The wrong optimization target
An AI product can be extraordinarily good at making a child return. That does not make the return good for the child. The incentives that made adult feeds compulsive do not become benign because the voice is softer.
Little Pines starts from a refusal: the product never speaks first. Silence is a feature. Boredom is allowed to remain available. The object should support a child's existing world, not become the most responsive thing in it.
Privacy is architecture
Transcripts are the wrong primitive for early childhood. The system is designed to keep inference on-device and send patterns rather than conversations. That choice changes what can be measured, how support works, and which business models remain possible.
The product is screen-free and designed for ages three to seven. A working prototype is targeted for late 2026, with target retail in holiday 2027. Those are targets, not a disguised ship announcement.
What is being tested
The real question is whether a children's AI product can be economically viable without becoming an attention product or a surveillance product. The public benefit corporation structure is part of the answer because governance is product design when the user cannot read the terms.
The prototype has to prove the limits are felt in the interaction, not merely listed in a policy. If the boundaries only exist in a document for adults, they do not exist for the child.