AYO Opens a Privacy-Preserving Platform for Studying Children’s LLM Interactions

Summary:AYO (AI Youth Open Science) is an open-source research platform for studying how children and adolescents interact with large language models. It combines a child-friendly interface, configurable age-based safety guardrails, researcher monitoring, and privacy-preserving data workflows, addressing major obstacles in minors’ AI research: safety, consent, control, and reproducible experimental design. The platform supports cross-sectional, longitudinal, and experimental studies of trust, epistemic

---

AYO (AI Youth Open Science) turns research on children’s interactions with large language models into an open-source, supervisable experimental platform. It does not introduce a new foundation model or training algorithm. Instead, it addresses infrastructure problems that are often overlooked in minors’ AI research: age-appropriate interaction, configurable safety guardrails, researcher monitoring, consent, data minimization, and reproducible experimental control.

The Unit of Study Is Interaction, Not Just the Model

Much existing work relies on surveys asking children whether they trust AI, understand model responses, or enjoy using chatbots. Surveys are inexpensive, but they cannot reliably capture immediate behavior during real interactions: how children formulate questions, when they accept or challenge model outputs, whether they continue after an incorrect answer, and how interfaces or guardrails alter exploratory behavior.

AYO shifts the unit of analysis toward the child–LLM interaction process. The platform is designed to support cross-sectional, longitudinal, and experimental studies of trust, epistemic judgment, curiosity, and related developmental outcomes. This is important because liking a chatbot does not establish calibrated understanding, while a one-time trust score cannot substitute for observations of questioning, rejection, deception, and dependence.

Platform Architecture

According to the public abstract, AYO combines four main capabilities.

1. Child-Friendly Interface

A children’s interface must control reading difficulty, interaction burden, and prompt structure. Age adaptation is not merely a matter of shortening text. It should also account for anthropomorphism, dependency-inducing language, and the risk of presenting model outputs as authoritative. Ideally, cognitive load, accidental interaction, and explanatory feedback should be treated as experimental variables rather than fixed interface properties.

2. Configurable Age-Based Guardrails

Guardrails can operate at the input, output, and orchestration layers:

• Input filters identify requests involving self-harm, exploitation, or age-inappropriate content.
• Output filters block unsafe or unsuitable responses and provide age-appropriate fallbacks.
• Orchestration controls session length, tool access, memory writes, and exposure of high-risk functions.

Age-specific policies are more realistic than one global filter, but they introduce classification errors at age boundaries. An LLM-based age or risk classifier can itself be vulnerable to prompt injection, role-play, and long-context evasion. Critical controls should therefore combine deterministic state machines, policy rules, output classifiers, and human escalation rather than relying entirely on another generative model.

3. Researcher Supervision and Auditing

Supervision may include session-status visibility, risk alerts, anomaly flags, intervention or pause controls, and researcher-operation logs. Yet visibility should not imply unrestricted access. Supervisor permissions should be minimized and role-based, with access times, actors, and data scopes logged so that research does not become a new source of continuous surveillance.

If complete conversations can be exported, teams should distinguish research data, operational logs, and debugging telemetry. Provider-side logs, crash reports, vector-store replicas, and annotation interfaces can each create additional copies of sensitive information.

4. Privacy-Preserving Data Workflows

Privacy protection requires more than removing names. A stronger workflow would:\

• Avoid logging raw prompts and complete responses by default.
• Perform entity recognition and pseudonymization before the data reaches analysis layers.
• Separate identity records from conversational content.
• Apply access and publication thresholds to small demographic groups.
• Restrict long-term memory and cross-session linkage.
• Audit data access, export, and deletion by the research team.

Differential privacy can reduce the risk of inferring sensitive attributes from aggregate results, but it weakens statistical power for small age groups, rare behaviors, and longitudinal change analysis. A more defensible approach is to minimize collection first and then apply controlled privacy budgets only to aggregates that genuinely require disclosure.

Engineering Trade-offs

Child–AI studies contain an inherent conflict. Stricter guardrails reduce risk but move the interaction farther from natural product use. Permissive guardrails preserve ecological validity but increase exposure to harmful content.

Configurable guardrails make this trade-off explicit. Yet if researchers frequently intervene by blocking or rewriting responses, model effects become inseparable from the moderator’s judgment. A stronger design records every intervention, substitution, and escalation event, while versioning guardrails alongside model and system-prompt versions.

Longitudinal studies face another issue: attrition bias. Children may stop using the system as novelty fades, and dropouts may differ systematically from continuing users. Treating only completed sessions as the analysis population can exaggerate apparent trust formation and sustained engagement. Withdrawal, refusal, and mid-study exit should therefore be recorded as outcomes rather than discarded as operational noise.

How to Evaluate AYO

The abstract does not report quantitative results. For a research platform, standard model leaderboard scores are less important than validating the full research chain.

1. Safety block and false-block rates: Report them by age, language, prompt technique, and risk category rather than relying only on a global average.
2. Prompt-injection robustness: Test role-play, encoding, nested-story, and delayed multi-turn attacks.
3. Reproducibility: Pin model version, system prompt, guardrail configuration, interface version, and sampling parameters in an auditable manifest.
4. Construct validity of developmental measures: Trust, epistemic judgment, and curiosity require measures validated with children, not simple proxies such as response length or conversation count.
5. Privacy leakage audits: Inspect raw text, exception logs, analysis exports, and model-service traces for residual identifiers.

Guardrails can also change the behavior researchers are trying to measure. If risky requests always trigger a fixed refusal, children may learn to hide them rather than develop safer judgment. Evaluation should test not only whether a request is blocked, but whether the system helps children understand why it cannot answer and where to seek help.

Practical Deployment Guidance

AYO is best treated as a research environment, not as a ready-to-deploy general-purpose chatbot for children. Teams should begin with the research question, then select models and guardrail configurations rather than assuming that a black-box chat frontend is sufficient.

A minimum deployment should include:

• A separate safety orchestration layer between the interface and model service, with versioned policies.
• Separation of raw conversations from derived features, favoring structured events over complete transcripts.
• Executable guardian-consent, withdrawal, and deletion workflows.
• Role-based access control and access logging for researcher interfaces.
• Red-team testing, privacy-impact assessment, and small-scale usability testing before formal studies.
• Explicit tracking of provider, guardrail, and interface updates as experimental covariates.

Value and Limitations

AYO’s principal contribution is moving children–LLM research from questionnaire-based observation toward controlled, longitudinal, and auditable platform studies. It could provide a stronger evidence base for age-appropriate AI design, education, and child-safety policy.

The current abstract is not enough to establish that the hardest questions have been solved. External validity depends on whether guardrails remain safe without distorting behavior, whether privacy claims withstand independent audit, and whether the platform covers different languages, cultures, and vulnerable groups. AYO is better understood as a reusable research starting point than as a final answer about how AI affects children in the wild.

---

Source: ArXiv AI (cs.AI) (2026-11-10)

— END —

Tags: #AYO Follow #Child AI Safety Follow #LLM Interaction Research Follow #Privacy-Preserving Research Follow #Open-Source Platform Follow