones Schultz 1998. Habit formation in the basal ganglia is driven by this same prediction-error signal under repetition; the result is automaticity — the behaviour fires from a cue (notification, idle moment, mild anxiety) without conscious decision.

Two design layers stack on top of the variable-reward core. (1) Trigger inflation: push notifications, badges, vibration, and social cues ("X started following you") manufacture cues that didn't exist before. Fogg's model says behaviour requires a trigger at the moment motivation × ability is high enough — apps subsidise the trigger side aggressively. (2) Friction removal: autoplay, infinite scroll, one-tap login, swipe-to-next eliminate the natural stopping points that would let a user notice they wanted to stop. Eyal's Hook model adds an "investment" phase — every like given, every account followed, every story posted increases switching costs and reload urgency on the next visit Eyal 2014.

evidence

The strongest causal evidence comes from three converging study designs.

Allcott et al. 2020 ran an RCT paying ~2 800 Facebook users to deactivate for four weeks before the 2018 US midterms Allcott et al. 2020. Deactivation freed roughly 60 minutes/day, reduced post-experiment news knowledge and political polarization, and produced a small but significant improvement in subjective wellbeing (~0.09 SD on a wellbeing index combining happiness, life satisfaction, depression and anxiety scales). 80% of participants reported the experience was positive; ~20% chose to use Facebook less afterward.

Hunt et al. 2018 randomised 143 undergraduates to either limit Facebook, Instagram and Snapchat to 10 minutes per platform per day for three weeks, or use normally Hunt et al. 2018. The limited group showed significant reductions in loneliness and depressive symptoms relative to controls, with the largest effects in participants who entered the trial with higher baseline depression.

Braghieri, Levy and Makarin 2022 exploited Facebook's staggered college-by-college rollout 2004–2006 as a natural experiment, comparing mental-health outcomes (National College Health Assessment) before vs after Facebook arrived on each campus Braghieri et al. 2022. Facebook introduction caused an increase in severe depression of ~7% and in anxiety disorder of ~20% relative to baseline; effects were larger for students predicted to be susceptible to unfavourable social comparison.

The mental-health literature on adolescents is more contested. Twenge et al. 2018 documented post-2010 increases in US adolescent depressive symptoms, suicide-related outcomes and suicide rates that tracked the smartphone adoption curve Twenge et al. 2018. Orben & Przybylski 2019 applied specification curve analysis across three large datasets and found digital technology use explained only ~0.4% of variation in adolescent wellbeing — comparable in magnitude to eating potatoes Orben & Przybylski 2019. The debate is partly about what's measured: aggregate self-report screen time has small effects; passive social media consumption specifically (the design-heavy use pattern) has larger ones. Haidt 2024 synthesises the alarmist position, arguing the early-2010s phone-based-childhood transition is causal in the adolescent mental-health collapse Haidt 2024; the academic debate over effect sizes remains active.

attention

Ward et al. 2017 showed that the mere presence of a participant's own smartphone, silent and face-down on the desk, reduced working-memory capacity and fluid-intelligence scores relative to leaving the phone in another room — even though all participants reported no impact on their performance Ward et al. 2017. The proposed mechanism: ongoing inhibitory effort to not attend to the phone consumes the same executive resources the task requires.

Stothart et al. 2015 demonstrated that simply receiving a notification (no checking, no looking) tripled errors on a sustained attention task — the cost was comparable to actively using the phone Stothart et al. 2015. Ophir, Nass & Wagner 2009 found heavy media-multitaskers performed worse on task-switching, filtering of irrelevant stimuli, and memory tasks — possibly trait differences, possibly consequences Ophir, Nass & Wagner 2009. Leroy 2009 established the attention-residue mechanism: switching tasks leaves a cognitive residue of the prior task that impairs performance on the new one, with the magnitude depending on how much the prior task was completed Leroy 2009. App design generates frequent involuntary switches; the residue tax compounds across the day.

