No training, no rules, no apparent crisis: artificial intelligence is finding its way into humanitarian organisations through a myriad of individual initiatives that almost no one notices. A qualitative study conducted by the two authors reveals the hidden costs of this silent adoption – lost knowledge, exposed data, inequalities. Before a crisis that may be on the horizon?
The humanitarian sector has long prided itself on its ability to adapt to crises with agility and innovation. However, when it comes to artificial intelligence (AI), this is yet to be seen. In 2025, a sector-wide survey revealed a striking paradox: while most humanitarian workers had already integrated AI into their daily tasks, only a small fraction did so within any form of organisational governance framework.[1]Ka Man Parkinson, Madigan Johnson and Lucy Hall, “Artificial Intelligence in the Humanitarian Sector: Mapping Current Practice and Future Potential”, Humanitarian Leadership Academy, 2025–2026, … Continue reading
This article draws on a qualitative study of humanitarian practitioners to explore not whether AI is being adopted, but how, under what conditions, and with what consequences.[2]This article draws on seventeen semi-structured interviews conducted between March and May 2026 with practitioners from a variety of organisations (including Médecins Sans Frontières, CARE, Save … Continue reading The findings suggest that the current trajectory of AI use is not merely unstructured; it is actively producing new forms of risk, inequality, and silence within organisations. At the heart of this scenario lies a fundamental question: what happens when adoption of technology outpaces the governance structures meant to guide it? And most importantly: what changes are required to break from these self-perpetuating dynamics and allow the sector to move from individual adaptation to collective transformation?
The architecture of ungoverned adoption
Artificial intelligence is already embedded in humanitarian practice, but often without the organisational frameworks needed to guide its use.[3]This article does not explore other critical ethical dimensions of AI technology adoption, including environmental sustainability and geopolitical power asymmetries. Across interviews, practitioners consistently described AI adoption as unfolding in conditions of uncertainty, assumption and improvisation rather than deliberate institutional design. Many reported the routine use of commercial AI tools to process sensitive information, including protection data, health-related content and community feedback, without clear organisational guidance or oversight.
In one case, a team handling highly confidential information continued using unvetted platforms for more than a year before a formal risk assessment was conducted. When it finally took place, the exercise was so brief that staff questioned whether meaningful scrutiny had occurred at all.
This absence of governance was not limited to formal policy. It extended to basic organisational visibility. As one experienced practitioner observed: “There is no official policy … I think, although I am not certain, that colleagues are using it individually, by their own initiative.”
This uncertainty is not a neutral condition. A governance vacuum does not simply leave risks unmanaged; it actively generates them. Unstructured AI adoption does not remain static. It deepens over time, as practices normalise, habits solidify and increasingly consequential decisions are shaped by tools operating beyond institutional scrutiny.
As the following sections will show, this governance vacuum produces effects that go well beyond compliance risk. It shapes how practitioners learn, what kinds of practices become normalised, which forms of expertise remain visible or hidden, and ultimately how humanitarian organisations define the value AI is meant to create.
Therefore, what emerges is not merely a compliance problem, but a structural trust deficit: practitioners use tools they do not always fully understand, organisations hesitate to endorse practices they cannot confidently oversee and the communities affected by humanitarian decisions remain largely uninformed about systems that may increasingly shape those decisions. This creates what we see as a self-reinforcing dynamic: informal adoption expands because it appears useful; governance lags because the system continues to function. The longer this continues, the harder responsible correction becomes.
The learning paradox
The overwhelming majority of practitioners interviewed for this study had learned to use AI on their own initiative. Almost none described receiving structured organisational training; where support existed, it was limited to basic tool orientation rather than meaningful capability-building. As one participant summarised bluntly: “There is no training being provided. I think there is an assumption that everyone knows how to use it, which is not true. There is no support.”
And yet learning is happening. This creates a paradox. Humanitarian organisations are already benefiting from AI capability they did not intentionally build. In effect, they have outsourced AI literacy to individual workers, creating a fragile model of capability accumulation that depends on personal motivation, informal experimentation and knowledge that remains largely invisible to the institution. When that motivation fades – through burnout, departure or simple exhaustion – the knowledge disappears with it.
