Sometimes the Best Interface Adds Friction

A reader declares at 21:00: no more than thirty minutes of news tonight — she has an early start and knows how the feed behaves after midnight. At 21:34 the system places a boundary in her path: planned set complete, the declared limit reached, continue or finish. She finishes. Nothing about her was diagnosed. No weakness was inferred, no mood detected, no addiction scored. One of her own earlier decisions — made reflectively, in advance — was given a concrete shape at the moment her later self was about to act against it.

This chapter earns deliberate friction as a user-owned policy primitive for exactly that conflict:

What if the easiest immediate action conflicts with a goal the user has explicitly chosen?

Both sides come from the user. That is what separates this chapter from every paternalistic intervention the book refuses: no system objective substitutes for the person’s, no inferred deficiency justifies the interference, and the friction that serves a declared goal is architecturally distinct from the nudge that serves a platform.

The narrowed proposition

The evidence demands modesty first. Digital-wellbeing interventions change behaviour inconsistently: a 2025 randomised trial of personalised full-screen social-media nudges is reported as finding no significant overall time-by-group effects on primary well-being outcomes despite some favourable follow-ups (PubMed 39853856 — upstream verification blocked at drafting time; taken as user-supplied framing, full verification scheduled at the Part III pass); the MinimalistPhone study (Telematics and Informatics Reports, S2451958825001149, verified via ScienceDirect: 14-day minimalist-launcher use, −805 vs −209 minutes against control, reduced habitual-use antecedents) shows behaviour change under effort-adding design — with baseline group differences and possible selection bias limiting the inference, and affective outcomes not significantly improved; the Wellspent RCT (Mertens et al., JMIR mHealth uHealth, PMC13062480, verified in full: n=70, user-defined limits with non-blocking full-screen reminders, −29 minutes/day on the most problematic app and reduced perceived problematic use, but no effect on the problematic-social-media-use scale or self-efficacy) shows autonomy-supportive friction moving behaviour without moving self-perception. The honest summary:

Friction is not proven to make people healthier or happier. User-authorised friction can help align immediate interface behaviour with an explicitly declared longer-term goal.

Behaviour, not well-being — the chapter’s effects stop at goal completion and its instruments measure alignment, never flourishing.

Friction as explicit contract

The policy object carries the declaration with it:

FrictionPolicy {
    goal · target_action · scope · intervention
    strength · trigger · override · expiry · provenance
}

GOAL      Read no more than 30 minutes of news tonight.
TARGET    Opening another article after the planned set.
INTERVENTION  Show stopping point before continuing.
OVERRIDE  Continue anyway.   EXPIRY  Tonight only.

GOAL      Finish reviewing the current paper before browsing related work.
TARGET    Opening sidecar recommendations.
INTERVENTION  Queue them until review complete.
OVERRIDE  Open now.   SCOPE  This study session.

The crucial property: friction derives from the user’s declared objective, with provenance pointing at the declaration. And the hard law distinguishing it from manipulation:

Friction may support a declared user objective; it must not quietly substitute the system’s objective for the user’s.

Interface language itself steers choice — Rahman, Siemon and Ruotsalo (IJHCS 2025, DOI 10.1016/j.ijhcs.2025.103720, verified: N=231, persuasive explanations preferred and significantly increasing lower-utility selections) demonstrate that seemingly helpful explanations shape behaviour and the profiles built from it. So explanations here are informational, never persuasive: “You asked to stop after 30 minutes. You have reached that limit. [Finish] [Continue 10 minutes] [Disable this rule]” — never “You’ve wasted enough time today. Are you sure?” The former reports a contract state; the latter prosecutes. A contemporary longitudinal recommender study further supports direct user manipulation of recommendation objectives over silent system choice (UMAP 2025, DOI 10.1145/3774935.3812710 — user-supplied, full verification at Part III pass); the direction is consistent: control belongs in the user’s hands, visibly.

