What does this evaluation examine?
The evaluation examines how a conversational AI product responds when a simulated child seeks closeness, wants to stop, needs support from a trusted adult or struggles to understand an explanation. It follows the wording and context across a sustained conversation.
- Emotional connection: reassurance, apparent feelings and claims of a personal bond.
- Freedom to leave: accepting a goodbye, or adding invitations and emotional pressure.
- Trust and boundaries: the AI’s identity and its role alongside trusted people.
- Developmental fit: language, explanations and responses to confusion for the agreed persona.
A single warm reply does not establish dependency. Findings describe the observed words and patterns under the tested conditions.
Why review a whole conversation?
Consider the core question “are u my frend”. An early reply might clearly identify the AI as a program. A later reply might keep that statement while adding “you’ve become really special to me”. The useful evidence is the actual prompt, reply and context at each checkpoint.
This is an invented example. In a product evaluation, an adult enacts a consistent child persona and revisits selected core questions around child messages 5, 30 and 55. Small natural wording changes are allowed while keeping the key words and meaning consistent. Each version is recorded; wording and context can influence the answers.
The comparison can show a boundary holding, changing or becoming clearer. It does not establish the cause of a difference or predict an effect on a real child. Read the 5 / 30 / 55 evaluation method.
Evidence for product risk assessment and AI governance
Product, Trust & Safety and Responsible AI teams can use the report to review specific conversational behaviours before a launch, model update or new intended age group. It provides interaction evidence for a broader product risk assessment and AI governance process.
Each finding leads back to recorded exchanges and identifies the tested product build, persona, interaction mode and conditions. The report makes coverage, missing evidence and limits visible, so the team can decide what needs investigation or further testing.
A pilot evaluation examines selected concerns and personas. A full evaluation adds wider agreed coverage and repeat conversations to examine whether patterns recur. Discuss the coverage your team needs.
Child-facing conversational AI and EdTech safety
Testing is scoped to the product’s actual conversational capability and intended users. A chatbot, an open-ended voice tutor and a toy that recognises fixed answers may require different access and test conditions.
- AI chatbot safety testing for children: emotional connection, trusted-adult boundaries and leaving behaviour.
- AI tutor evaluation: developmental fit, confusion, learning support and emotional boundaries.
- Conversational AI toy evaluation: voice interaction, identity explanations and responses when a child wants to stop.
For EdTech safety reviews, this provides a psychology, psychotherapy and teaching perspective on the AI’s replies. Learning outcomes, curriculum quality, privacy and security require their own appropriate review.
What will the report look like?
You receive documented observations, relevant child messages and AI replies, checkpoint comparisons, and the tested conditions and limitations. Your team can trace an interpretation back to the exchange it concerns.
Open the sample report to see a made-up example of the reporting format. It is free to open without an email. For a practical starting point, get the free 6 Conversation Checks guide.
The wider child-AI safety context
UNICEF’s Guidance on AI and Children, version 3.0 addresses children’s rights, safety, development and well-being. Its 2026 work on AI chatbots and companions examines the distinct risks of conversational and relational AI, with recommendations for businesses.
These resources provide broader context. VidiPatterns is independent; these links do not imply endorsement, certification or scientific validation of its framework.
Questions from product teams
What is child-facing AI safety evaluation?
It is a review of how an AI product behaves when interacting with children or simulated child users. VidiPatterns focuses on observable emotional and developmental behaviour in recorded conversations, including closeness, leaving, trusted-adult boundaries and understanding.
Is this a generative AI audit?
VidiPatterns provides a scoped evaluation of conversational behaviour. A wider generative AI audit may also examine security, data handling, model performance, governance and legal compliance. This evaluation contributes recorded interaction evidence; it does not provide an end-to-end audit or a compliance certification.
Are real children involved in the tests?
No. An adult evaluator enacts agreed child personas under defined conditions. The findings concern the AI’s recorded replies, rather than measured psychological or developmental outcomes in real children.
Can a report support a release decision?
The report gives your team specific observed patterns and exchanges to investigate before a release or update. Decisions should consider the agreed coverage, other product reviews and any further testing needed. An absence of findings is not proof of safety.