This does not mean competence is absent. Quite the opposite: the range of actual AI practice appears far wider than leadership may assume. However, in the same humanitarian ecosystem (and sometimes under the same organisational non-policy), AI is being used both for sophisticated, orchestrated workflows and for unchecked, context-lacking copy-and-paste output submitted without revision. The issue is therefore not simply uneven skill levels, but the absence of any shared understanding of what good practice looks like. Without common standards, vocabulary or feedback loops, organisations cannot distinguish meaningful innovation from poor use.
The consequences are significant. When learning remains entirely informal, capability becomes difficult to see. A practitioner who has spent months developing advanced workflows may appear indistinguishable from a colleague who occasionally uses a chatbot to draft emails. Organisations, as a result, design responses for the users they think they have (typically cautious beginners), rather than the much more complex reality beneath the surface.
The silence of the capable
If the previous section reveals a learning paradox, this one exposes why that paradox persists. Several practitioners described a professional culture in which openly using AI remains socially ambiguous, not because the technology is formally prohibited, but because its legitimacy remains contested. For some, the concern was not ethical discomfort or fear of job displacement, but social judgement. One practitioner explained: “We are still in a culture where people are valued more for the effort we think they have put in than for the result.” Another was even more direct: “I get the impression that using AI is like being lazy.”
This matters because the practitioners most capable of anchoring organisational learning may be precisely those least inclined to make their practices visible. If using AI risks undermining professional credibility, then competence becomes something to conceal rather than share. As a result, the expertise organisations most need to learn collectively remains socially hidden.
This dynamic has broader organisational consequences. Research from other sectors suggests that peer exchange is among the most effective ways to build meaningful AI capability.[4]Riya Sahni and Lydia B. Chilton, “Beyond Training: Social Dynamics of AI Adoption in Industry”, arXiv, 18 February 2025. Yet in humanitarian settings, this potential remains constrained. Informal knowledge-sharing does occur, but unevenly, invisibly and without institutional recognition. Rather than creating a culture where advanced practice can circulate and mature, organisations risk reproducing informal ecosystems in which average practices spread more easily than strong ones. The issue, then, is not simply a lack of technical training. It is the absence of social and organisational conditions that legitimise learning in the open.
The cost of efficiency
If there is one promise consistently associated with AI in humanitarian practice, it is efficiency. Practitioners across interviews described tangible gains in drafting, translation, summarisation, reporting, formatting and repetitive administrative work. In this sense, the productivity case for AI is real and it matters.
The humanitarian sector is currently operating under extraordinary pressure. Humanitarian needs continue to rise while funding shrinks, forcing organisations into increasingly difficult trade-offs between mission, compliance and survival.[5]For a broader discussion of the current humanitarian funding crisis and its systemic implications, see Pierre Micheletti, “2025: the humanitarian movement’s great depression?”, Humanitarian … Continue reading In such an environment, tools that reduce the administrative burden are understandably attractive. As one humanitarian director put it candidly: “If we weren’t in crisis, efficiency would be about giving more to beneficiaries. Right now, efficiency is about continuing to exist.” As seen, the promise of efficiency is no longer simply about doing better humanitarian work. In some cases, it is about organisational survival in a contracting system.
Yet this is precisely where caution is needed. Efficiency gains do not necessarily translate into transformation. Several practitioners questioned what time savings actually produce in practice. One respondent framed it starkly: “Does this mean I have more time for my family? […] To do sport? No. It means more time to work even more. […] I can clearly see that, in a way, it’s a headlong rush forward.”
If AI helps accelerate reporting, but reporting systems remain structurally excessive and disconnected from operational reality, what has changed? If AI reduces the burden of donor compliance, but leaves compliance architecture untouched, is this a transformation, or is it adaptation? If overstretched teams simply absorb productivity gains into ever-higher output expectations, AI may not reduce pressure. It may normalise it. The question is not whether AI saves time, because it clearly can. The question is whether those gains create space to rethink broken organisational systems or, rather, merely help humanitarian actors endure them for longer.
