A research platform for studying how children interact with AI

Understanding how children use AI requires more than asking whether they have tried a chatbot. Researchers also need to observe what children ask, how they respond to answers and whether their judgments change over repeated conversations. Yet allowing open-ended exchanges means researchers cannot specify every response a child might encounter in advance.

Huiwen Alex Yang, Rachel Skwersky, Bill D. Thompson and Celeste Kidd at the University of California, Berkeley describe an approach to this problem called AYO, short for AI Youth Open Science. Their November 2026 arXiv preprint, titled “AYO: A Safe and Privacy-Preserving Platform for Studying Children's Interactions with LLMs” and labeled for the NeurIPS 2026 Child Safety in AI workshop, presents research infrastructure rather than results showing how AI affects children’s development.

A chatbot interface with research controls

AYO has two main components: a participant-facing chat interface and a researcher dashboard. It builds on LibreChat, which supplies a configurable conversation interface, and LiteLLM, which manages requests to language-model providers.

The child-facing interface includes text-to-speech and file uploads, while exposing fewer menu options than a fully configurable system. Participants can revisit their chat history, but the authors say no memory is preserved across separate chats. Those are different features: a saved conversation can remain accessible without the model carrying information from it into another conversation.

Researchers can invite families through unique sign-up links, organize participants into groups and configure experimental conditions. The dashboard allows them to inspect conversations and export records with associated metadata. Models can run on researcher-controlled infrastructure, or researchers can use commercial models under appropriate research-specific agreements.

This setup could support comparisons across age groups, repeated observations over time and experiments that change aspects of the interaction. Questionnaires delivered immediately after a conversation could help connect what a child did with what they thought about the exchange.

Safety combines filters and people

The authors report evaluating 22 child-safety keyword-blocking guardrails from LiteLLM. Their testing included prompts intended to be allowed or blocked, attempts to bypass restrictions and edge cases. They say gaps identified during testing led them to add a semantic guardrail: an evaluator prompt intended to assess meaning and context rather than keywords alone.

Its categories include self-harm intent, romantic content or persona adoption, stereotyping, graphic violence and sexual content. Depending on the conversational context, a triggered safeguard can produce a redirection message or a hard block, with an alert recorded for researchers. Guardrails can also be configured by age group.

Human oversight is another layer. Trained research staff manually review platform activity and maintain a log of flagged incidents. A tiered crisis protocol specifies responses that can include monitoring, guardian notification, suspended access and legally required reporting. High-severity incidents call for phone and email outreach, alongside crisis resources for families.

These are described procedures, not proof that harmful exchanges will always be detected. The paper does not provide quantitative detection results, rates of inappropriate blocking or a measured comparison demonstrating the added guardrail’s effectiveness.

Privacy does not mean conversations disappear

The platform separates identifying information, including consent materials, from conversation data. The authors also describe a guardrail intended to block requests containing personally identifiable information before they reach a model provider. Its reliability is not quantified in the paper.

For participants under 13, the team proposes video calls to verify the parent–child relationship and obtain parental consent. Consent documentation would be retained, but video recordings would not. The authors say their ongoing studies will limit data collection to six weeks per participant.

Although the paper invokes a zero-data-retention policy, it also explicitly describes researcher storage and export of conversations. That phrase should therefore not be read as a promise that no records exist anywhere. Provider arrangements and researcher-held records remain important distinctions.

Infrastructure, not developmental findings

AYO could help researchers investigate trust, curiosity and how children decide whether information deserves belief. But this paper does not establish that the platform improves those outcomes, prevents harm or reproduces children’s ordinary chatbot use outside a study.

Its contribution is a configurable environment for asking those questions while incorporating safeguards and oversight. Whether those protections work sufficiently well—and what children’s conversations reveal about development—requires evidence beyond the platform description.