The World Economic Forum sets out seven practices to make AI safe for children without shutting them out of learning
A September 2026 document from the Global Coalition for Digital Safety takes children’s well-being, and not merely the absence of incidents, as its goal, and proposes practices for those who design AI and incentives to ensure they are followed.
“Children’s Safety and Well-Being in the Age of AI: Best Practices and Incentives” is a World Economic Forum (WEF) white paper dated September 2026. It stems from the Global Coalition for Digital Safety, a Forum initiative, and the foreword is signed by Agustina Callegari, head of technology governance initiatives, and Cathy Li, head of the Centre for AI Excellence. The disclaimer explains that the conclusions are the product of a collaborative process and do not necessarily represent the views of the Forum or its members, and that generative AI tools were used to support drafting, with review by the authors.
It is not a classroom guide. It is aimed at those who shape children’s experience of AI: developers and model providers, app stores, regulators, families and also educators. It defines a “child” as anyone under 18, in line with the Convention on the Rights of the Child, and warns that this is a legal category and not a developmental profile. Hence its central thesis: there is no model that fits every case (“There is no one-size-fits-all model”), and the response cannot be reduced to blanket bans or age barriers.
The document argues that age alone does not make a service safe. Age verification is useful for applying differentiated safeguards, but it does not determine maturity, moderation or redress for harm. Nor does it accept that broad restrictions are free of cost: they can block benefits, reduce opportunities to acquire skills and push children towards less controlled services. It also points out that generative and agentic AI change the nature of the interaction, because they can sustain continuous exchanges, remember, personalise and act autonomously, and that humanlike language can lead minors to overestimate the system’s understanding or care.
The seven practices it proposes for organisations are: design for safety, age and development; encourage beneficial use while preserving one’s own judgement and human relationships; prevent harmful content, conduct and interactions, including child sexual exploitation and AI-mediated manipulation; protect privacy, dignity and informed understanding; strengthen learning, resilience and critical thinking; establish accountability, human oversight and continuous improvement; and promote inclusion, child participation and shared responsibility. Among the controls it mentions for minors are reducing, pausing or switching off an AI feature, declining memory or the reuse of conversations, and asking for a different explanation.
For education, its idea of learning is of particular interest: exposure does not create competence, and the use of AI “should not systematically displace the effort, feedback, discussion, play and social learning through which children build durable knowledge and skills”. It also argues that the test of a tool is relational: whether it helps the young person return to people, activities and decisions with more confidence, or whether its design prioritises itself as the preferred relationship.
The third part, on incentives, is what sets the document apart. It starts from the premise that protection will not be implemented consistently if it is costly or unprofitable, and proposes a portfolio of measures: comparable safety reports, procurement and distribution requirements, financial and insurance signals, shared testing infrastructure, stronger government capacity, safe channels for responsible disclosure and internal accountability tied to measurable outcomes for children. The authors acknowledge that this is a shared foundation and not a closed technical standard, and that the evidence on medium- and long-term effects is still limited.
In the library it is thematically related to the report for UNESCO on the right age for AI (our article) and to the narrative review “Inteligencia artificial y derechos del niño” (“Artificial intelligence and children’s rights”) (in the library). A teacher or school leadership team can use it as a checklist of criteria when evaluating or procuring tools: what controls students have, what data is kept, what happens when something goes wrong. It is best to start with the introduction and the first chapter (pages 5 to 10), which define the aim and the challenges, and continue with the section on learning and critical thinking in the second (page 16).
The World Economic Forum sets out seven practices to make AI safe for children without shutting them out of learning
A September 2026 document from the Global Coalition for Digital Safety takes children’s well-being, and not merely the absence of incidents, as its goal, and proposes practices for those who design AI and incentives to ensure they are followed.
“Children’s Safety and Well-Being in the Age of AI: Best Practices and Incentives” is a World Economic Forum (WEF) white paper dated September 2026. It stems from the Global Coalition for Digital Safety, a Forum initiative, and the foreword is signed by Agustina Callegari, head of technology governance initiatives, and Cathy Li, head of the Centre for AI Excellence. The disclaimer explains that the conclusions are the product of a collaborative process and do not necessarily represent the views of the Forum or its members, and that generative AI tools were used to support drafting, with review by the authors.
It is not a classroom guide. It is aimed at those who shape children’s experience of AI: developers and model providers, app stores, regulators, families and also educators. It defines a “child” as anyone under 18, in line with the Convention on the Rights of the Child, and warns that this is a legal category and not a developmental profile. Hence its central thesis: there is no model that fits every case (“There is no one-size-fits-all model”), and the response cannot be reduced to blanket bans or age barriers.
The document argues that age alone does not make a service safe. Age verification is useful for applying differentiated safeguards, but it does not determine maturity, moderation or redress for harm. Nor does it accept that broad restrictions are free of cost: they can block benefits, reduce opportunities to acquire skills and push children towards less controlled services. It also points out that generative and agentic AI change the nature of the interaction, because they can sustain continuous exchanges, remember, personalise and act autonomously, and that humanlike language can lead minors to overestimate the system’s understanding or care.
The seven practices it proposes for organisations are: design for safety, age and development; encourage beneficial use while preserving one’s own judgement and human relationships; prevent harmful content, conduct and interactions, including child sexual exploitation and AI-mediated manipulation; protect privacy, dignity and informed understanding; strengthen learning, resilience and critical thinking; establish accountability, human oversight and continuous improvement; and promote inclusion, child participation and shared responsibility. Among the controls it mentions for minors are reducing, pausing or switching off an AI feature, declining memory or the reuse of conversations, and asking for a different explanation.
For education, its idea of learning is of particular interest: exposure does not create competence, and the use of AI “should not systematically displace the effort, feedback, discussion, play and social learning through which children build durable knowledge and skills”. It also argues that the test of a tool is relational: whether it helps the young person return to people, activities and decisions with more confidence, or whether its design prioritises itself as the preferred relationship.
The third part, on incentives, is what sets the document apart. It starts from the premise that protection will not be implemented consistently if it is costly or unprofitable, and proposes a portfolio of measures: comparable safety reports, procurement and distribution requirements, financial and insurance signals, shared testing infrastructure, stronger government capacity, safe channels for responsible disclosure and internal accountability tied to measurable outcomes for children. The authors acknowledge that this is a shared foundation and not a closed technical standard, and that the evidence on medium- and long-term effects is still limited.
In the library it is thematically related to the report for UNESCO on the right age for AI (our article) and to the narrative review “Inteligencia artificial y derechos del niño” (“Artificial intelligence and children’s rights”) (in the library). A teacher or school leadership team can use it as a checklist of criteria when evaluating or procuring tools: what controls students have, what data is kept, what happens when something goes wrong. It is best to start with the introduction and the first chapter (pages 5 to 10), which define the aim and the challenges, and continue with the section on learning and critical thinking in the second (page 16).
