Understand

Artificial intelligence is steadily changing how the world responds to crises.

Governments use it to analyse satellite imagery after natural disasters, humanitarian agencies are testing AI to anticipate displacement before it overwhelms fragile communities, and researchers are applying machine learning to uncover patterns buried inside millions of records that no human could process alone.

These advances, while useful, expose a problem worthy of scrutiny.

Most AI systems are built on data that tells only part of the world’s story. Since they are trained largely on information from North America, Europe, and parts of Asia, they know relatively little about the realities of conflict in parts of Nigeria like Borno, Zamfara, Katsina, Diffa or the wider Lake Chad Basin. They struggle with Hausa, Kanuri, Fulfulde, Igbo and hundreds of other African languages. In addition, they often miss the informal power structures, survival economies, historical grievances, and social relationships that determine whether a community remains stable or slips into violence.

For example, despite advancements in AI content flagging on social media platforms, HumAngle’s reporting and research have consistently found that moderation systems that perform effectively in English often fail to detect incitement expressed in Hausa, Kanuri, Igbo, and other Nigerian languages. Language models frequently miss local expressions, figurative speech and cultural nuances, allowing harmful content to evade automated moderation. That is not simply a technical limitation. It is a human security challenge.

Africa has long faced a deficit in research, reporting, documentation, and structured knowledge about local realities. That gap is becoming one of the continent’s biggest constraints in the AI era. Large language models (LLMs) can only learn from information that exists. If communities, conflicts, languages, histories or ecosystems remain poorly documented, those blind spots are carried directly into the systems built on them. Building culturally informed and representative AI therefore begins long before model development. It begins with producing more reliable knowledge.

Kunle Adebajo, editor of Code for Africa’s African Academy for Open Source Investigations (AAOSI), has this to say: “Artificial intelligence companies have successfully scraped most of what exists on the Internet to train their models. But a vast amount of data lives outside of the shores of the Internet, especially in Africa. AI systems, for example, will struggle with geolocation because many parts of Nigeria are still not properly mapped. They will struggle with translating languages spoken by minority ethnic groups. Even languages spoken by tens of millions of people, like Amharic, are considered low-resource by AI companies and therefore aren’t prioritised, let alone those spoken by only a few million. There’s, of course, also the lived experiences of conflict victims, which aren’t reported online and which you can only understand through fresh field interviews. Even when all this information is integrated into the AI systems, we still have the risk of hallucinations. In matters of security, we cannot take any of these shortcomings and risks lightly, because lives are at stake. Security actors who are incorporating AI tools into their operations must do their due diligence and maintain strong human oversight over the processes.” 

The decline in funding for research and investigative reporting therefore comes at a critical moment. Every research project left unfunded, every environmental change left undocumented, and every local story that never enters the public record represents knowledge that remains unavailable not only to policymakers and researchers today but also to the AI systems increasingly shaping future analysis. The quality of future AI will depend, in part, on the quality and diversity of the knowledge produced today.

If the data is incomplete, the analysis will be incomplete. If the context is missing, the conclusions may be wrong. In fragile environments, inaccurate conclusions can delay humanitarian assistance, distort policy responses and, in the worst cases, cost lives.

Communities see the signals first

Conflict rarely arrives without warning, and the warning signs are rarely dramatic. They are usually ordinary changes that occur over time, not all at once. A weekly market suddenly stops operating, or fishermen abandon parts of Lake Chad. Parents stop sending children to school weeks before an attack while farmlands expand into traditional grazing routes, or herders begin avoiding ancestral migration paths that have remained open for generations. HumAngle’s reporting has repeatedly documented these patterns.

Months before major offensives around the Lake Chad Basin, local communities described unusual restrictions on fishing, increased taxation by armed groups after governments introduced ransom-payment restrictions, shifting movement routes, and growing pressure on traders. These were not isolated observations. They formed part of a broader pattern.

