UNICEF Policy Guidance on AI for Children (2021)

Per UNICEF AI for Children: Support Children's Development and Well-being + Best Interests of the Child + Age-Appropriate Design.

• UNICEF AI evidence for UNICEFAI-1

• best interests + age-appropriate partial

Per UNICEF AI: Ensure Inclusion of and for Children + Accessibility + non-discrimination + bias prevention.

• UNICEF AI evidence for UNICEFAI-2

• best interests + age-appropriate partial

Per UNICEF AI: child privacy + safety + Empowerment + child agency + voice.

• UNICEF AI evidence for UNICEFAI-3

• best interests + age-appropriate partial

Per UNICEF AI: transparency + explanation for children + adult capacity building + create enabling environment.

• UNICEF AI evidence for UNICEFAI-4

• best interests + age-appropriate partial

Map AI system design choices and outcomes against the four general principles of the Convention on the Rights of the Child: non-discrimination, best interests of the child, right to life and development, and the right to be heard.

• CRC alignment matrix for the AI system
• Best-interest determination notes
• Participation log capturing children's input
• Mitigation actions linked to specific articles

• No reference to CRC in design records
• Best-interest analysis treated as a one-off form
• No participation of children in design

Set defaults, features, and interactions according to the age and developmental stage of the child user, with the most protective settings on by default.

• Age-tier configuration matrix
• Default privacy and safety setting record
• User journey map by age band
• QA results verifying default behaviour

• Single configuration applied across all ages
• Protective defaults disabled at install
• No verification of age tier behaviour in QA

Provide caregivers with meaningful and usable controls over a child's interaction with AI systems, including visibility into data use and the ability to adjust or withdraw consent.

• Caregiver control panel design
• Consent withdrawal workflow
• Visibility report for caregivers
• Usability testing with caregivers

• Controls buried in account settings
• No way to withdraw consent without deleting the account
• Caregiver dashboard lacks data on AI use

Involve children in the design, evaluation, and review of AI systems that affect them, using methods appropriate to their age and capacity.

• Co-design session plans and notes
• Child participation policy with safeguarding controls
• Evidence that input changed the design
• Ethics approval for child research

• No children involved at any stage
• Participation tokenistic and not reflected in design
• No safeguarding controls for research sessions

Provide additional safeguards for children in vulnerable situations, including children in humanitarian settings, children with disabilities, refugee and migrant children, and children in alternative care.

• Vulnerability risk register for AI deployment
• Adapted protocols for humanitarian or low-connectivity contexts
• Safeguarding plan referencing local child protection authorities
• Records of access provisioning for assistive needs

• Single design assumed for all contexts
• No safeguarding contact for humanitarian deployments
• Accessibility needs treated as edge cases

Restrict profiling, behavioural advertising, and recommendation systems that may exploit child users, and disable such features by default.

• Profiling policy for child users
• Recommender system risk assessment
• Default state record for behavioural advertising
• Independent audit of recommendation outcomes

• Behavioural advertising on by default
• Recommender amplifies harmful content to children
• No audit of recommender impact

Require suppliers of AI systems used by or affecting children to demonstrate compliance with child-centred requirements before contract award and during the term.

• Supplier questionnaire referencing UNICEF requirements
• Pre-award evaluation memo
• Ongoing supplier review schedule
• Termination rights for safeguarding failure

• Generic supplier questionnaire used
• No re-evaluation after contract start
• Contract lacks safeguarding termination clause

Provide accessible mechanisms for children, caregivers, and third parties to raise concerns about AI systems that affect children, and ensure independent oversight of high-risk uses.

• Independent advisory or ethics board terms
• Child-friendly complaint channel
• Case log with resolution times
• Reporting to oversight body

• Complaint channels not visible to children
• No oversight body for high-risk uses
• Cases unresolved or undocumented

Continuously monitor AI systems affecting children for safety, fairness, well-being, and rights impacts after deployment, and act on findings.

• Monitoring plan with child-specific indicators
• Periodic monitoring reports
• Trigger thresholds for intervention
• Corrective action log

• Monitoring focused on engagement metrics not well-being
• No thresholds defined for intervention
• Findings not acted on

Design and deploy AI systems that actively support the cognitive, social, emotional, and physical development of children, and do not displace activities essential to healthy growth.

• Child development impact assessment
• Age-appropriate design specification
• Evidence of consultation with paediatric or developmental experts
• Pilot study results with child cohorts

• Adult-oriented product simply rebranded for children
• No developmental experts consulted
• No measurement of effect on child well-being

Make AI systems inclusive of children with disabilities, of different genders, languages, cultures, and socio-economic backgrounds so that no child is excluded or disadvantaged by the system.

• Accessibility conformance report (WCAG 2.2 AA)
• Multilingual coverage list
• Bias and exclusion testing results across demographic segments
• Adaptations for children with disabilities

• Only one language supported
• No accessibility testing with assistive technologies
• Demographic gaps in training and validation data

Test and mitigate bias and discrimination in AI systems that affect children, including bias by age, gender, disability, ethnicity, language, and socio-economic status.

• Fairness metrics report by protected attribute
• Bias mitigation plan with corrective actions
• Disparate impact analysis
• Reviewer notes on contested outputs

• No fairness testing performed on under-13 cohorts
• Bias detected but not remediated
• Single fairness metric used in isolation

Collect only the minimum personal data needed from children, secure it appropriately, restrict secondary use, and design privacy controls that are understandable to children and caregivers.

• DPIA covering child users
• Data minimisation map
• Retention and deletion schedule
• Child-friendly privacy notice
• Caregiver consent records

• Adult-style privacy notice used for children
• Excess data collected for non-essential personalisation
• No deletion path for child accounts

Prevent AI systems from exposing children to harmful content, manipulation, harassment, or unsafe interactions, including through generative outputs and recommendation behaviour.

• Safety policy for child users
• Content filters and red-team test results
• Incident response plan with safeguarding escalation
• Logs of moderated outputs

• No red-teaming for child-targeted harms
• Filters tuned to adult sensitivities only
• No safeguarding escalation path

Explain to children and caregivers in age-appropriate terms when AI is being used, what it does, what data it uses, and how to challenge or correct outcomes.

• Age-tiered AI disclosure notices
• Child-readable explanation for automated decisions
• Caregiver dashboard for oversight
• Appeal and correction workflow

• AI use undisclosed inside the product
• Legalese explanations not adapted for children
• No way for a child or caregiver to contest an output

Build internal capability and stakeholder awareness so that policy makers, developers, procurers, and operators can apply child-centred AI requirements consistently.

• Child-centred AI training curriculum
• Stakeholder briefing pack
• Procurement guidance referencing UNICEF requirements
• Records of internal awareness sessions

• No training tailored to child-centred AI
• Procurement teams unaware of guidance
• Knowledge concentrated in one person

Provide children with age-appropriate AI literacy resources so they can understand, question, and engage with AI systems they encounter.

• AI literacy module aligned to age groups
• In-product learning prompts
• Partnership records with schools or youth organisations
• Evaluation of literacy outcomes

• No literacy content provided
• Literacy material aimed only at adults
• No measurement of comprehension

Adopt policies, procurement criteria, contracts, and oversight mechanisms that make child-centred AI the default rather than an optional add-on.

• Internal policy mandating child-centred AI review
• Supplier code of conduct referencing UNICEF guidance
• Contract clauses on child safety, privacy, and inclusion
• Independent oversight or advisory board terms

• Policy exists but is not enforced in procurement
• Contracts silent on child-specific obligations
• No external oversight of high-risk child-facing AI