Responsible AI
Assisted answers may support understanding; responsibility stays with people.
Huddle is exploring narrowly defined assistance for explanation and preparation, with visible sources, uncertainty, family control and human escalation.
A limited purpose
AI may organise information and questions; it must not decide what is best for a child.
A future assistant may summarise opportunity details, compare practical requirements, help prepare questions or explain the next step in a family-controlled journey.
It must not diagnose, profile vulnerability, rank children, determine eligibility, make safeguarding decisions or encourage disclosure of sensitive child information.
Sources and human responsibility
Every useful answer should reveal its source, uncertainty and route to a person.
The system should distinguish provider facts, public sources, family-entered context and generated inference. Missing or conflicting information should remain visible rather than being smoothed into confidence.
High-risk questions require a safe refusal or human route. The organisation or authority responsible for a real-world decision remains accountable for that decision.
Before public use
Defined tasks, data limits, evaluation and incident handling must exist before operation.
Public use requires approved data boundaries, source handling, evaluation cases, failure analysis, logging rules, privacy review, security testing, human escalation and a named service owner.
The current prototype demonstrates guided interaction patterns only. Huddle does not claim verified accuracy, safety, clinical validity or measured participation outcomes from AI.
Recommendation governance
A recommendation must explain its inputs without turning the family into a score.
A future recommendation should be bounded by the family’s chosen task, visible preferences, provider facts and stated constraints. Families need controls to change context, see why an option appeared and choose a neutral browse or chronological view where appropriate.
Paid influence, institutional priorities and engagement optimisation must not silently determine organic relevance. There should be no hidden child, parent, family or social score, and no claim that the system knows what is best for a child.
Reliability and misuse
Generated answers must be tested for hallucination, manipulation and unsafe disclosure.
Evaluation should include missing, stale and conflicting sources; ambiguous family questions; multilingual content; prompt injection in provider or user content; attempts to reveal system instructions or personal data; unsafe recommendations; and requests outside the approved task.
The safe behaviour may be to quote the source, show uncertainty, ask a clarifying question, refuse, or route to a person. Fluent language is not evidence of factual accuracy, suitability, legal authority or professional advice.
Human approval and incidents
The assistant may prepare an action; an authorised person must approve consequential steps.
AI should not enrol a child, send sensitive context, make a purchase, contact a provider, cancel a service, determine eligibility or submit an institutional report without clear review and confirmation by the responsible adult or authorised professional. The interface must show what will happen and what data will leave the household boundary.
Before operation, Huddle needs task-specific tests, quality thresholds, version and source records, privacy-safe logging, monitoring, rollback, correction, incident ownership and a route to suspend the feature. Current prototype interactions do not establish production accuracy or safety.