In northwestern Nigeria, our reporting has shown how armed groups shifted from pickup vans to motorcycles when security operations disrupted their mobility. We documented how kidnapping networks increasingly demanded ransom in CFA francs after Nigeria’s cashless policy disrupted access to large amounts of naira. We traced the emergence of new ungoverned forest frontiers, the expansion of illegal mining sites and the growing links between critical minerals, transnational migration and armed violence. These adaptations rarely appear in conventional datasets, yet they are recognised by communities living with insecurity every day.

Researchers increasingly argue that AI systems become more reliable when such verified local observations are analysed alongside conventional datasets, helping analysts identify patterns that would otherwise remain fragmented. 

Joshua Olufemi, Founder of Dataphyte, a data and research analytics company, believes that agenda-setting is critical to this. Speaking to HumAngle about some practical pathways, he mentioned “Multistakeholder agenda setting and strategic partnerships on critical infrastructure and local data modelling and innovation. ⁠Newsroom Viability and business case initiatives to build and sustain the ecosystem.”

Building AI that understands fragile societies

The proposed HumAngle AI and Human Security Lab is designed to explore how verified field reporting, historical conflict data, humanitarian information, environmental indicators, satellite imagery, where appropriate, and open-source intelligence can be organised to identify areas where risk is changing.

The objective is not to predict the exact time or location of an attack. Conflict is too complex for deterministic prediction, and responsible AI requires transparency about uncertainty. Instead, the approach focuses on identifying shifts in risk while ensuring that assessments remain explainable, transparent, and subject to human review.

For humanitarian organisations, even a few days of additional warning can allow food supplies to be prepositioned before roads become inaccessible, schools to strengthen protection measures before attacks occur, vulnerable communities to relocate before violence escalates or emergency resources to reach areas before displacement intensifies.

Journalism is structured evidence

Journalism is often viewed as a record of past events. Increasingly, it also provides structured evidence that can help explain how crises evolve over time.

For more than six years, HumAngle has reported on insurgency, communal violence, kidnapping, humanitarian crises and governance failures across Nigeria and the Lake Chad region.

Its investigations have exposed mass arbitrary detentions, documented attacks on schools, mapped displacement, tracked ransom economies and examined the evolution of Boko Haram, ISWAP and other armed groups.

Reporting on school attacks has gone beyond casualty figures to examine how fear empties classrooms, deepens inequality and reshapes communities. Investigations into kidnapping networks have documented how criminal economies evolve, how ransom systems adapt and how insecurity expands into areas once considered relatively stable. Each story contains verified interviews, timelines, locations, local terminology and community knowledge. They represent one of Africa’s richest public interest records documenting how insecurity has evolved across the Lake Chad Basin and the wider Sahel.

Adebajo argues that this type of journalism becomes even more valuable when its evidence is intentionally structured for computational analysis.

“Credible investigative journalism products and community knowledge can be used to train AI models that provide intelligence for crisis signalling and data-driven predictive research. But it has to be structured. The good thing is that AI tools can help with that structuring, so it doesn’t consume as much time. However, collecting information for AI training and collecting information for Investigative reporting are not the same thing.”

According to the investigative journalist and AI researcher at Code for Africa, “if you set your mind on achieving both objectives, it’ll also inform the kind of data you collect and how you standardise the process and schema across different scenarios. When you have such a big dataset from years of journalism, you can combine it with other security, satellite, and reporting datasets to draw out patterns and outliers. You can determine which areas are more susceptible to attacks and multiple displacements, which seasons, what time of day, etc.”

Beyond the statistics of political violence, you can also draw out patterns in the nature of the attacks and the experiences of victims. All these can help with better allocation of scarce resources in tackling the problem of insecurity, Adebajo argued.

Kofoworola Odozi, a media analyst and AI hobbyist, agrees. She believes deliberate archiving is central to this work. “Newsrooms and CSOs have to build a digital library that has local information and that they completely own and control. What this would do for us is that instead of relying on data that big tech companies gather, we collect more materials like notes from townhalls, local newspapers, budget hearing notes–things like that that you would not typically find in the data that big tech companies have. CSOs and newsrooms can gather that and use open source to document them.”

