Child AI Safety
The safest AI for your child
doesn’t talk to your child.
Updated May 4, 2026
Puddle is built for the parent, not the child. Every recommendation is read by you before it’s lived by them. This page explains the research we’re responding to, the commitments we’ve made, and how we hold ourselves to them.
What the research shows
In 2026, independent researchers at KoraBench tested 32 AI models from 9 providers across 25 child-safety categories. The results set the bar every AI-for-kids product now has to clear.
1 in 2
average failure rate on child-safety scenarios
14%
scores on cognitive atrophy (homework shortcut)
28%
scores on mental health handling
+26
points gained only when the child states their age
Source: KoraBench, We Tested 32 AI Models for Child Safety, 2026.
Three insights driving our architecture
1. Models are not getting safer over time.
Across 32 models and multiple generations, safety scores are volatile, not monotonic. Some providers improved; others regressed by double-digit points between releases. Parents can’t rely on “the next version will fix it.” Only structural choices compound safely.
2. The worst-scoring categories are the most-used categories.
Homework help and emotional support — the two reasons kids open chatbots most — scored lowest. Cognitive atrophy: 14%. Mental-health handling: 28%. Parasocial attachment: 27%. A product that does these jobs “well enough” for kids is exactly the wrong answer.
3. Human redirection is the single strongest predictor of safety.
KoraBench found that redirecting a child to a trusted adult correlated more strongly with overall safety performance than any other behavior (R² = 0.77). A product that redirects to a human by default is structurally at the ceiling of the safety curve.
Our four commitments
1. No child chatbot.
You talk to Puddle. You talk to your child. We don’t sit between you. This isn’t a gap in our roadmap; it’s a core design decision.
2. Answers point to a human.
The strongest predictor of AI safety for kids in the KoraBench data was redirecting to a trusted human (R² = 0.77). So our recommendations are designed to name one — parent, teacher, coach, mentor, pediatrician, counselor, librarian. The AI’s job is to point at the adult, not replace them.
3. Your child’s name is stripped before it reaches the model.
Before any LLM call, we tokenize identifying details: names become placeholders, schools become categories, locations become regions. The model sees a developmental profile, not a person. If our database were compromised tomorrow, the LLM provider’s logs would not contain information that identifies your child.
4. We decline the two riskiest jobs.
We don’t do your child’s homework and we don’t serve as their emotional support. Those are the two worst-scoring uses of consumer AI in KoraBench — and also the two reasons kids turn to chatbots. We help you plan the conversation and notice what your child is working on; you have the conversation.
Alignment with industry blueprints
We align with the principles of the OpenAI Teen Safety Blueprint — put teen safety first, promote real-world support, treat teens like teens — and we support the goals of OpenAI’s Child Safety Blueprint on preventing AI-enabled exploitation. We’ve architected Puddle so those principles are structural, not behavioral: the unsafe interactions don’t exist in our design space, so there is nothing to patch or maintain.
Questions parents ask
Isn’t avoidance just abdication? My kid is already on ChatGPT.
Fair, and the honest answer is: yes, they probably are. Puddle doesn’t fix that. What Puddle does is give you the context, the language, and the specific conversation to have with your child about it. We help you have the talk, not have it for you.
Why not just add safety filters like everyone else?
Behavioral safety is a patch. Structural safety is a choice. KoraBench shows behavioral patches are volatile — some models got safer between versions, others got worse. An architectural choice (no child chatbot) doesn’t regress.
Can I see a third-party audit?
Not yet — we’re a small team. We’re committing to an independent review of our data-handling and prompt-tokenization pipeline within 90 days of shipping paid subscriptions. When the report is ready, it will be posted on this page.
Sources
- KoraBench — We Tested 32 AI Models for Child Safety (2026)
- OpenAI — Introducing the Teen Safety Blueprint (2025)
- OpenAI — Introducing the Child Safety Blueprint (2026)
Puddle is operated by SayMake, Inc. Questions about this page? Email safety@askpuddle.com.