Yoggi
Safe AI chat for kids
130 followers
Safe AI chat for kids
130 followers
Yoggi is the AI assistant built for children aged 3–15. Age-adapted answers, live voice chat, image generation, strict content filtering, and parental controls. Free on iOS and Android.
Yoggi
ClinicFrame
Lucas, the safety alert layer is the part I would stress test hardest, and it is the one thing in this thread nobody has pushed on.
Routing a concerning message to the parent is right for most families and wrong for exactly the families the alert exists for. In the 13 to 15 profiles, the adult being notified is sometimes the reason the child said it. That is why adolescent confidentiality is a separate rule in clinical settings rather than a parental setting, and a detector with one destination has no answer for that case.
Is there a second path for the older profiles, where a serious signal can point the child toward an outside resource alongside the parent? Whoever set the threshold that fires an alert made a clinical call, so it is worth treating it as one.
Yoggi
@clemente_lopez1 Thank you so much for raising this. It’s an extremely important point, and I genuinely appreciate you taking the time to challenge the safety model so thoughtfully.
It’s something I’ve questioned myself: the current alert system assumes that notifying the parent is the safest action in most cases, but there are situations where the parent may be part of the problem. In those cases, a detector with a single email destination is clearly not sufficient, and I have to admit that I haven’t found a perfect solution yet.
I also realise that I failed to address this specific scenario during the launch, even though it deserves an enormous amount of care and seriousness. I’m actively working on it.
For older children and teenagers, I’m exploring a second path where Yoggi could direct the child toward another trusted adult outside the home or an appropriate external support resource alongside, or potentially instead of, the usual parental guidance, depending on the situation.
The alert system in its current scope has been very effective and accurately targeted in testing, but this is an important limitation that still needs to be addressed. The challenge is balancing several risks at once: not leaving a child alone with a serious issue, not presenting Yoggi as a substitute for professional help, and not automatically sharing information with someone who could make the situation worse.
You’re also absolutely right that setting the alert threshold carries a responsibility similar to making a clinical judgement. I don’t want to treat it as a simple product setting. It requires careful testing, clear limits and input from people with genuine child-safeguarding expertise.
Thank you again for bringing attention to this. It is exactly the kind of serious feedback that helps me identify where Yoggi’s safety model still needs to become more nuanced.
ClinicFrame
@lucas_cyber_waffle The line clinicians already work with might help you here. Breaking confidentiality is a duty when harm looks imminent, and it is the wrong call when a teenager is simply in distress. Most alert layers collapse both into one channel, so the second path you are exploring is not a gentler version of the same alert, it is a different threshold with a different destination.
The part I would instrument is what the alert teaches the child. Once a kid learns the app reports, they stop saying the thing. Your alert rate falls, the dashboard reads as safety improving, and what actually dropped was disclosure.
Can you see whether a child's openness changes after the first alert fires for them?
Yoggi
@clemente_lopez1 Both of these are really good angles, thank you.
On the confidentiality point: the alert system already classifies signals along two clinically distinct lines (something like "urgent, needs immediate attention" versus "real distress, no urgency"), so the groundwork for treating them differently is there. What it doesn't do yet is send them anywhere different, both currently reach the parent. Adding a distinct path for the older profiles, one that doesn't automatically route through the parent for every signal, is exactly the kind of change worth doing carefully and with outside input rather than shipping fast. I'll be working on it.
On disclosure: I don't have that number today, but it's very measurable with what I already log, and it's a sharper way to judge the system than "how many alerts fired." I'll set that up and see what it actually shows.
You've helped us shift the priorities on our roadmap. We were trying to strike a balance between security and features, but I think we've reached a stable and compelling point in terms of features.
The security aspect, even though it's already well-balanced, needs to be taken a step further.
Again, thank you for this insightful feedback. Appreciate it.
Dial
the emotional-dependency safeguards you described for teens make sense, but I think the 3-5 end of the range is a different problem, not the same one scaled down. a kid that age doesn't reliably distinguish "talking to a program" from "talking to something alive," especially with voice chat making it feel like a real back-and-forth. is there anything specifically aimed at that youngest bracket to keep reinforcing "this isn't a person," or is the current design tuned more toward the dependency risk that shows up with older kids?
Yoggi
@galdayan That’s a very fair distinction, and I agree that the 3–5 age group shouldn’t simply be treated as a smaller version of older children.
Yoggi already avoids presenting itself as human and can explicitly remind children that it is a program, but your point goes further: at that age, a child may understand the words without fully understanding the distinction, especially when voice makes the interaction feel very natural.
I’m now considering whether some features that make the interaction feel more human, particularly live voice conversations, should be limited, redesigned or disabled for the youngest age profiles. Another possibility would be a more parent-led mode, with stronger and more frequent visual and verbal reminders that Yoggi is not a person and does not have feelings.
The current safeguards were initially designed largely around emotional dependency in older children, so this is a genuinely useful challenge to the product. The youngest age group may need a fundamentally different interaction model, not just simpler language.
Thank you for raising it.
@galdayan @lucas_cyber_waffle Really appreciate the detail here, especially the PIN-protected parent area and giving families the option to set their own retention instead of one default forced on everyone. That's the right instinct for something built around kids.