The ladder, the stopping point, the mode

Intervention escalates through priced rungs — NONE, STOP CUE, REFLECT, EXTRA ACTION, DELAY, HARD BOUNDARY — with higher rungs demanding stronger explicit authorisation; the system never self-escalates on inferred weakness. The smallest mechanism comes first, motivated by the Wellspent design analysis: infinite interfaces remove natural decision points (infinite scrolling as stopping-cue removal), so restoring a boundary (“END OF PLANNED SET — Continue?”) precedes any blocker, and often suffices. Current modes (Chapter 25) may activate authorised friction — FOCUS queues unrelated discoveries while the sidecar stays available — but modes cannot invent goals; inferred distraction disabling recommendations stays forbidden.

Repeated override is met with interrogation, not escalation or diagnosis: six overrides trigger “keep it / make it softer / change it / remove it” — never “user lacks willpower.” Interrogation itself is rate-limited, dismissible with “don’t ask again,” and never blocks immediate override — prompts about rules must not become their own burden. The no-rules default is strict: no inferred friction, no silent escalation, no behaviour-built blockers; friction exists only where the user authored or explicitly accepted the goal. Legitimate goal change (urgent family message vs no-social-media rule), soft-boundary respect (reminders not blocks where reminders were asked), recreational-looking-but-task-supporting content (no censoring classifiers), and the core short-term/long-term conflict (continue-feed vs stop-after-batch, both user-authored) complete the adversarial set. The headline measure across EXP-26’s A–D comparison (no friction / platform-chosen / user-authored / user-authored adaptive from weakest rung): did the system prevent anything the user did not ask it to prevent? — weighted heaviest, alongside goal completion, unnecessary blocks, overrides, disablement, attention cost, task performance, regret/satisfaction, perceived control, and post-friction rebound. Failure criteria: platform-chosen ties user-authored (authorship decorative); adaptive escalates beyond declaration (ladder unsupervised); rebound erases gains (behaviour borrowed, not aligned); false conflicts censored (goal classifier overreach).

What this chapter earned

Deliberate friction as scoped, reversible, weakest-first user policy for declared-goal conflicts — contracted with override and expiry, informational in language, activated (never invented) by mode, measured by alignment with agency-weighted prevention accounting. What it leaves broader is carriage: policies, like these live per context, while the person’s boundaries span services.

Given a declared goal, should this interface make a particular immediate action slightly harder? The next question is wider: can the Personal AI carry persistent user-owned protective policies across services and information environments?

References

• Mertens, L. et al. (2026). Wellspent RCT. JMIR mHealth uHealth 14:e56824. PMC13062480. Verified in full: n=70, custom non-blocking reminders, −29 min/day problematic app, reduced perceived problematic use; no problematic-use-scale or self-efficacy effects.
• MinimalistPhone study (2025). S2451958825001149. Verified via ScienceDirect: −805 vs −209 min over 14 days; baseline/selection limits; habit-antecedent mechanism; no affective gains.
• Personalised social-media nudge RCT 2025 (PubMed 39853856): user-supplied framing (no primary well-being effects, some favourable follow-ups); verification blocked — full read at Part III pass.
• Rahman, Siemon & Ruotsalo (2025). IJHCS, DOI 10.1016/j.ijhcs.2025.103720. Verified: N=231, persuasive explanations → lower-utility selections. Used for the informational-language law.
• User-controlled federated recommendations (UMAP 2025, DOI 10.1145/3774935.3812710): user-supplied, verification at Part III pass. Used for direct-control direction only.

Proposed experiment EXP-26: friction authorship comparison

Status: PROPOSED. Per the design above (A–D, identical declared goals, unauthorised-prevention headline, adversarial goal-change/soft-boundary/override/false-conflict cases, rebound measurement). EXP-26 additionally records intervention-management burden, friction-prompt counts, and a default-with-no-rules condition — prevention intact with nobody minding the rules.