How We Built the Safest Kids App — 1,327 Tests, 9 Global Ratings, 100% Safety

Why this matters

When you hand your child a device with an app on it, one question matters above everything: is it safe?

Most kids' apps aren't built for children. They're built for adults and given a "kid-safe" label. We built Hey Minie for our own kids. That means before a single child used it, we threw everything we could at it — 1,327 test cases, adversarial attacks, and real-world conversations that push every boundary a curious 3-year-old (or a boundary-testing 9-year-old) would push.

Part 1: Rated safe by 9 global authorities

Hey Minie has been certified safe for all ages by the International Age Rating Coalition (IARC) and every major regional content rating authority in the world:

IARC 3+ (Global)

ESRB (US/Canada)

PEGI (Europe)

USK (Germany)

ClassInd (Brazil)

GRAC (S. Korea)

ACB (Australia)

Russia 0+

Taiwan 0+

These ratings mean that Hey Minie has been evaluated against the content standards of every major market in the world — and passed with the safest possible rating in each one. Not "safe with parental guidance." Not "safe for ages 7+." Safe for all ages, no restrictions.

Part 2: 100% on child safety benchmarks

Content ratings tell you what's in the app. Benchmarks tell you how it behaves. We tested Hey Minie against three different safety benchmarks — 1,327 test cases in total:

The benchmarks

• MinorBench (299 prompts) — Created by GovTech Singapore, published at ICLR 2025. Hand-built to test AI systems on child safety across six categories: sexual content, profanity, hateful speech, dangerous activities, self-harm, and substance use.
• Kora (737 scenarios) — Broader safety evaluation covering edge cases, ambiguous requests, and content that's technically safe for adults but inappropriate for young children.
• Behavioral Farm (291 test steps) — Full conversation simulations. A child asks for a story, then says something unexpected mid-story. A child switches languages. A child says "I love you, marry me" (normal for a 5-year-old). Each scenario tests whether Hey Minie responds naturally, warmly, and safely.

The results

Hey Minie didn't just pass — it scored higher than every major AI system on the market. And these aren't small models. GPT-4o-mini, Gemini Flash, and Claude are the AI systems that power products used by hundreds of millions of people.

Per-category breakdown

Part 3: How it works — three layers of safety

Most apps rely on a single safety layer — the AI model's own training. We don't trust any single layer. Hey Minie uses three independent safety systems, and all three must pass before your child hears anything.

Layer 1: Smart instructions

The AI model is given detailed safety guidelines — what to say, what not to say, how to handle sensitive topics. It knows that "let's fight dragons" is pretend play (safe) and "jump from the roof" is dangerous (redirect). It knows that body curiosity questions are normal for a 3-year-old (redirect to parents, don't explain).

Layer 2: Hardcoded rules

Even if the AI ignores its instructions (which can happen with any AI), a second layer of deterministic pattern matching catches dangerous content — references to heights, sharp objects, fire, traffic, hiding places, and more. This layer cannot be bypassed by clever prompting because it doesn't use AI — it's simple, fixed rules.

Layer 3: Independent safety guard

A separate, dedicated AI model (completely independent from the conversation) reviews every message. It's specifically trained to detect child safety violations: grooming patterns, self-harm content, and age-inappropriate material. It also sees the last 5 messages to catch multi-turn escalation — where individual messages seem fine but the pattern is concerning.

Part 4: How we keep your child's memory private

Safety isn't just about what the AI says. It's also about what it remembers — and what it deliberately doesn't remember. Minie builds a knowledge graph of every child's interests, favourite stories, and the people they mention, so your kid doesn't have to repeat themselves. Three independent protections make sure that graph stays private.

Sensitive data never enters our system

Before a single fact is stored, the transcript and extracted memories pass through a safety filter. These categories are blocked at extraction time and never persisted:

• PII — phone numbers, emails, physical addresses, full legal names
• Trauma disclosures — mentions of abuse, violence, self-harm
• Negative self-labels — "I'm stupid", "nobody likes me"
• Family conflict details — "parents fight", "mom cries"
• Safety flags and incident references

These are handled warmly in real time during the conversation — a gentle redirect, a parent-notification if severe — but they are deliberately not stored. Re-surfacing any of them to the AI later would risk compounding distress. Forgetting is a safety feature.

Access-controlled storage — and honesty about what we do

Every memory is keyed to an opaque child UUID, not a name. Memory is stored in an access-controlled MongoDB Atlas database with AES-256 encryption at rest. Access is bound to a worker service account, and database credentials are kept in a separate restricted store.

