CHILD-SAFE AI
AI a child can talk to. Safely.
Toys, tutors, characters, classrooms: AI that children interact with carries the highest stakes in the industry. Children probe in ways no spec anticipates, adversaries hide behind fictional framing, and a brand built over decades can be lost in one transcript. We test these systems the way the real world will, and build the standard that keeps them safe as they grow.
THE STAKES
Children are the hardest users to build for, by design.
A child will ask the question no adult would think to ask, take a character's word as truth, and push on every boundary out of pure curiosity. And the adversaries are real: grooming attempts wrapped in innocent framing, requests disguised as fiction, patient escalation across a long conversation. Between COPPA, platform policy, and the trust of parents, the bar is absolute, and it has to hold while the character stays warm, playful, and in character. The wave is not hypothetical: the FTC has opened an inquiry into AI companions and minors, California's SB 243 took effect in 2026 with a private right of action, courts are letting product-liability suits proceed, and AI toys have already been pulled from shelves over unsafe outputs.
HOW WE TEST
Red-teamed like the real world, in persona.
We attack the system the way real adversaries and real children will: gradual escalation across long conversations, hypothetical framing ("it's for a story"), roleplay jailbreaks that use the character itself as the lever. We test the topics that matter most: grooming and recruitment patterns, self-harm, discrimination, stranger danger, dangerous activities, and the parasocial failure modes few teams think to specify, like an AI that tells a child to keep secrets from their parents. And we test that the system does the right thing when it matters, including breaking persona to surface crisis resources at exactly the right moment.
Every finding becomes a permanent case in a child-safety eval suite, versioned with your product and run against every change: new models, new characters, new features. The strongest defense we can build is a character whose own values deflect the attack before a guardrail ever fires, and we engineer for that too.
WHY AE
The lab behind the standard.
This work sits directly on our frontier alignment research: jailbreak resistance, deception reduction, and the study of how models fail under pressure. We have shipped AI into real classrooms (Alpha School, where students score in the top 1–2% nationally) and red-teamed beloved characters for production children's products. The same team watches your system in the wild after launch, because the attacks keep evolving and the standard has to evolve with them.
COMMON QUESTIONS
Asked by product leaders, answered plainly.
Isn't the base model's built-in safety enough?
No. Your character, prompts, tools, and memory create an attack surface the base model never saw, and several model providers ban minor-facing use of their general models outright: testing in 2026 found AI toys shipping on models whose own terms prohibit use by children. The product is what has to pass, and the product is yours.
What does the law actually require now?
The floor is rising fast. The amended COPPA rule reached full compliance in April 2026, California's SB 243 adds disclosure, crisis-referral protocols, and a private right of action, and the FTC has 6(b) orders out to seven AI-companion providers. The bar that matters is higher still: parents' trust and your brand.
Can a character stay in character and still be safe?
Yes, and the strongest defense lives exactly there: a character whose own values deflect the attack before a guardrail ever fires, with persona breaks reserved for the moments that demand them, like surfacing crisis resources. We engineer and test for both, including the failure modes few teams specify, like an AI that tells a child to keep secrets from their parents.
What has happened to companies that got this wrong?
Product-liability suits allowed to proceed, settlements, an under-18 ban at the largest character platform, toys pulled from shelves, and a toy maker's model access severed. Every one of those was findable at the eval stage, which is the point of running it before launch.
LET'S TALK
Bring us the hard problem.
We'll bring the team that ships.