Governance as enablement: from efficiency to mission-oriented value creation
If the preceding sections describe a self-reinforcing cycle of informal adoption, silence and reactive efficiency, we come to ask ourselves: what changes when governance exists?
A limited number of organisations described efforts to move beyond unstructured AI adoption, not by treating AI simply as a technical tool to regulate, but as a strategic organisational capability requiring deliberate stewardship. In these cases, the shift was not driven by technological sophistication. It was driven by a recognition that inaction also carries risk. Governance meant creating visibility where uncertainty previously prevailed: clarifying acceptable use, identifying secure environments, supporting practitioner learning, creating internal tools and ensuring organisational permission for responsible experimentation.
However, governance should also mean reflecting on the fundamental purpose of AI adoption. As earlier sections showed, in a sector under severe operational strain, efficiency gains can easily be absorbed into a logic of survival – doing more reporting, processing more compliance and sustaining overstretched systems. Yet practitioners working under a more structured environment articulated a different ambition: using AI not simply to do existing work faster, but to redirect human time towards activities where humanitarian added value is greatest. These activities include critical analysis, contextual judgement, relationship-building, creative problem-solving and deeper engagement with affected communities.
One operational example illustrates what this can look like in practice. In Jordan, a practitioner described a pilot scheme in which AI was used to redesign community data collection rather than merely accelerate internal workflows. Affected communities could engage in one-to-one AI-supported interviews at a time that suited them, reducing the burden of conventional survey scheduling while improving accessibility and responsiveness. Human oversight remained essential for interpretation and contextual judgement, but the practitioner reported both faster data collection and stronger reporting output.
What makes this example significant is not the technology itself, but the conditions behind it: a bounded use case, sufficient practitioner literacy, human oversight and an explicit design choice to create operational value rather than simply internal efficiency. This is what governance makes possible.
Meaningful AI use does not emerge automatically from access to tools. It requires organisational conditions that allow thoughtful experimentation, shared learning and clear accountability. Governance, in this sense, is not a barrier to adoption. It is the enabling condition for responsible transformation.
Towards a new conversation
The humanitarian AI paradox can be described as a familiar story: technological change outpacing institutional adaptation. But the findings presented here suggest something more structurally troubling.
The sector appears trapped in a self-stabilising loop: individual adaptation is sufficient to keep the system functioning, which removes the urgency that might otherwise force collective response. Practitioners integrate AI into their work: sometimes thoughtfully, sometimes improvised, often invisibly. Output continues to be produced. Leaders do not necessarily perceive a crisis.
Meanwhile, the costs accumulate below the surface: knowledge that cannot be shared, quality that cannot be assured, data that is insufficiently protected, risks that remain poorly understood, and inequalities that quietly deepen. None of these necessarily appear as singular events. They emerge gradually, distributed across countless individual choices, until they become crises.
This is why governance cannot be understood only as an enabler of innovation. It is also a preventive imperative. Some risks are immediate operational concerns: security, ethical use and regulatory compliance. Others are broader but no less consequential: the environmental footprint of AI infrastructures, or the possibility that humanitarian AI reproduces existing geopolitical and epistemic power asymmetries rather than challenging them.
But governance is not only defensive. It also determines what kind of value AI creates. If AI merely helps humanitarian organisations survive by accelerating administrative overload, then its contribution may remain limited to adaptation within an already overstretched system. But if governance enables AI to redirect human time towards judgement, creativity, relationship-building and deeper engagement with affected communities, the stakes become far more transformative. Therefore the conversation must shift from efficiency to value: defining what we are trying to achieve with AI, and for whom.
The future of AI in the humanitarian sector has yet to be written. Whether this future is intentionally shaped by the humanitarian community or passively inherited will depend on the choices made today.
Author’s diligence statement: In the development of this qualitative study, we used generative AI to assist in interview transcription, thematic coding and text refinement. All AI-assisted content included in this work was rigorously reviewed, edited and curated by the authors. The final text reflects our own expertise, interpretations and intended meaning and we retain full responsibility for the content, its accuracy and its presentation. This statement is provided in the interest of transparency and to openly share how artificial intelligence technologies are used to support intellectual work.
Picture credit: Matheus Bertelli – 11949553456 @BertelliFotografia