The media analyst also believes that media and CSOs have to move to a unified digital infrastructure. “Media and CSOs are very individualistic in the ways we work, even though we are all working towards the same thing. We should work towards an industry-wide consortium.”

Adebajo says there are some strategies working at Code for Africa.

“One thing we’re doing at Code for Africa in our fight against information disorder is to go beyond fact-checking to train people who edit Wikipedia articles and train AI models, so that we can put the fact-checks into a pipeline that feeds into quality AI data, because we understand that more and more people are getting their information directly from these models,” he said. “That’s another thing to consider when reporting on conflict, especially in the fight against misinformation and conspiracy theories. How do we get these dozens and hundreds of reports to be assimilated reliably into AI systems, which are not only disseminating information directly to people but are also feeding indirectly into other knowledge production mechanisms?”

Context matters more than computing power

Much of the AI conversation focuses on larger models, faster processors, and bigger datasets. Yet model performance ultimately depends on the quality and relevance of the information used to train it.

A global AI system may recognise the word “flood.” It may identify references to drought, migration or armed conflict. Yet it may completely miss why a cash policy disrupted kidnapping networks, why changes in seasonal border patrols altered irregular migration routes or why the discovery of lithium, gold, and other strategic minerals reshaped local security dynamics.

It may recognise extremist vocabulary without understanding how concepts such as hijrah, bay’ah or takfir carry different meanings depending on who is using them, where they are used, and the historical context in which they are invoked.

It may classify a dispute as ethnic when local communities understand it primarily as competition over grazing routes, irrigation access, taxation or political representation.

Without local knowledge, AI systems risk misunderstanding complex societies despite increasing computational power.

The next battlefield is information

Conflict no longer unfolds only on the ground. It also unfolds across phones, messaging platforms, and social media.

False rumours spread within minutes as old videos are recycled as evidence of new attacks. AI-generated images and synthetic audio are becoming increasingly difficult to distinguish from authentic documentation, and armed groups continue adapting their propaganda while misinformation and disinformation deepen mistrust between communities and weaken confidence in legitimate institutions.

HumAngle has documented how Boko Haram and ISWAP compete not only through violence but also through narratives that seek legitimacy, exploit grievances and shape public perceptions of governance, justice and religion.

Understanding these information ecosystems is becoming increasingly important for humanitarian response and conflict analysis.

The proposed HumAngle AI and Human Security Lab also explores how publicly available information can be analysed to better understand AI-enabled misinformation, extremist narratives, and digital conflict across African contexts, while respecting privacy, freedom of expression, and the principles of independent journalism.

Human judgement must remain at the centre

The growing use of AI in humanitarian and security settings has been accompanied by claims that algorithms will predict wars or replace analysts. Current evidence suggests a more limited role.

AI can support human decision-making by organising large volumes of information, identifying patterns and highlighting anomalies. Human judgement remains essential for interpreting those findings, questioning assumptions, and understanding context. Transparency about uncertainty, limitations, and data quality remains fundamental in fragile environments where analytical errors can carry significant consequences.

Africa and the future of humanitarian AI

The debate over AI in Africa has increasingly shifted from digital access to questions of data, governance, and representation.

Africa possesses extensive community knowledge, linguistic diversity, and historical experience that remain underrepresented in many AI systems. As governments, humanitarian organisations and researchers expand the use of AI in crisis response, these forms of knowledge are becoming increasingly important sources of evidence.

The proposed HumAngle AI and Human Security Lab’s focus is not AI for its own sake. It is the development of analytical tools that help humanitarian organisations act earlier, journalists identify emerging patterns, researchers organise complex evidence and policymakers make better-informed decisions.

If these approaches prove effective, AI may contribute to more timely humanitarian action, stronger evidence-based policymaking and improved protection for vulnerable communities.