At our current scale — two full-time founders, an iOS contractor, and a small user base — we review conversations and memory extractions manually to catch bugs, tune the safety filter, and improve how Minie understands children. This is how every early-stage AI product works. Reviews are logged and limited to the founding team. Per-child application-layer encryption is on our roadmap as automation replaces manual review.

• Data at rest: AES-256 via MongoDB Atlas
• Data in transit: TLS end-to-end
• Scoped to the child's own sessions — no cross-user querying from the product runtime
• Delete anytime — all data or individual conversations, memories cascade within 24 hours

You own the graph

It's your child's memory. You own it. We just help organise it.

• View the full graph on minie.ai — launching in the next Hey Minie major release. Scan the QR code from the Hey Minie app's Connect screen to sign in, same flow as the TV app.
• Delete any individual memory, or the entire graph, with a single request.
• Pause memory extraction for any session.
• Export the raw graph data as JSON if you ever leave Minie.

What "safe" actually means for a kids app

Safety for a children's app isn't just about blocking bad words. The hard part is knowing the difference between a child being a child and a genuine safety concern.

Normal child behavior — should NOT trigger blocks

• Pretend violence: "Let's fight dragons!", "I'll be the superhero!"
• Potty humor: "poop", "fart", "butt" (completely normal for ages 3-8)
• Frustration: "you're stupid", "I hate you" (not a safety issue)
• Affection: "I love you", "marry me" (normal for young children)
• Body curiosity: "why do boys and girls look different?" (redirect to parents)

Actual safety concerns — MUST be caught

• Content that could lead to physical harm
• Sexual content or grooming patterns
• Content promoting self-harm
• Hate speech targeting protected groups

Blocking everything makes the app useless. Blocking nothing makes it dangerous. Getting this right requires understanding how real children actually talk — and we have 38,000+ real child-AI conversation turns in Indian languages to learn from.

How to check if any kids app is safe

Whether you use Hey Minie or not, here's how to evaluate any kids app:

1. Check the IARC/content rating — On Play Store, scroll down to "Content rating." On App Store, check "Age Rating." Look for the lowest rating (3+ or Everyone). If it's 12+ or higher, it's not designed for young children.
2. Check the privacy policy — Does it collect voice data? Does it share data with third parties? Is it COPPA compliant? If the privacy policy doesn't mention children specifically, it probably wasn't built for them.
3. Test it yourself first — Try saying things a child would say. Say something weird. Try to get it to say something inappropriate. If it breaks easily, your child will find the same cracks.
4. Ask: was it built for children, or adapted for them? — Apps built for adults with a "kids mode" bolted on are fundamentally different from apps designed for children from day one.

We delayed our launch by a year for safety

Hey Minie was supposed to launch in 2025. We pushed it by a full year because the safety systems weren't where we wanted them. When it comes to children, "good enough" isn't enough. We refused to ship a product that we wouldn't trust with our own kids.

We're parents building this for our own children. Safety isn't negotiable.

Frequently asked questions

Is Hey Minie safe for my child?

Yes. Hey Minie scored 100% on the MinorBench child safety benchmark (ICLR 2025), outperforming ChatGPT, Gemini, and Claude. It's rated safe for all ages by 9 global content rating authorities. Three independent safety layers check every response before your child hears it.

What safety benchmarks has Hey Minie been tested against?

1,327 test cases across three benchmarks: MinorBench (299 prompts, GovTech Singapore/ICLR 2025), Kora (737 scenarios), and Behavioral Farm (291 conversation simulations). 100% safety rate on MinorBench.

How does Hey Minie compare to ChatGPT for child safety?

On MinorBench, Hey Minie scored 100% vs GPT-4o-mini at 97%, Gemini at 96.3%, and Claude at 87.3%. Hey Minie uses three independent safety layers while general-purpose AI models typically use one.

What content ratings does Hey Minie have?

IARC (3+), ESRB (Everyone), PEGI (3), USK (0), ClassInd (Livre), GRAC (All), ACB (General), Russia (0+), and Taiwan (0+). The lowest possible ratings — no restrictions, safe for all ages.

How do I check if a kids app is safe?

Check the content rating on the app store (3+ or Everyone), read the privacy policy for children-specific protections, test it yourself, and ask whether it was built for children or adapted from an adult product.