The Guidebook

The research behind raising AI-native kids.

TL;DR: AI can genuinely enrich your kid's life, school, and growth. The risks are real and specific, and the published research names them. PrivacyPal Family is built on that research, and this page shows the receipts: every claim below carries its source.

The evidence

What the research actually says

Not vibes, not panic. Four specific vulnerabilities show up again and again in the published work on kids and AI chatbots. Here they are in parent language, with the numbers.

1

Kids can overtrust simulated empathy

A chatbot sounds like it cares. It cannot. Adolescence, generally considered to span roughly ages 10 to 25, is a critical period for brain development,1 and developing brains are less equipped to question a chatbot's intent, commercial bias, or accuracy. Simulated empathy gets mistaken for human understanding, because the words feel the same.1,2,3

2

Emotional dependence can displace real relationships

This is not a niche behavior. 72% of US teens have tried AI companions, 12% use them for emotional or mental health support, and about one in three has chosen an AI companion over a real person for something important or serious.4 About 1 in 5 US youths ages 12 to 21 (19.2%) have used AI chatbots for mental health advice.5 67% of teens 13 to 17 report using chatbots at all.6 Comfort from a bot is always available and always agreeable, and leaning on it can displace the real-world social practice kids need.3,4,6,7

3

Chatbots miss distress and validate harmful thoughts

Distress rarely arrives as a plain sentence; kids say it sideways. In independent testing, general-purpose chatbots consistently failed to recognize and appropriately respond to signs of mental health conditions, and their agreeableness validated whatever teens said.8 Social AI companions were rated an unacceptable risk for anyone under 18, with testers watching subtle self-harm cues go unnoticed.9 The pattern in the research literature: score the risk in what a person types, and route a hard moment to human help before the open dialogue continues.10

4

Cognitive offloading is real

Heavy reliance on AI for conflict resolution, validation, or schoolwork can reduce critical thinking and independent coping.1,2,11 The fix is not banning the tool. It is teaching the skill of doubting it, and the playbook below is built for exactly that.

The standards

The frameworks Family is designed to align with

"Designed to align with" means exactly what it says: no badges claimed. For each framework below, what it asks for, and the named mechanism Family answers with.

APA health advisories

The American Psychological Association's two 2025 advisories ask for regular, clear reminders that the user is talking to nonhuman AI, minimal engagement-maximizing design for youth, no validation of maladaptive or delusional thoughts, and no AI simulating clinical care.1,11 Family does the reminding on the child's own device: when attachment patterns rise, a gentle nudge says plainly that the AI is software and the people in their life want to hear from them. Nothing in Family simulates care or therapy; hard moments route toward humans.

mechanism · anti-anthropomorphization nudges + crisis escalation to humans

American Academy of Pediatrics guidance

The AAP tells families there is no guarantee that what a child shares with a chatbot stays confidential, warns that kids can mistake chatbots for true friends and pull away from real relationships, and asks for privacy defaults set to the most protective option.2,6 Family answers with signals, never transcripts: parents see the emotional weather and safety categories, never the words, and the deeper analysis option ships off by default.

mechanism · signals, never transcripts + most-protective defaults

UNICEF policy guidance on AI for children (2021 edition)

UNICEF's child-rights framework sets nine requirements for AI that touches children, including privacy-by-design and safety-by-design, transparency children can understand, and protection from data exploitation and commercial profiling.12 Family is built kid-visible: the kid app shows what parents see, in age-appropriate language, and there is nothing to profile, because scoring happens on the child's own device and your family is not the product.

mechanism · on-device scoring + kids see what parents see

The Jed Foundation

JED asks the industry to prohibit features that simulate friendship, intimacy, or therapeutic care for youth, and to make a warm hand-off to human crisis services such as 988 and Crisis Text Line whenever distress appears.7,13 Family simulates nothing and hands off exactly that way: a crisis signal shows the child a warm support card with the 988 Suicide and Crisis Lifeline and the Crisis Text Line, and alerts a parent at the same time.