sleep

Chang et al. 2015 ran a within-subjects trial comparing five evenings of pre-bed light-emitting eReader use to five evenings of print book Chang et al. 2015. The eReader nights produced delayed sleep onset by ~10 min, suppressed evening melatonin by >55%, delayed circadian phase by ~1.5 hr, reduced REM sleep, and impaired next-morning alertness despite equivalent total sleep time. The mechanism is two-pronged: short-wavelength light suppresses melatonin via intrinsically photosensitive retinal ganglion cells; pre-sleep cognitive arousal from algorithmic feeds delays the natural settling that precedes sleep onset. Observational evidence converges: bedtime smartphone use predicts longer sleep latency, shorter total sleep time and reduced sleep quality across age groups, with the largest effects in adolescents whose phase is already drifting later for developmental reasons.

relationships

Przybylski & Weinstein 2013 ran two field experiments showing that the mere presence of a mobile phone during a face-to-face conversation reduced felt closeness, conversation quality and perceived empathy from the partner — even when the phone wasn't in use and the conversation was about a personal topic Przybylski & Weinstein 2013. Roberts & David 2016 surveyed 145 adults in committed relationships and found "phubbing" (partner phone-snubbing) predicted lower relationship satisfaction; the path ran through conflict over phone use and then through depressive symptoms Roberts & David 2016. The displacement mechanism is straightforward: attention is rivalrous, and intimate conversation requires the kind of sustained, undivided attention that a phone in the field of view structurally undermines.

protocol

Direct trials of attention-restoring interventions are thin; most evidence is mechanistic and behavioural. The shape of an effective protocol follows Fogg's model in reverse — to reduce a behaviour, attack triggers and increase friction Fogg 2003. Practical levers, ranked roughly by leverage per unit of effort:

• Notifications off by default. Allow only human-to-you communication (calls, direct messages from a small whitelist). Strips the trigger side of B = MAT for every other category. Stothart's data implies large gains: ~3× error reduction on attention tasks corresponds to a real felt-experience lift.
• Remove the worst offenders from the homescreen. Move social and video apps to a folder behind a search; the swipe-to-open habit dies when the cue isn't there. Logging out adds a 30-second password barrier — Fogg's "ability" axis — that breaks the automaticity loop.
• Delete the app, keep the browser. Browser versions are intentionally degraded: no push, no algorithmic notification re-engagement, slower load, no infinite scroll on some platforms. The friction differential is large.
• Phone out of the bedroom. Removes the bedtime use loop (sleep evidence above) and the morning-scroll opening of the day. Cheap analogue alarm clock replaces the phone function.
• Grayscale display. Removes colour reward salience; the literature is mostly observational, but the engagement-design rationale is sound — colour is part of the variable-reward payload.
• Hard time limits (Screen Time / Digital Wellbeing). Modest evidence, because the override path is one tap; the value is more as a metric than a wall.
• External commitment devices. App blockers like Freedom, Opal, One Sec impose friction the user can't easily lift mid-session.

The bedroom-phone-removal lever pairs with the established sleep-hygiene literature (light exposure, fixed wake time). The notification audit pairs with the deep-work / focus literature; attention residue (Leroy) means a cleared notification panel still costs working memory until the next deliberate context switch.

contraindications

None medical. The closed contraindication vocabulary in this catalogue (pregnancy, cardiac, kidney etc.) doesn't apply to a behavioural-avoidance entry. Eating-disorder history is an aggravating factor — image-driven feeds (Instagram, TikTok) escalate eating-disorder risk via social comparison and pro-ED content discovery, documented in internal Facebook/Meta research surfaced in the 2021 Haugen disclosures and in independent literature on appearance-comparison platforms. This belongs in the article as a population-variability note rather than a meta-level contraindication.

misconceptions

Three common misreadings.

