Overview

This has been a busy year, as AI technology has continued to evolve rapidly even as the global world order has experienced significant changes. Four key trends for 2025 have included our team’s shift towards more building and prototyping; more generally, the increasing empowerment of AI models with tools and information; within the UN system, the emergence of a more strategic approach to leveraging AI; and within the development and humanitarian community, the need to articulate how AI can help in a time of system-wide funding crises. 

Looking forward,  we’re excited about the growth in AI capabilities that we’ve seen over the last year. We’re optimistic about the implications for UNICEF and our ways of working – whether this means developing effective human-in-the-loop workflows in which staff work alongside AI to balance efficiency and control; empowering staff to experiment and prototype with the help of new AI coding tools; or better articulating where the concrete opportunities lie to incorporate AI into UNICEF programmes (from education to health), with a particular focus on supporting frontline workers. 

What’s been going on

A Year of Prototyping

In the Frontier Technology Team, we’ve tried to keep up with the emerging technology via a “learn by building” mindset, creating prototypes for tasks such as: 

• machine translation, a.k.a. our Eshu prototype -- which allows users to pass their documents to multiple open and commercial machine translation models and helps them evaluate which model is best suited to their use case;
• text simplification, via a prototype which allows users to simplify documents according to different standards and scores the change in complexity after this transformation;
• easy read, via a prototype which transforms text into Easy Read format (including retrieval of suitable accompanying images/symbols), and allows the user to edit and modify the resulting content as needed. 

While much of our work focuses on advising and empowering others to build, we’ve found that we can learn a lot about the capabilities and limitations of AI through creating these simple demos, and that they can act as introductions to deeper conversations with teams about how AI can address their needs.  

A focus on equipping AI models with the tools and information needed to succeed

From a technical standpoint, two big trends have shaped the development of AI systems this year: AI agents, and context engineering. In practical terms, this has marked a shift from AI models that “know about the world in general” to more tailored systems that “have access to information and capabilities relevant to your specific problem” (a trend already started by the rise of retrieval-augmented generation in 2024). 

AI agents – “models using tools in a loop to achieve a goal” -- are generative AI models equipped with a set of tools and empowered with discretion around when to use them in service of a user request. A tool could perform a function (e.g. conducting a calculation), retrieve information (e.g. checking today’s weather), or take action (e.g. updating a file). The rise of agents has made us think about our prototypes differently; not as standalone applications, but as services that can be integrated into different applications and AI-powered front ends. To this end, we have begun exploring the creation of a UNICEF-specific MCP registry, which is a collection of servers hosting different tools (like translation, text simplification, and easy read) capable of exchanging information with a user-facing LLM. This will make it possible to answer requests like “take this PDF on the importance of eating healthy food, convert it into easy read format, and translate it into Bengali” using a single chatbot interface – and for other UNICEF teams to incorporate our tools into their own interfaces. Ultimately, we envision that the registry will include tools developed by other UNICEF teams, as well as external tools that might be useful to UNICEF users – such as tools that receive data or maps from outside sources, or invoke predictive models built by start-ups we invest in. 

The move towards more tailored AI systems and the rise of AI agents aligns with a growing emphasis on context engineering – equipping LLMs with curated, relevant  information to help them succeed, in a format they can effectively use. Context could include information about available tools as well as documents, knowledge bases, or logs and summaries of the conversation history. Context is one asset that UNICEF teams have plenty of; for example, this year we started working with our Nutrition in Emergencies team to explore how AI can make the Global Nutrition Cluster’s resources – which include over 8,000 documents in multiple languages, as well as online course content – accessible in bite-sized formats. We look forward to identifying opportunities for effectively structuring and leveraging this context to power additional, meaningful applications in 2026.  

In UNICEF and beyond, an increasingly strategic approach to AI

While the Office of Innovation and many of our colleagues have been experimenting with AI-powered technologies and solutions for years, in 2025 we finalized our organization-wide AI strategy. The strategy identified three priority areas where AI can deliver results for UNICEF: (1) programmatic use cases that directly impact our work with communities on the ground; (2) operational use cases that support UNICEF staff to do their work more effectively; and (3) advocacy use cases to promote safe and beneficial AI systems for children. It also established an AI hub to coordinate colleagues across the organization via working groups focused on key capability areas (value realization, risk and governance, etc.). This work was complemented by UNICEF’s release of its Digital Transformation Strategy. 

At the UN system level, we saw efforts to realize some of the promises set out in 2024 via the Global Digital Compact. A General Assembly Resolution established two key bodies – the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance—to help inform the UN of the promises and risks of AI and coordinate its policy and governance response. On a more practical level, ITU established the UN AI Resource Hub, which builds on years of work compiling annual reports on AI activities across the UN system. 