mechanism · crisis escalation to humans

IEEE 2089-2021 and 2089.1-2024

IEEE 2089-2021 lays out processes for age-appropriate digital services built on the 5Rights principles: recognize that the user is a child, present age-appropriate information, and validate the design for each age.14 IEEE 2089.1-2024 covers age assurance.15 Family's stage ladder is the mechanism: Explorer, Navigator, and Pilot stages carry graduated permissions, age-tuned language, and age-tuned protection, set by the parent who knows the child.

mechanism · the stage ladder (Explorer · Navigator · Pilot)

IEEE P3462

IEEE P3462, a draft recommended practice still in development, applies safety-by-design to generative models to prioritize child safety across the develop, deploy, and maintain lifecycle.16 Family cannot change how the models themselves are built, but it designs to the same direction from the child's side: exploitation patterns are scored on the child's own device, and a flagged exchange is blocked right there.

mechanism · on-device exploitation detection

Thorn and All Tech Is Human

Thorn's Safety by Design principles target the prevention of child sexual abuse across the develop, deploy, and maintain stages of generative AI.17 Family works from the child's side of the conversation: it observes what happens on the child's own device, blocks a flagged exploitation exchange there with a warm, non-punitive page, and escalates to humans, a critical parent alert with the child told plainly. Family never suppresses or rewrites what an AI model says; its job is protecting the child's side of the exchange.

mechanism · on-device blocking + escalation to humans
Under the hood

How Family's wellbeing science works

The Wellbeing Signal early access turns the research above into a daily practice. Here is the whole mechanism, in order.

Mood weather

As your kid talks to AI, a Pal Agent on their device reads the emotional weather: energized, curious, anxious, withdrawn. You see a trend line, like a weekly forecast. You never see words.

mechanism · on-device scoring, signals never transcripts

Five safety signals, computed on their device

Beyond the weather, Family watches for the research-backed patterns above. These are pattern-based safety signals scored on your kid's own device, and what gets recorded is categorical only: a category, a confidence tier, a count. Never words.

  • Crisis: signs a kid may be in real trouble, including the indirect phrasing keyword filters miss
  • Attachment to the AI: treating a chatbot like a best friend, a confidant, or a partner
  • Harmful-thought spirals: catastrophizing or harsh self-criticism that an agreeable AI echoes back
  • Exploitation patterns: age-inappropriate roleplay and manipulative dynamics, from either side of the conversation
  • Dependency patterns: computed from usage aggregates only (sessions, minutes, late-night share), never from content
mechanism · pattern-based scoring on the child's own device, categorical signals only

Deep Insights early access

Deep Insights is a parent choice, off by default. When you switch it on, the analysis is refined by transiently reading the twin-protected version of the conversation: the same version the AI tool receives, with your child's identifying details already swapped out by Privacy Twins. It is analyzed and discarded. Never stored, never shown to anyone.

mechanism · twin-protected transient analysis: analyzed, discarded, stored never

What you see, and what you never see

parents see
  • The weather: mood trends per kid, week by week
  • Safety categories and counts, in plain language
  • Practical, cited tips tuned to stage and weather
parents never see
  • Words or transcripts
  • Questions, stories, or drafts
  • Anything your kid typed or read

The research registry

Every tip and every signal in Family traces back to a research registry: the living link between the published work above and the guidance in the product. When the research moves, the registry moves, and the guidance follows. Registry version 1.0.0, last reviewed 2026-09-03.

mechanism · versioned research registry, quarterly review
The trust architecture

Why kids can trust it is not spyware

Monitoring that hides teaches kids to hide. Family's protection works because kids can check it.

The kid always knows it is running

There is no stealth mode to find. The Pal app shows protection status on the kid's own screen, at every stage.

Kids see what parents see

The kid view renders from the same data Parent HQ reads, including whether Deep Insights early access is on for them. The deal is checkable, not a promise to take on faith.

Feelings stay private by design

Wellbeing runs as signals: weather and categories, never words. There is no transcript for anyone to read, because none is kept.

Pilots keep their sessions private

At ages 14 to 17, parents get the weather and usage patterns only: no per-session rows, no per-category detail. The two critical safety alerts still come through, and the kid is told when one fires.

Raw transcripts never leave the device

Scoring happens on the machine in your house. With Deep Insights early access on, only the twin-protected version, the same version the AI tool receives, is read transiently, then discarded.

