In 2025, Artificial Intelligence (AI) has become nearly ubiquitous in humanitarian work. Yet few fully trust it. According to recent research, 93% of humanitarian practitioners report using AI tools, but only 38% believe these tools improve decision-making. Less than half agree it has improved operational efficiency, and nearly 30% remain neutral or uncertain about AI's benefits altogether. This contradiction—high adoption but low confidence—is what researchers call the “humanitarian AI paradox” (Humanitarian Leadership Academy & Data Friendly Space, 2025).
This paradox defines much of my daily work as a humanitarian analyst. Every day, I oversee the use of GANNET, Data Friendly Space (DFS)’s AI-powered platforms that process humanitarian data at scale, ensuring its outputs truly reflect realities on the ground. My task is not to replace AI, but to guide it: to decide what is relevant, what needs verification, and what deserves deeper exploration. In other words, I am the “human in the loop” who ensures AI works for people, not just about people.
The process starts with data collection. AI systems ingest information from dozens of sources across multiple languages. For the GANNET Myanmar Situation Hub, that means processing reports in Burmese alongside English, French, and Arabic materials. The system synthesizes displacement figures, protection incidents, and food security data, among others, faster than any human analyst could manage.
Then comes AI analysis. The system identifies patterns, flags anomalies, and summarizes key trends. On a good day, it surfaces connections I might have missed, an emerging displacement route correlating with new checkpoints, or a spike in protection concerns not yet visible in formal reporting. The speed and coverage are genuinely valuable; when you are tracking multiple crises simultaneously, that initial synthesis matters.
But this is where the critical work begins: human verification. Every output requires expert validation. I verify sources, cross-check claims, assess relevance, and determine what the analysis is actually telling us about conditions in the field. This human verification process is crucial, as it ensures that the AI's findings are not taken at face value but are critically examined for accuracy and relevance. Sometimes the AI flags something as minor that turns out to be a serious protection concern. Other times, it highlights data points that, while accurate, miss the broader context that humanitarian responders need.
Consider an example from the 2025 cholera outbreak in Chad (DFS, October 2025). The AI-generated analysis correctly captured quantitative indicators, such as how the distance to water sources creates "additional concerns regarding protection", but did not elaborate further. A human analyst immediately recognized the more profound implications: women and girls walking over 5 kilometers for water in displacement contexts face heightened exposure to gender-based violence. At the same time, the time burden prevents children (particularly girls) from attending school. This transforms the issue from a technical WASH gap into an urgent protection and education concern requiring a coordinated multi-sectoral response. The AI flagged protection as a concern, but could not translate distance metrics into the lived realities of risk and vulnerability that drive humanitarian decisions.
After verification comes customization; the same analysis looks different for an inter-agency situation update, a protection-focused brief, or a donor report. Human analysts adapt outputs for specific audiences and purposes. We adjust terminology, emphasize different findings, and interpret implications based on who needs to make what decisions. This adaptability of AI-generated analysis, when guided by human analysts, ensures that the insights are tailored to the specific needs of different stakeholders, enhancing their utility and relevance.
When this process works as intended, the result is actionable insights: faster, more comprehensive, and genuinely useful for humanitarian decision-making.
AI brings clear advantages. It processes information quickly, handles large volumes of data, works across languages, and helps identify patterns that would otherwise take days to detect. Yet it also shows consistent weaknesses. It often misses nuances, offers plausible but superficial answers, and struggles to interpret context. For instance, an algorithm might not distinguish between a protest that is part of regular civic activity and one that signals escalating tensions requiring protection monitoring.
During situation monitoring in Lebanon, AI-generated summaries tended to emphasize numerical data, such as the volume of aid delivered, while paying little attention to critical protection incidents. Brief mentions of checkpoint harassment carried far greater operational significance than many of the figures highlighted. An algorithm cannot judge such nuances; a human analyst, guided by contextual understanding and protection principles, can.
Beyond analytical limitations, practical barriers also limit the usefulness of AI tools in humanitarian settings. In areas with poor connectivity or low digital literacy, advanced tools often remain out of reach. The sector needs solutions that operate effectively in the field, not only in coordination hubs. Some encouraging progress has been made, such as the offline AI assistant for operational security developed by an NGO in Lebanon (Humanitarian Leadership Academy & Data Friendly Space, 2025), but these remain the exception rather than the rule.
The issue is not that AI is unreliable, but that it requires skilled human oversight. Algorithms process information according to patterns in data; humanitarian judgment depends on context, ethics, and accountability to affected communities.
A growing concern is that many professionals are using AI tools before their organizations have established clear policies or training frameworks for their responsible use. Only 22% of survey respondents reported that their organizations have formal AI policies in place, and just 8% said AI is widely integrated with adequate institutional support (Humanitarian Leadership Academy & Data Friendly Space, 2025). This lack of clear guidelines and support can lead to decisions affecting populations in vulnerable situations being informed by tools that few people know how to use responsibly, within institutions that haven't fully grappled with their ethical implications.
This means decisions affecting populations in vulnerable situations are often informed by tools that few people know how to use responsibly, within institutions that haven’t fully grappled with their ethical implications. As NetHope's Humanitarian AI Code of Conduct warns, "AI technology has the potential to entrench inequality further, deepen existing divides (including the digital and gender divides), and worsen conflict and fragility" (NetHope, July 2025).
