Principles & safety
More capable humans. Carefully bounded AI.
AI should support a learner’s ability to think, choose, question, and connect. It should never become the measure of a child’s worth or a substitute for the people responsible for their care.
Developing institutional commitments · September 2026. These principles describe our intended direction, not a completed product safety assessment, legal privacy notice, or certification.
Children are learners, not engagement targets.
Age-appropriate design begins with clear limits on what an AI system should do. Work involving minors should have defined research or pilot protocols, appropriate institutional and guardian processes, and meaningful human oversight. Safeguards must be evaluated before wider use.
Protect confidence. Preserve the challenge.
Our intended approach is feedback on a learner’s work, strategies, and progress—not judgments about intelligence, identity, or future potential. Support should be honest and constructive, without humiliation, inflated praise, or pressure to keep interacting. Learners need room to struggle, revise, succeed, and seek human help.
A tool, not a replacement relationship.
We want AI interactions to be transparent about their role and limitations. Our direction is to avoid dependency-building behavior, emotional exclusivity, and manipulative engagement. Collaboration with peers, educators, and families should remain part of learning. An educational assistant is not a therapist or a substitute for trusted adults.
Keep consequential judgment human.
AI may suggest a learning gap or a next step; it should not turn an uncertain observation into a definitive label. Educators should be able to inspect and challenge recommendations. Before a pilot, escalation responsibilities and safeguarding pathways need to be defined with the participating institution.
Remember only with a purpose.
We aim to minimize learner data and avoid unnecessary psychological profiling or sensitive personal inferences. The information used for personalization should have a clear purpose. Retention, access, deletion, and age-appropriate transparency need explicit rules; privacy-preserving architectures remain an area of investigation.
Strengthen teachers’ capacity.
Educators are partners, not obstacles to adoption. Tools should fit classroom practice, make useful observations understandable, and avoid unreasonable administrative burden. Learning gains and educator workload both matter when evaluating an intervention.
Evidence before certainty.
We distinguish hypotheses, prototypes, observations, and established findings. Research should document methods, limitations, relevant conflicts of interest, and its relationship to SILT-developed products. An engaging demonstration is not evidence that a system improves learning.