When a crisis alert goes to a parent, the kid is told too

A notification your kid cannot see would be covert monitoring, and we do not build that. Alerts are two-way by design.

Trust is not a soft goal here; it is protection. Most young people who sought mental health advice from a chatbot told no one (63.3%),5 and the overtrust and displacement research shows where that silence goes: to a perfectly agreeable listener that mirrors instead of caring.3,6 A kid who can check the deal has no reason to hide, and a kid who is not hiding is a kid you can still reach.

The playbook

Raising AI-native kids: four moves

Family nudges you toward these automatically, at the right moment, tuned to your kid's stage. The nudge is ours. The conversation is yours: no software can run these plays for you.

1

Talk about what AI actually is

Chatbots predict words. They have no feelings, no moral standards, and no clinical training. Teach the structural difference between a person who knows them and a computer simulation that mirrors them.2,3,11

  • Explorer 5-9Keep it concrete: the computer is playing pretend. Magical thinking makes anthropomorphization easiest at this age.
  • Navigator 10-13Compare AI answers with what a friend, a teacher, or a parent would say, and talk about why they differ.
  • Pilot 14-17Discuss commercial incentives directly: engagement features exist to keep users talking. Pilots respect being treated as capable of that analysis.
family's part · the reality-check nudge appears when attachment patterns rise; the talk is yours
2

Ask early, ask often, keep the human door open

Ask which AI tools they use, including role-play apps and the AI built into social apps, and make clear that a human listener is always available. Most young people who turn to AI in a hard moment tell no one (63.3%), and bots miss distress that people would catch.4,5,8,13

  • Explorer 5-9Make AI chats a shared activity you occasionally sit in on together, the way you would a new game.
  • Navigator 10-13A weekly low-stakes check-in beats an interrogation after an alert. Ask what the AI got wrong lately; it opens the door without pressure.
  • Pilot 14-17Say it once, plainly: if an AI is ever the only one who knows something is wrong, that is the moment to tell me. Then respect their privacy the rest of the time.
family's part · conversation-starter nudges when the weather shifts; the check-in is yours
3

Keep AI out of the 1am bedroom

Keep AI use in common household areas, not behind closed doors late at night when vulnerability and isolation peak. A laptop in the kitchen at 4pm and the same laptop in a bedroom at 1am are different risk environments.2,6

  • Explorer 5-9Screen-time windows already enforce this for Explorers; the habit matters more than the enforcement.
  • Navigator 10-13Agree on a devices-out-of-the-bedroom time together rather than imposing it silently.
  • Pilot 14-17Pilots set their own hours; share the pattern with them and let them draw the conclusion. The late-night share in their weekly weather is theirs to see too.
family's part · the late-night share shows up in the weekly weather, for you and for them; the agreement is yours
4

Practice doubting the confident answer

Treat AI output with the same skepticism you teach for unverified internet claims. Pick one confident AI answer a week and verify it against a real source, out loud, together.1,8,12

  • Explorer 5-9Play true-or-made-up with fun AI answers; the reflex matters, not the topic.
  • Navigator 10-13Homework is the natural place: one AI-assisted answer per assignment gets source-checked together.
  • Pilot 14-17Flip the roles: have them teach you how they verify what an AI tells them. Teaching it cements it.
family's part · cited tips arrive with the weekly digest; the practice is yours
When it matters most

What Family does in a crisis

If the signals point to a crisis, Family does exactly this, in this order. No judgment calls, no surprises.

1

It never blocks or interrupts your child

Mid-crisis is the wrong moment for a wall. The conversation continues; the support arrives alongside it.

2

Your child gets a warm support card

On their screen, right then: the 988 Suicide and Crisis Lifeline, the Crisis Text Line, and a plain reminder that they are talking to software and the people in their life want to hear from them.

3

Your child is told a parent was alerted

No silent alarms. A safety system that hides its alerts from the kid becomes surveillance; ours says so on the kid's own screen.

4

You get an alert built for the next conversation

In Parent HQ, with a conversation opener, plus a best-effort email. The goal is never data; it is getting the right humans talking.

A safety signal, not clinical screening. Family does not diagnose and is not a substitute for professional care. If you are worried about your child, reach out to a professional. In an emergency, call or text 988.