• "It's a willpower problem." Fogg's model says behaviour = motivation × ability × trigger, and you cannot will yourself against adversarial trigger-and-friction design with intelligence-tier teams iterating on your psychology. The lever is environment, not willpower Fogg 2003.
• "Screen time is the metric." Aggregate screen time has small effects (Orben & Przybylski). Passive consumption of algorithmic feeds has larger ones. A 3-hour video call with a friend is not the same intervention as 3 hours of TikTok. The articles that find null effects often pool both Orben & Przybylski 2019.
• "A digital detox fixes it." Short cessations confirm the felt-experience benefit but revert to baseline once the device returns and the same triggers, app placements and notification settings reload the loop. Persistent architectural change beats time-bounded cessation.

failure-modes

Common backslide patterns. (1) Substitution: TikTok deleted, YouTube Shorts adopted; the variable-reward mechanic is the substance, not the brand. (2) Notification leakage via smartwatch — wrist haptics defeat phone-side notification audits. (3) Browser workarounds with no friction added (logged-in browser tabs that survive restarts). (4) Re-enabling for a "specific reason" (event, deal, message) and never disabling again — the architectural change has to be sticky. (5) Time-limit override fatigue: the "15 more minutes" tap becomes reflexive within a week.

stakes

The literature supports a felt-experience forecast at multiple time horizons. Week-to-week: working-memory decrement Ward et al. 2017, attention-task errors Stothart et al. 2015, sleep-onset delay and reduced REM Chang et al. 2015, partner-relationship friction Roberts & David 2016. Month-to-year: the wellbeing decrement Allcott measured at four weeks compounds with the time-use cost (one hour/day × 365 = ~15 waking days/year). Decade: the population-level shift Twenge documents for adolescents Twenge et al. 2018; for adults the long-run evidence is weaker but the time-displacement argument carries — what isn't built, learned, or maintained because of the hours.

payoff

Allcott's deactivation arm provides the cleanest quantification at four weeks: ~60 minutes/day returned, ~0.09 SD subjective wellbeing lift, ~20% post-trial behaviour change Allcott et al. 2020. Hunt's three-week limit RCT showed loneliness and depression reductions concentrated in the higher-baseline subgroup Hunt et al. 2018. Mechanism-derived predictions on shorter horizons: notification audit + bedroom-phone removal should produce a within-week felt-experience shift (sleep, morning mood, focus). Architectural changes (homescreen, app deletion) compound over months as automaticity is replaced by deliberate access.

The credibility range

Optimist case

The strongest pro-position: aggregate wellbeing effects are small (Orben & Przybylski's r ≈ 0.05; smaller than wearing glasses) and the public discourse is in a moral-panic phase that exaggerates risk. Connection benefits are real and underweighted — long-distance friendships maintained, isolated populations (LGBTQ+ teens in unsupportive areas, the homebound, geographically dispersed families) materially helped. Many people use phones moderately without measurable harm. The RCTs that find negative effects are short (3–4 weeks) and may capture novelty/withdrawal more than steady-state. Adolescent mental-health trends have plausible alternative drivers (sleep deprivation independent of phones, school pressure, post-2008 economic conditions, post-2020 pandemic isolation). Banning or restricting access for adolescents risks paternalism and may not work — substitution to other media is the historical pattern.

Skeptic case

The strongest counter-position: multiple study designs (RCT, quasi-experiment, observational, mechanistic) converge on small-to-moderate harms. The time use is enormous (4+ hours/day for adults, 7+ for teens) — even a small per-hour effect aggregates to a large total. Industry internal research (the 2021 Haugen disclosures on Instagram and teen-girl mental health) confirms the operators know the effects and have suppressed publication. Designers' revealed preferences (refusing the products for their own children) is data. The mechanism is well-characterised (variable-ratio + dopamine prediction error + trigger inflation) and not in dispute. The Braghieri natural experiment is identification-strong: Facebook's college rollout was staggered for reasons unrelated to mental health Braghieri et al. 2022, and the effects are not small. Felt-experience reports are nearly universal: people describe automaticity, regret, and inability to stop — which in any other category (substance use, gambling) would be sufficient evidence of addictive design.