Innovation at a time of challenging transformations

Even as UNICEF and the broader UN system have been laying out ambitious plans to scale and leverage AI, the humanitarian and development sectors as a whole have entered a period of austerity and restructuring, with a global funding crisis reducing the budgets of many agencies and organizations. This has led many to look to emerging technologies to “do more with less”. We’ve done some exciting work in this direction, including a pilot of causal machine learning in Kazakhstan that aimed to estimate the impact of cash transfers on different households in order to identify how these resources could be channeled to households that benefit the most; a pilot with our supply division to explore how to optimize production schedules in order to reduce the cost of developing education kits; and continuing our work on anomaly detection for cash transfers in Yemen. It's safe to bet that leveraging AI to generate efficiencies will be a continued focus in 2026. 

Where we’re headed

A return to workflows and humans in the loop

We’ve spent 2025 grappling with Amara’s law, which states that “we tend to overestimate the effects of technology in the short term, and underestimate them in the long term.” For example, a McKinsey study found that while 88% of enterprises use AI, 62% were still in the piloting and experimentation phase – despite the many promises of AI transformation, we still have a lot of work to do towards concrete value realization. Similarly, this year we’ve seen reversals of optimistic attempts to eliminate human jobs by replacing them with AI. Yet, it is clear that AI is here to stay and will increasingly be a part of organizational workflows. 

What does this mean in practice for UNICEF? A continued and reinforced effort to implement AI systems that leverage humans in the loop, with AI positioned as a tool that can empower humans to execute their jobs better – a concept that has been manifested in the form of partially autonomous apps. An interesting example comes from our collaboration with the Learning Innovation Hub’s accessible digital textbooks (ADTs) project, which in 2025 released an open-source tool for creating ADTs from PDFs with the help of generative AI and piloted this technology in Uruguay. While we started with an emphasis on automating ADT generation end to end, we’ve pivoted to focusing on ways to audit, modify, and intervene in that AI-powered workflow – including  building out tools for systematically evaluating how different steps in the workflow perform; experimenting with the creation of an agentic editor for ADTs; and building an interface that allows users to modify the outputs of the generative AI pipeline before publishing the results. 

A move towards AI-powered DIY

In line with the trend towards empowering humans by design in AI systems, this year has seen an increase in the ease with which anyone can build systems with the help of AI – or what I think of as a shift towards “do-it-yourself with AI”. The beginning of the year saw a meteoric rise in the popularity of “vibe coding” -- or developing software by instructing an LLM in plain language – which has translated into code agents such as Cursor and Github Copilot (oriented towards developers), as well as no-code interfaces such as Lovable (oriented to a broader user base). In the frontier tech team, we’re starting to look at ways to enable UNICEF staff to build and design their own tools directly with the help of AI – whether by helping them quickly create simple chatbots that leverage a custom set of our MCP registry tools, or by developing a conversational interface that allows them to code prototypes of their own. 

Increased clarity of vision around programmatic uses of AI

As the question of how to build becomes increasingly easier to tackle, a key focus of this year has been what to build. In this respect, we’ve seen a fortunate outcome of the organizational reforms driven by this year’s Future Focus Initiative – the Office of Innovation will move under UNICEF’s Programme Group.  This has brought us closer to colleagues that we are already collaborating with across sectors including health, education, and WASH.  

After spending much of 2025 identifying the most promising AI use cases for these different impact areas, we’ve emerged with some clear priorities in 2026 – such as a focus on uses of AI for empowering frontline workers, whether in the form of community health worker upskilling or tools for teacher support (building on previous and current investments in technologies like Angaza Elimu’s WhatsApp-based teacher support tool, Bookbot's app for interactive reading feedback, and a chatbot for supporting computer science education in Serbia). 

What this means for UNICEF/Children/Social Impact

Our hope is that this year will build on the progress we made in 2025, which has brought us closer to useful, informed, and empowering AI systems that advance outcomes for children, as well as provided some strategic clarity on how to best extract real value from AI advances. We’re optimistic about what we can achieve together as an organization and in collaboration with our partners. 

At the same time, we see big risks and challenges ahead. We’re thinking about the difficulties of effectively incorporating AI in learning given children’s neuroplasticity – since we’ve seen claims that using AI can reduce cognitive activity and potentially generate dependance even as it yields big gains for students – as well as how best to upskill students in this technology and prepare them for the jobs of the future. We are proud that UNICEF has released an updated version of its Policy Guidance on AI for Children that accounts for the transformational changes wrought by generative AI – but we know that there are many challenges ahead, including additional examples of the risks of children’s engagement with AI systems in the absence of sufficient guardrails and the generation of harmful content. 

As always, we look forward to another year of exploring how to make frontier technology work for children, mitigating risks and harms while maximizing benefits. We invite you to follow along with us as we continue to learn in 2026.