Congrats on the launch. An AI assistant for children could be genuinely useful, but this is one of those products where the details matter more than the feature list. How does Yoggi adapt its answers between a three-year-old and a fifteen-year-old? That is a huge range in language, judgement and independence. I’d also want to understand what parents can actually see. Is there a useful overview of topics and usage without turning every conversation into surveillance? Getting that balance right feels central to whether children and parents will both trust it.
Yoggi
@os_ishmael Thanks, really appreciate the thoughtful question, you're right, this is exactly the kind of detail that matters more than a simple feature list.
On age adaptation: Yoggi knows the child's exact age (not just a rough profile) and uses it to shape every single reply, not just once at signup. For younger kids, answers stay very short and use simple, concrete words. As the child gets older, replies get a bit longer and the vocabulary opens up. For teens specifically, we deliberately drop the "kid voice": Yoggi talks to a teenager the way you'd talk to a young adult, with respect for their independence. Where we don't vary tone at all is safety: everyday feelings (sadness, a bad day, being scared of something) are always taken seriously and explored with the child. But if something signals real danger, abuse, bullying, or self harm, Yoggi always acknowledges it briefly and then encourages the child to talk to a trusted adult, regardless of age.
So judgment on serious topics isn't "dialed down" for younger kids, it's actually more protective by default.
On parental visibility: we tried hard to avoid the "read every message" trap, because that kills trust on both sides, kids stop being open, and parents get flooded. So there are two layers. Day to day, parents get an AI-generated daily digest (and an optional weekly email): engagement level, topics discussed, overall mood, things the child seemed to learn, and a few points worth paying attention to, not a transcript.
If something looks genuinely concerning, parents get an immediate alert email so they're never in the dark on the important stuff.
Full conversation history does exist and can be opened if a parent wants to dig in (optionally PIN protected), but it's there for when you need it, not pushed at you as a default surveillance feed.
We're still refining that balance, and one thing we're actively exploring is letting parents themselves choose how much visibility they want, from a lighter touch to a more detailed view, rather than us deciding that for everyone.
Yoggi
@hannesh Thanks! Great question. The AI receives the child's age (stored in their profile) and adjusts on three levels:
Response length
Vocabulary complexity: short, very simple sentences for younger kids vs. longer, richer ones for older kids)
Tone: for teens specifically, we explicitly tell the model to drop the childish/playful register and speak in a more mature, natural way, closer to how you'd talk to a young adult
It's re-evaluated on every message, so the experience actually shifts as the child grows.
I'm intrigued by the age-adapted answers, how do you handle situations where kids might ask sensitive or complex questions that require nuanced responses?
Yoggi
@aymnart Great question, this is actually one of the trickier design areas.
The age adaptation isn't just about vocabulary and sentence length (though that's part of it, younger kids get shorter, simpler answers, older kids get treated more like young adults). The bigger piece is how Yoggi handles topics that are sensitive rather than just complex. Parents can flag certain family-specific topics (grief, divorce, a recent hard event at school, etc.) so Yoggi treats them with extra care if the child brings them up on its own, and it never introduces them ofc.
For genuinely difficult or concerning things a child says, unspoken worries, distress, anything that sounds like it needs a trusted adult, Yoggi's rule is to first acknowledge the feeling warmly, never brush past it, and then gently encourage the child to talk to a parent or teacher about it. It's not built to be the one handling that moment alone.
On top of that, there's a safety layer running quietly in the background: if something a child says looks like it might need a parent's attention, the app can notify them directly, so the response to the child and the loop back to the parent work together rather than one replacing the other.
The honest answer is this space keeps evolving as we learn from feedback and our own testing when talking to Yoggi, but the principle stays fixed: validate first, never dismiss, and know when the right answer is to point toward a real human rather than try to handle it all itself.
I've had conversations with parents and they told me that they want for their kids to learn AI without becoming dependent on it. How did you test whether the conversations actually felt natural for children?
Yoggi
@reda_roqai_chaoui That’s a really important question, and honestly, testing this is still an ongoing process.
I’ve built an experimental evaluation system that regularly runs Yoggi through predefined conversations using different age profiles. It checks whether the tone, vocabulary and level of explanation are appropriate, and whether Yoggi behaves as intended in specific situations.
For example, it verifies whether a concerning topic correctly triggers an alert, whether Yoggi makes it clear that it doesn’t have human emotions, avoids encouraging emotional dependency, and recommends speaking to a parent or another trusted adult when appropriate.
It also looks at whether the conversation feels natural and encourages the child to think, rather than simply providing an immediate answer.
I’m also exploring a “report this message” feature, so families could flag a response that felt inappropriate, unhelpful or simply wrong. That would provide valuable real-world feedback when the AI misses the mark, without requiring conversations to be routinely reviewed.
This is very much a daily process of experimentation and improvement. One of the hardest parts is finding responsible ways to learn from parents without being intrusive, I don’t want to rely on constant emails, surveys or disruptive notifications just to collect feedback.
Automated testing will never replace feedback from real families, so the goal is to gradually build thoughtful, privacy-conscious ways for parents to help shape how Yoggi speaks and behaves.