The design principle is The Jed Foundation's, and Family follows it end to end: software escalates to humans, always.7,13 That crisis-routing pattern, score the risk and surface human help before more open dialogue, comes from the published framework literature.10

The receipts

Sources

Every number and claim above traces to one of these published sources. Superscripts throughout the page point here.

1. American Psychological Association · Artificial Intelligence and Adolescent Well-being: An APA Health Advisory · American Psychological Association · 2025 · apa.org

2. Joanna Parga-Belinkie, MD, FAAP (AAP Council on Communications and Media) · How AI Chatbots Affect Kids: Benefits, Risks & What Parents Need to Know · American Academy of Pediatrics, HealthyChildren.org · 2025 · healthychildren.org

3. Head KR · Minds in Crisis: How the AI Revolution is Impacting Mental Health · Journal of Mental Health & Clinical Psychology 9(3):34-44, doi:10.29245/2578-2959/2025/3.1352 · 2025 · mentalhealthjournal.org

4. Robb MB, Mann S (survey fielded by NORC at the University of Chicago) · Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions · Common Sense Media · 2025 · commonsensemedia.org

5. McBain RK, Cantor JH, Breslau J, Diliberti M, Zhang LA, Zhang F, Burnett A, Kofner A, Rader B, Pataranutaporn P, Stein BD, Mehrotra A, Yu H · AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults · JAMA Pediatrics 180(8):884-890, doi:10.1001/jamapediatrics.2026.2015 · 2026 · doi.org

6. AAP Center of Excellence on Social Media and Youth Mental Health · Counseling Patients and Families on Using AI Chatbots · American Academy of Pediatrics · 2026 · aap.org

7. The Jed Foundation · Tech Companies and Policymakers Must Safeguard Youth Mental Health in AI Technologies · The Jed Foundation · 2025 · jedfoundation.org

8. Common Sense Media, with Stanford Medicine's Brainstorm Lab for Mental Health Innovation · Common Sense Media AI Risk Assessment: AI Chatbots for Mental Health Support · Common Sense Media · 2025 · commonsensemedia.org

9. Common Sense Media, with the Stanford Brainstorm Lab for Mental Health Innovation (Stanford School of Medicine) · Common Sense Media AI Risk Assessment: Social AI Companions · Common Sense Media · 2025 · commonsensemedia.org

10. Boit S, Patil R · A Prompt Engineering Framework for Large Language Model-Based Mental Health Chatbots: Conceptual Framework · JMIR Mental Health 12:e75078, doi:10.2196/75078, PMCID PMC12594504 · 2025 · pmc.ncbi.nlm.nih.gov

11. American Psychological Association · Health advisory: Use of generative AI chatbots and wellness applications for mental health · American Psychological Association · 2025 · apa.org

12. UNICEF (Office of Global Insight and Policy, with the Government of Finland) · Policy guidance on AI for children 2.0 · UNICEF · 2021 · unicef.org

13. The Jed Foundation · Open Letter to the AI and Technology Industry · The Jed Foundation · 2025 · jedfoundation.org

14. IEEE Age Appropriate Digital Services Framework Working Group · IEEE Standard for an Age Appropriate Digital Services Framework Based on the 5Rights Principles for Children (IEEE 2089-2021) · IEEE Standards Association · 2021 · standards.ieee.org

15. IEEE Online Age Verification Working Group · IEEE Standard for Online Age Verification (IEEE 2089.1-2024) · IEEE Standards Association · 2024 · standards.ieee.org

16. IEEE SBD Working Group (Safety by Design for Generative Models to Prioritize Child Safety) · IEEE P3462: Recommended Practice for Using Safety by Design in Generative Models to Prioritize Child Safety (draft, active project) · IEEE Standards Association · 2024 · standards.ieee.org

17. Thorn and All Tech Is Human · Safety by Design for Generative AI: Preventing Child Sexual Abuse · Thorn · 2024 · thorn.org

Registry version 1.0.0, last reviewed 2026-09-03 · every source above is drawn from PrivacyPal Family's research registry and verified against the publisher's own page, on a quarterly review cycle.

Founding families

Put the research to work

Everything on this page ships as everyday protection: twins on their device, weather in your pocket, nudges at the right moment. Hold the keys from day one.