The author's call

The entry lands on the skeptic side. The RCT and quasi-experimental evidence is good enough on its own; the mechanism is uncontested; the felt-experience reports are universal; and the time displacement is too large to ignore even at small per-hour effects. The Orben & Przybylski-style "potatoes" framing misreads what's being measured — aggregate screen-time variance is the wrong denominator for effects driven by specific design features in specific app categories. The action is avoid applied to the design pattern, not to phones or to the internet. The article framing emphasises architectural change (notifications, homescreen, bedroom) over willpower or detox cycles. Evidence rating: 4 (multiple RCTs + quasi-experimental + mechanism). Controversy rating: 3 (active debate among reasonable experts on effect magnitude, especially in adolescents).

Stakeholder + incentive map

• App companies (Meta, Google, ByteDance, X, Snap): revenue is engagement-time × ad load. Every design lever above is optimised against. Internal research is selectively released (Haugen disclosures the canonical example).
• Ex-insider critics (Tristan Harris / Center for Humane Technology, James Williams, Aza Raskin): formerly inside Google/Apple/Mozilla; now run advocacy organisations. Credible but selection-biased — the ex-insiders who didn't quit don't write the books.
• Persuasive-design researchers (Fogg, Eyal): built the playbook; Eyal partly recanted ("Indistractable" 2019 walked some of "Hooked" back). Their work is simultaneously the canonical mechanism description and the canonical industry training material.
• Alarmist academics (Twenge, Haidt): book-driven public arguments; Haidt running an advocacy push for school phone bans and under-16 social media restrictions.
• Moderate academics (Orben, Przybylski, Odgers): push back on effect magnitudes and emphasise heterogeneity; argue moral panic mis-allocates concern.
• Regulators: Australia (under-16 social media ban, 2024), UK (Online Safety Act), EU (DSA), US state-level (Florida, Utah, NY phone-in-school bans). Growing political appetite.
• Schools, paediatricians: increasingly recommend phone-in-locker policies; AAP guidance evolving.

Population variability

• Adolescents (12–18): highest risk. Developing prefrontal cortex, peak sensitivity to social-comparison feedback, sleep timing already shifted late developmentally. Female adolescents more affected on image-driven platforms (Instagram, TikTok); the Braghieri effects were largest for those predisposed to unfavourable comparison Braghieri et al. 2022.
• Young adults (18–29): highest absolute use; the Allcott and Hunt samples sit here. Strong effects on time use and focus; mental-health effects present but smaller than in adolescents.
• Working adults (30–60): the cognitive-attention literature applies; phones in workplaces measurably degrade output even when not in active use Ward et al. 2017. Relationship-quality effects load here (Roberts & David sample) Roberts & David 2016.
• Older adults (60+): generally lower problematic use; Facebook stickier than newer platforms. Connection benefits (against isolation) larger in this group.
• Pre-existing depression/anxiety: bidirectional risk amplification. Hunt found largest depression reductions in higher-baseline subgroup Hunt et al. 2018.
• Eating-disorder history: image-driven feeds escalate via comparison and content-discovery loops; the population-specific advice is stronger here than the general protocol.
• ADHD / executive-function differences: Fogg's friction argument predicts disproportionate vulnerability; the protocol's architectural interventions are correspondingly more useful.

Knowledge gaps

• Long-run (5–10 year) RCTs are impossible; identification rests on natural experiments (Braghieri) or pre/post cohort comparison (Twenge).
• Which design features cause which effects — most studies measure aggregate use, not specific feature exposure. Disentangling notifications vs feed algorithm vs autoplay vs social comparison is largely unstudied.
• Children under 13 — ethics constrain experimental work; almost all causal evidence is in adolescents or adults.
• Effect of partial interventions (one app deleted, notifications muted) vs total cessation — RCTs use binary deactivation; the architectural-change protocol most readers will actually run sits in between.
• Effect of replacement activities — Allcott found the freed hour was redistributed (offline social, TV, reading) but the redistribution matters for outcomes and is understudied.
• Whether ex-insider self-reports (designers' children's phone restrictions) reflect informed risk-assessment or status signalling.
• What would change the author's call: a well-powered, multi-month RCT with a partial-architectural intervention arm finding no effect on subjective wellbeing or attention. Or a comprehensive replication of the Braghieri natural experiment with a null result.

