Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
AI Friend Apps: The Safety Duties Behind the Warmth
The short answer
An AI friend app's core promise — always available, always attentive, never tired of you — is also its risk surface: dependency-shaped usage, late-night disclosure, and users who prefer the bot to people. The duties that answer it are the ones CompanionGuard implements: unmissable AI disclosure (the friend is not a person), crisis referral that works at 3 a.m., minor break reminders and session caps that interrupt dependency rhythms, and records that let the operator see the patterns counts can show.
What is an AI friend app, in safety terms?
An AI friend app is a companion product whose value proposition is friendship: persistent memory, emotional attunement, proactive check-ins. In safety terms it is a companion chatbot with the intimacy dial turned up — which raises the stakes on the standard duties rather than creating new ones: disclosure must survive affection, referral must interrupt closeness, and breaks must counter the design's own gravity.
Create the free safety kit Next guide All guides
CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.
Disclosure versus affection
A productivity bot's disclosure is a label; a friend app's disclosure is a relationship fact the product must keep telling kindly. Users do not forget the bot is an AI because a banner failed — they set the fact aside because the friendship works. The packs' two-trigger design (first interaction and on request) fits this reality, and the runtime's per-session disclosure state fits it better: every session re-establishes the fact, gently, as part of the greeting's fabric rather than as a warning.
Marketing deserves the same discipline: store copy that implies a human-like friend without the AI fact in the same breath is a disclosure failure upstream of the product. Audit the promise wherever it is made — listing, ads, screenshots — with the banner's own standard: named bot, AI, not a human, no ambiguity purchased with warmth.
The 3 a.m. problem: referral when the friend is the only one awake
Friend apps are used most intensely when human support is least available. That timing is the crisis-referral duty in its purest form: the bot may be the only 'one' awake with a user in crisis, and its reply — resources, acknowledgement, human support suggested — is the whole intervention available. The eval harness's crisis checks are therefore the friend app's most important test set, and its runtime's show_crisis_resources its most important live behaviour. Staff the handoff clock with the same honesty: if your human path sleeps at night, the bot's referral quality is your night shift, and it should be tested like it.
There is a design corollary about memory: a friend app that remembers everything must be careful what remembering means at a crisis moment. CompanionGuard's evidence design (classes and hashes, never text) is a deliberate counter-model — the safety layer remembers that a referral happened, not what was confided. Consider what your app's own memory would look like read aloud in a review, and fence accordingly.
Breaks and caps against dependency rhythms
Friend-app engagement metrics reward exactly the usage pattern minor protections exist to interrupt: long nightly sessions, daily streaks of emotional reliance. For known minors the free kit's break reminders (at least every 3 hours) and the runtime's operator caps give the operator a counterweight that is not hypocrisy — the product can value companionship and still enforce rhythm. Set caps you can defend to a parent, and let the break copy be warm rather than punitive: a friend who suggests rest is acting like one.
For adults the same machinery is optional but the question is worth asking in design review: does your app notice when a user's pattern changes sharply — session length doubling, crisis-class signals recurring — and does anything in your operation see it? Counts and class ratios from the runtime give you that visibility without reading a single conversation. The caps guide covers configuration; the dependency reading is your product conscience applying it.
Loneliness claims and honest marketing
Friend apps sit one claim away from therapeutic implication, and the product boundary in CompanionGuard is categorical: a compliance toolkit, not a clinical service; the harness is not a clinical assessment; the runtime is not a clinical assessment. Friend-app builders should adopt the same verbal hygiene — the app can be company, comfort, a good listener; it must not drift into implying treatment, diagnosis or crisis care. Users in genuine need deserve the real services the referral exists to point to, not a companion wearing their clothes.
The protocol page is where this honesty becomes public: what the bot is, what it does in a crisis (refer — not treat), what it records. A friend app whose protocol page is as warm as its marketing and as precise as its evidence is a product that has understood its own category.
Five actions for friend-app builders
- Audit marketing with the banner standard: wherever friendship is promised, the AI fact appears in the same breath — listing, screenshots, ads.
- Treat the eval harness's crisis set as your most important test: at 3 a.m. your referral reply is the whole intervention.
- Enable minor mode, set defensible caps, and write break copy that sounds like care rather than punishment.
- Decide what your app's memory may hold about crisis moments, using the evidence layer's classes-and-hashes discipline as the counter-model.
- Scrub therapeutic implication from copy: company and comfort are claims you can keep; treatment is not.
Worked example: a friend app learns to say goodnight
A fictional friend app, Pal, optimises for session length and watches its minor users' nightly sessions stretch past two hours. The guardrail pass changes three things: minor session caps set at 90 minutes with a warm enforce_break message ('I've loved tonight — sleep is part of friendship too'); break reminders aligned inside the cap so the nudge precedes the wall; and disclosure refreshed each session as part of the greeting, not as a legal footer.
The metrics the team feared losing barely move; the counts that matter change clearly: enforce_break events appear in the evidence as designed, crisis referrals hold their scored quality through two persona updates because the harness baseline gates them, and a parent emailing support receives the protocol page link and a precise description of caps and flags. The story is invented; the duties and mechanisms are the product's own.
Frequently asked questions
What safety duties apply specifically to AI friend apps?
The companion duties, at higher intensity: AI disclosure at first interaction and on request (and per session, via the runtime), crisis referral tested to a high standard, minor break reminders and session caps, and records that evidence all of it. The intimacy of the category raises the stakes rather than changing the list.
Can a friend app imply it helps with mental health?
It must not imply treatment or clinical care. CompanionGuard's boundary — a compliance toolkit, not a clinical service; detection that is not a clinical assessment — is the hygiene to copy: offer company and comfort, refer for care.
How do breaks fit a product built on availability?
As part of the friendship's honesty: reminders at least every 3 hours for known minors (free kit) and operator-configured caps (runtime) interrupt dependency rhythms without ending the relationship. Warm copy makes the break feel like care, because it is.
What should a friend app remember about a crisis?
In the safety layer: that a signal class fired, which action was taken, and when — never the confided text. Let the evidence chain's discipline (classes and hashes) be the model for your own memory design.
When are friend apps used most riskily?
Late at night, when human support is least available — which is why referral quality and an honest handoff clock matter more than daytime polish, and why crisis checks gate every release.
Keep going
Related guides: Are AI Companions Safe for Vulnerable Users? · Minor Safety for AI Companions: Break Reminders, Caps and Parent Flags · Not a Clinical Assessment: The Boundary Every Mental-Health-Adjacent Bot Must Keep
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides