Share & more
“Is it legal for the US to strike this target?”
“Where should displaced families seek assistance?”
“Who started the conflict?”
These are no longer questions reserved for diplomats, aid workers, or journalists. Increasingly, they’re being posed to chatbots. In fragile states where institutions are fractured and official channels unreliable, large language models (LLMs) are quietly becoming part of the information infrastructure. Their responses, shaped by opaque training data and shifting system updates, can influence what people believe, how they move, and whom they trust.
As generative AI becomes increasingly integrated into the digital infrastructure of everyday life, its reach is extending into places they were never explicitly designed for – fragile and conflict-affected states, regions of proxy warfare, and humanitarian crisis zones. The technology is even showing up in countries that lack the capacity to build or train these models embedded in translation tools, search engines, and community knowledge hubs. Increasingly, AI tools are developed on behalf of aid organisations hoping for efficiency gains, or simply drawn to the promise of new tech.
This technology doesn’t arrive neutral. It brings with it biases, blind spots, and narrative preferences shaped by the data it’s trained on and the institutions that develop it. And in volatile environments, those narrative shifts matter.
What happens when AI-generated responses begin to shape how conflicts are understood? What happens when the models give subtly different answers depending on the user’s language, the phrasing of a prompt, or the geopolitical leanings encoded in their training data?
This is model drift, and it’s a real humanitarian risk.
I borrow the term from machine learning, where it describes a loss of predictive accuracy when real-world data conditions change. In technical fields, model drift typically refers to a drop in predictive accuracy when the data a model sees in the real world changes over time. However, in fragile states, the concern isn’t just statistical accuracy – it’s narrative volatility.
LLM outputs can shift in subtle or strategic ways during fast-moving crises: A model’s description of an airstrike, a refugee movement, or a foreign intervention might change depending on the language used, the political framing of the prompt, or the timing of the query. But here, I use it to refer to the shifting behaviour of language models as they respond to complex, politically sensitive situations. It’s not just about factual accuracy. It’s about narrative instability: the way AI systems describe events, assign blame, suggest responses, or omit key actors.
In conflict zones and fragile states, truth is already contested. Model drift adds another layer of ambiguity. And because language models are designed to sound authoritative, their outputs often carry an implicit legitimacy that can confuse users, reinforce dangerous narratives, or distort public perception of humanitarian efforts.
Consider a displaced family in the Sahel asking an AI tool where to seek assistance. If one model lists an aid organisation affiliated with a Western power, while another flags it as biased or unsafe, who do they trust? Or think of a journalist in Ethiopia querying about responsibility for a recent atrocity. Does the model provide the same framing in Amharic as it does in English? If not, why not?
Rather than treating model drift as a purely technical problem, we should recognise it as a humanitarian communication challenge. And we should bring the insights and infrastructure of humanitarian diplomacy to bear on AI governance.
These aren’t hypothetical risks. They are already emerging as generative AI tools are quietly adopted in fragile information ecosystems. And unlike official public statements or verified databases, LLMs are designed to be conversational, fluid, and adaptive – qualities that make them appealing, but also unpredictable.
Humanitarian actors know what happens when communication breaks down during crises. Rumours spread. Aid is blocked. Lives are lost. Over the past two decades, organisations have developed rapid-response tools to manage these risks: rumour-monitoring systems, multilingual messaging campaigns, trusted messenger networks, and real-time feedback loops with communities on the ground. These tools don’t just transmit information – they build trust, maintain coherence, and prevent harm.
Yet in the growing conversation around AI safety and model governance, these field-tested strategies are almost entirely absent. Alignment is discussed in terms of machine ethics, policy compliance, or content filters – but not in terms of how people in crisis settings actually receive and respond to information.
There’s a missed opportunity here.
Rather than treating model drift as a purely technical problem, we should recognise it as a humanitarian communication challenge. And we should bring the insights and infrastructure of humanitarian diplomacy to bear on AI governance – especially in the regions most likely to suffer from misinformation, geopolitical distortion, and underregulated digital deployments.
Drawing on my own experience in humanitarian diplomacy and operational research in fragile contexts, I believe these systems can inform AI safety frameworks in three ways:
- First, by offering tools for narrative consistency across languages and platforms, which is critical when multiple agencies rely on shared messaging.
- Second, by developing community-centered mechanisms for detecting and correcting misalignment in real time, using existing trusted messenger models.
- And third, by advocating for localised oversight and participation in how AI tools are deployed, even when those tools originate elsewhere.
In Syria, for example, international organisations have already deployed AI-powered systems like Hala Sentry to predict airstrikes, and SKAI to assess damage via satellite imagery. These tools are lifesaving – but they are also managed externally, without long-term integration into local systems. The same risks apply to LLMs. Without mechanisms for local review, accountability, and redress, AI deployments risk repeating a familiar pattern: innovation without consent, infrastructure without ownership, and technology that reinforces dependency rather than resilience.
The most important contribution AI can make to crisis response may not be through high-tech interventions or drone-enabled logistics. It may come from something simpler but more fundamental: not making things worse.
This isn’t just an ethical issue; it’s a security one. When LLMs are inconsistent in how they frame conflicts, describe alliances, or suggest actions, they can erode the diplomatic narratives that humanitarian actors and governments depend on. They can be exploited to spread disinformation, sow mistrust, or polarise vulnerable communities.
Model drift matters and needs to be addressed – not after the next crisis, but before.
This requires interdisciplinary collaboration. Humanitarian responders, AI researchers, and security analysts must work together to define what alignment looks like in fragile contexts. This means designing AI systems not just for fluency or safety in abstract terms, but for reliability, cultural sensitivity, and policy coherence during real-world emergencies.
It also means shifting the way we talk about “AI for good”. The most important contribution AI can make to crisis response may not be through high-tech interventions or drone-enabled logistics. It may come from something simpler but more fundamental: not making things worse.
Model drift isn’t inevitable. It can be anticipated, detected, and corrected. But only if we recognise it for what it is: a humanitarian risk, not just a technical glitch.