The risks are real. When AI systems are trained on biased or incomplete data, they risk reproducing those biases; silencing particular perspectives, misrepresenting local contexts, or reinforcing structural inequalities. Generic language models cannot automatically capture the political nuances of Sudan’s conflict or the protection implications of military checkpoints in Myanmar. As the United Nations University observed, "while AI personas might make information more accessible, they could simultaneously distance decision-makers from the lived realities of crisis-affected populations" (UNU, July 2025).
To mitigate these risks, DFS prioritizes the curation of data from diverse sources and perspectives. Lessons learned from each analysis cycle are documented and used to strengthen methodologies. DFS analysts work closely with the Assessment & Analysis Cell during both UNDAC and non-UNDAC activations, ensuring that analytical outputs remain technically rigorous and operationally relevant. Beyond these deployments, DFS also collaborates regularly with NGO forums and coordination bodies across contexts such as Sudan, Syria, and Lebanon, and has previously supported diverse humanitarian mechanisms in multiple countries to strengthen analysis, coordination, and decision-making processes.
Even with these safeguards, human oversight remains essential. Accountability cannot be automated. When AI contributes to humanitarian analysis, it is ultimately human analysts who bear responsibility for the outcomes. If automated outputs distort protection risks or overlook marginalized voices, it is the role of analysts and institutions to identify and correct those gaps.
Genuine accountability means keeping affected communities at the center of every process. It requires ensuring that AI serves humanitarian principles rather than efficiency alone, and that its role and limitations remain transparent. As the UN Office for the Coordination of Humanitarian Affairs notes, “AI has the potential to drive operational gains for humanitarians through enhanced efficiency... but must be balanced with risk management, transparency, and oversight.” (OCHA, April 2024). That responsibility falls on human decision-makers, not algorithms.
So, what does responsible AI supervision actually look like in practice?
It is not about letting algorithms run unchecked; it is about creating a constant dialogue between machine outputs and human judgment. In the GANNET SituationHub, analysts act as active curators of information rather than passive consumers of AI outputs. We select and validate data sources, ensuring trusted local sources are included alongside international media. We refine AI-generated analysis to align with established humanitarian frameworks, including the Humanitarian Programme Cycle, protection mainstreaming, and sectoral analysis standards. A number of our projects are also grounded in the Joint Intersectoral Analysis Framework (JIAF), which has guided the analytical structure and ensured coherence, comparability, and operational relevance across multiple SituationHub deployments.
This collaboration allows each to do what it does best: the system collects and synthesizes data from multiple sources, filtering and organizing it efficiently, while analysts interpret the patterns in context for people on the ground.
The reverse can also happen. In the Philippines Earthquake Situation Assessment (October 2025), GANNET initially produced summaries that accurately reflected the immediate impacts of the disaster but missed how pre-existing vulnerabilities, such as poverty, marginalization, and prior displacement, would exacerbate humanitarian needs. Human review reframed the analysis, shifting it toward a more protection-sensitive understanding of the crisis.
This continuous cycle — AI analysis, human verification, and collaborative refinement — is what makes the human-in-the-loop model effective. But its success depends on skilled analysts, strong institutional support, and sustained investment. As the International Committee of the Red Cross emphasizes, “human control and judgment must be integral in any decision affecting human life or dignity” (ICRC, July 2023). Responsible use of AI is not without cost; it requires training, resources, and a genuine organizational commitment to ethical practice.
The humanitarian sector stands at a turning point. AI adoption is already happening; the question is whether it will happen responsibly. Evidence shows that individual use is far outpacing institutional readiness. Practitioners turn to AI to manage overwhelming workloads, yet many do so without guidance, training, or ethical frameworks. In practice, this means large parts of the sector are experimenting with AI in real time, without adequate safeguards.
The way forward is not overly complex, but it does require investment. Humanitarian organizations need formal AI policies and training that go beyond technical skills to include ethics, bias awareness, and quality control. Human AI oversight should be recognized as a professional competency, not an add-on to an already full workload.
Donors and policymakers also play a role. Responsible AI is not about replacing people with technology; it is about enabling skilled humanitarians to use it safely and effectively. Funding should support organizations developing human-in-the-loop models, ensuring they have the resources needed to implement responsible and effective AI practices. Practitioners, too, need to advocate for institutional safeguards and transparency, making sure AI complements, not replaces, human judgment.
Data Friendly Space is among those investing seriously in the human-in-the-loop model, not because it is easier or cheaper, but because it reflects how responsible AI should work in humanitarian contexts. Yet scaling this approach requires broader sector commitment and resources for training, policies for ethical oversight, and institutional structures that value human expertise alongside technical capability. The humanitarian AI paradox is not inevitable; it results from rushing to adopt new tools without building the institutional foundations to support them.
AI will not replace humanitarians, but it can amplify what they do if the humanitarian sector invests in the right kind of partnership. Humanitarian work is fundamentally about serving people, not processing data. Ensuring that technology serves communities rather than the other way around remains a distinctly human responsibility.
