UNICEF's 2021 AI guidance named chatbot risks to children

The design

In November 2021, UNICEF's Office of Global Insight and Policy published version 2.0 of 'Policy Guidance on AI for Children,' a document its own title page describes as a 'working document ... prepared to facilitate the exchange of knowledge and to stimulate discussion.' The guidance itself states it followed a draft version 1.0 released in September 2020 and incorporated feedback from AI developers, companies, governments, civil society and, separately, consultations with children; the current version 3.0, from 2025, has since superseded it, as UNICEF's own current guidance landing page confirms.

What the evidence says

The 2021 document devotes a named section to 'AI-powered voice assistants and chatbots,' stating that 'millions of children are being shaped by them either emotionally or behaviourally' while citing, as a specific example, a 2018 BBC test of two mental-health chatbots that 'failed to properly handle children's reports of sexual abuse' despite both being considered child-suitable; UNICEF attributes that finding to the BBC's testing, not to its own research. The same section quotes a separate UNICEF briefing's finding that 'when not designed carefully, chatbots can compound rather than dispel distress,' again naming that claim's source rather than presenting it as this document's own original research.

What it asks of people

The guidance sets out nine 'requirements for child-centred AI,' each paired with a first-person line the document attributes to consultations with children, such as 'I need to be safe in the AI world' under the safety requirement; it asks governments and companies to apply these requirements whenever an AI system interacts with children, regardless of whether the system was designed for them. The document states its child-consultation process involved nine workshops with children mainly aged 14 to 16, a specific but limited method the guidance names openly rather than presenting as a comprehensive survey of children's views.

Privacy and safeguards

Requirement four, 'Protect children's data and privacy,' is stated in the child's-voice framing as 'Ensure my privacy in an AI world,' and the surrounding text raises voice assistants' reliance on stored voice recordings as a specific, named concern about unclear company data-retention policies and consent; the document does not audit any single company's practice against this requirement, since it is guidance for future policy and product design rather than an enforcement record.

• Which named chatbot or voice-assistant product, if any, has been assessed against UNICEF's nine requirements since 2021?
• Has the 2018 BBC test UNICEF cites been repeated against current-generation companion chatbots?
• How did the 2025 version 3.0 guidance revise the chatbot-specific concerns this 2021 version raised?

The 2021 guidance is a policy document, not a product audit: it names risks, cites external testing and a UNICEF briefing for specific claims, and sets requirements rather than measuring compliance. Reading it as a source means citing what it names and attributes, not treating its requirements as evidence that any particular chatbot meets or fails them.