When data and AI shape public services, children experience their consequences first and longest. This piece by Judith Hanan, UNICEF Technology for Development Manager, explores why a child‑centred lens fundamentally changes how governments must think about data, AI, and accountability.
For UNICEF, data sovereignty is not an abstract debate about servers, jurisdictions, or technical architecture. It is experienced through the processes, tools, and systems that sit at the foundation of digital government, and through the tangible and sometimes intangible ways they shape children’s lives over time. It is visible when a birth registration system cannot be accessed during an emergency, or when a child protection case is routed by a system no one fully understands, and less visible when families begin to lose trust in public services because decisions feel distant, automated, or impossible to question.
When governments collect data about children, it is rarely a matter of choice. Health records, education histories, social protection registries, and child protection case files exist because a child needs support. Children do not consent in the way adults do, and they cannot meaningfully challenge how their data is used. That imbalance is why questions of control, accountability, and governance carry more weight when children are involved.
Across countries, we see growing pressure on governments to modernise public services through data and AI. Specialised models are being proposed to improve targeting, speed up decisions, and stretch limited resources. At the same time, governments are expected to guarantee children’s rights, protect highly sensitive data, and remain accountable when systems fail. Data sovereignty sits at the core of this tension.
From a child rights perspective, sovereignty is not really about where data sits. It is about whether the state can still show up for children as the duty bearer, especially when decisions are hard and trade-offs are real. That responsibility is anchored in law, exercised through the daily operation of public services, and safeguarded in the choices governments deliberately retain for the future.
Governments must be able to enforce national law and uphold children’s rights even when data moves across systems or borders. They must be able to keep essential services running during emergencies, political change, or vendor failure. They must also retain the ability to adapt systems over time without losing control of children’s data or institutional memory. Children grow up inside the systems designed today, and the consequences of poor decisions do not disappear when a contract ends.
Where do specialised models fit within this landscape?
Specialised AI models can play a constructive role in this landscape. In some contexts, they genuinely help governments deliver better outcomes for children. Models trained on local language, policy frameworks, and service workflows can better reflect how children and caregivers actually interact with public services. This matters in education, health, and protection systems where context and cultural understanding can reduce error and harm.
When designed with care, specialised models can also strengthen safeguards. Processing sensitive children’s data within government-controlled environments, with clear governance, auditability, and human oversight, can support privacy by design rather than undermine it. For children, principles like minimisation (minimising the collection of data), purpose limitation (collecting data for specific purpose), and traceability (tracking data and how it moves through a system) are not technical preferences. They are expressions of rights.
There is also a very practical benefit. Teachers, health workers, social workers, and case managers are often overwhelmed by administrative tasks. Well-designed tools can support summarisation, forecasting, and routine analysis, allowing frontline workers to spend more time with children and less time navigating systems.
... when a child is registered for school, that moment could also connect to their vaccination history or social protection record, allowing government to see the child as a whole rather than as separate entries across systems.
When does specialisation become a risk?
The risks appear when specialised models are treated as shortcuts rather than public systems. When models arrive as black boxes with no visibility into internal logic or structure, accountability becomes blurred. Decisions that determine whether children receive services are harder to account for and harder to fix when they go wrong. In child protection and education, opacity is not an abstract risk: it has real and harmful consequences.
Fragmentation is a less obvious harm, but its consquences are no less serious. When ministries or programmes build models in isolation, integration fails by design. Children do not move through digital public services as separate systems, and when digital connections are weak, transitions become fragile and children are more likely to fall through the gaps. The approach should be focused on end-to-end delivery, delivering services across ministerial boundaries. For example, when a child is registered for school, that moment could also connect to their vaccination history or social protection record, allowing government to see the child as a whole rather than as separate entries across systems.
There is also a risk of mistaking proximity for control. Hosting a model locally does not automatically mean a government governs it. If licensing terms, updates, evaluation methods, or monitoring remain externally controlled, sovereignty exists only on paper. When something goes wrong, children are still exposed to decisions made elsewhere.
What about bias and exclusion?
As with anything related to AI and data, there should be a relentless focus on bias – the unintended risks it can create deserves particular attention. Children are often underrepresented or unevenly represented in administrative data. Marginalised children are especially likely to be missing. Specialised models trained on partial or uneven data do not just reflect exclusion. They harden it into the system, reproducing it at speed and at scale. Without strong governance and testing, specialisation can magnify existing inequalities rather than reduce them.
The most sensitive children’s data requires the strongest governance
Can models replace or compensate for weak systems?
The simple answer is no. None of these issues can be solved by models alone. No algorithm can compensate for weak civil registration systems, poor interoperability, fragmented data standards, or unclear data ownership. For children, these foundations are what allow continuity of care through the delivery of public services to meet their changing health and educational needs throughout childhood.
A simple lens helps clarify whether a specialised model strengthens or weakens a government’s ability to serve children. Can the government change or exit the model without disrupting essential services for children? Can decisions that affect children be explained, audited, and challenged. Does the model reinforce shared systems or create another isolated solution? These answers determine whether children are protected or put at risk.
A child-centred approach does not argue for rejecting technology or retreating into isolation. It argues for deliberate choices grounded in rights, resilience, and long-term public value. The most sensitive children’s data requires the strongest governance. Specialised models should connect to shared digital public infrastructure rather than bypass it, and investment in algorithms must be matched by investment in data governance, interoperability, and public sector skills.
For UNICEF, data sovereignty ultimately comes back to a simple question. When systems change, vendors move on, or crises hit, can governments still keep their promises to children? If the answer is yes, technology is doing its job. If the answer is no, no amount of sophistication will make it right.