Scope. The brief named six consequences — attention, sleep, mood, time use, relationships, self-perceived control. The article covers all six, weighted by evidence strength: attention is the centre of gravity (focus 5 in meta), sleep and mood carry their own paragraphs with cited trial data, time use sits inside stakes and payoff as the "hour a day" anchor, relationships ride the Przybylski / Roberts-and-David pair in stakes, and self-perceived control is the dek's opening frame ("the loop you blame on your own willpower").

Narrowing. The adolescent case is touched in evidence (Twenge / Orben & Przybylski / Braghieri / Haidt) and misconceptions but not as its own substance. The reader-facing voice in this entry assumes an adult agent who can change their own architecture; the under-18 case has different stakeholders (parents, schools, regulators) and different leverage points (age-gating, school phone bans, family device policy). Flagged below as a separate-entry candidate rather than wedged in.

Rating difficulties.

• focus = 5 vs 4. Pushed to 5. The Ward / Stothart / Leroy stack measures the cost even when the user is trying not to attend; the substance is functionally an attention substance, and the meta would mis-signal at 4. Defensible at 4 on effect-size grounds alone, but the dominance criterion ("the kind of result that defines the entry") fits 5.
• mood = 4 vs 3. Three converging designs (Allcott RCT, Hunt RCT, Braghieri quasi-experiment) is more than three for "clear effect". Stayed under 5 because the per-study effect sizes are small-to-moderate, not transformative.
• controversy = 3 vs 4. Held at 3. Mechanism is uncontested; the live argument is over magnitude in adolescents, which is a real but bounded fight, not a foundational paradigm split.
• evidence = 4 vs 5. Stayed at 4. Strong RCT + quasi-experimental coverage, but no Cochrane-tier review and no consensus guideline yet — 5 is reserved for that.
• longevity = 1. Indirect, via the sleep / loneliness / chronic-stress pathways that each carry their own mortality signal. Defensible at 0; chose 1 to keep the multi-decade picture honestly on the table without claiming a direct effect.

Excluded.

• Industry deep-dive (Haugen disclosures, internal Meta research on teen-girl mental health, congressional hearings). Real and load-bearing for the skeptic case in research, but not what the reader does on Monday. Mentioned in the dossier's credibility range and stakeholder map; kept out of the article.
• Platform-by-platform comparison (TikTok vs Instagram vs X). The substance is the design pattern; brand comparisons date fast and miss the point.
• Children under 13. Ethics constrain the literature; the action lever isn't with the child reader anyway.
• ADHD-specific protocol. Touched obliquely (the architectural-friction argument is more useful in executive-function variability); not given its own audience block because the protocol generalises and the audience scoping would over-medicalise it.
• Specific app-blocker comparison. Named four options in protocol; deeper comparison rots quickly and belongs in a "tools" sidebar if at all.

Separate-entry candidates.

• Adolescent social media — different stakeholders, different protocol (parental, school, regulatory), different evidence base (Twenge / Haidt vs Orben / Odgers / Przybylski). Worth its own entry.
• Notification audit — the highest-leverage piece of the protocol could be its own short entry, linked from here.
• Doom-scrolling — the specific bedtime / news-during-crisis variant has a different felt-experience profile and could earn its own treatment.
• Phone-free bedroom — pairs with sleep hygiene, could live as a tight standalone.

Future links. sleep-debt, morning-sunlight / light-exposure, deep-work, meditation, boredom-tolerance — wire out-of-scope to these once they exist. phubbing and the adolescent social media entry above are forward candidates from stakes.

Voice note. The stakes / payoff pair leans heavily on social-mirror voice ("people you barely know stop asking…", "your partner notices…"). That's the per §5c bar; flagging it because the temptation in this category is to drift into "you'll feel transformed" wellness-influencer cadence, and the editor should push back on any line that crosses there in revision.