KIDS implementation for AI services
A safe interface needs model evaluation, memory controls and operational safeguards as well as an age gate.
Which AI systems are in scope
The proposal defines AI companions and general conversational chatbots in Article 3. It excludes certain specialised task-limited chatbots from the latter definition. Educational or public-authority exceptions also have specific conditions; a marketing description alone does not establish an exemption.
Article 14 priorities
- Avoid relationship simulations likely to create emotional dependency and designs encouraging excessive use.
- Do not use prior interactions in subsequent interactions by default, subject to the article’s narrow safety/settings exception.
- Enable under-13 access only through guardian tools.
- Evaluate child risks before market placement or putting into service.
- Provide post-market monitoring under the article’s conditions; its small/micro-enterprise exception is specific, not a blanket exemption.
- For AI embedded in a platform or game, avoid automatic activation and promotion to minors, and offer an easy opt-out.
Connecting AI access and parental choices
A provider-integrated gateway could apply service-specific permissions, control tool access, keep conversation context transient and manage deliberately enabled memory. The provider would need to verify that its underlying model and suppliers honour the same settings.
Turning off memory in a user interface does not establish that a model supplier stops retaining prompts, logging requests or processing them for another purpose. Contracts, API settings and observed behaviour must agree.
Partnership
Discuss your integration. Explore the integration options with the Valvur team.
Valvur’s AI integration proposition
Combine age and guardian permissions with an access gateway, deliberate feature activation, purpose-limited context and provider-enforced memory settings. Valvur’s local-language safety work and parent guidance can support the experience, subject to evaluated quality and actual connector availability.
Keep three controls separate: the child’s ability to access the service, optional conversation memory, and the model supplier’s retention or training of prompts. A consent switch cannot settle all three. Valvur’s own AI Guardian also needs classification and dependency-risk assessment.