A new study has raised questions about possible bias in how major artificial intelligence platforms select and present sources when explaining African humanitarian crises.
The African Humanitarian AI Citation Index (AHACI) analysed 1,083 verifiable citations used by ChatGPT, Gemini, DeepSeek and Perplexity when answering humanitarian questions about ten African countries.
Only 5.5 per cent of the citations came from organisations based in the country being discussed, while68.8% came from United Nations agencies or other multilateral organisations.
“The crisis may be local, but the visible authority explaining it to AI users is usually international,” said Hezron Ochiel, Founder of Hezron Insights and researcher behind the study.
The findings do not prove that AI platforms deliberately discriminate against African sources. They raise a different concern: could AI be reproducing existing inequalities in which some sources are much easier to find, verify and cite than others?
Kenya stands out
Kenya performed very differently from most of the countries studied.
Of the 60 country-local citations identified across the ten-country dataset, 31 came from Kenyan sources, meaning Kenya accounted for more than half of all local citations found.
Kenyan media organisations, humanitarian institutions and public agencies were among the sources appearing in AI-generated answers.
Nigeria also recorded relatively stronger local visibility, while the Democratic Republic of the Congo and Sudan recorded no country-local citations in their recoverable datasets.
The study does not establish why Kenya performed better. Language, publishing practices, website authority, search visibility and technical accessibility may all influence whether local information is easy for AI systems to discover and attribute.
Is AI inheriting the bias of the web?
Large international organisations often have established websites, extensive archives, dedicated publishing teams and strong institutional authority.
Local organisations may produce valuable evidence from affected communities but have fewer resources to publish and distribute it at the same scale.
AI systems operate within this unequal information environment. If they repeatedly surface organisations that are already highly visible online, they may reinforce existing patterns of authority.
In that sense, AI may be inheriting the biases of the information environment it searches.
The study found that the five most-cited knowledge producers accounted for 40.3 per cent of all citations, while the top ten accounted for 58.8 per cent.
UNICEF was cited most frequently, followed by UNHCR, OCHA/Humanitarian Country Team products, OCHA and the International Organisation for Migration.
A separate analysis of ChatGPT showed the same broad pattern. Across all 90 production questions, 7.2% of citations were country-local while 70.9% came from UN or multilateral organisations.
For Kenya, the findings are both encouraging and cautionary. They show that African organisations can become visible in AI-generated answers, but they also suggest that producing useful knowledge is no longer enough. That knowledge must also be easy to discover, verify and attribute.
Humanitarian localisation has long asked who receives funding, delivers programmes and participates in decisions.
AI introduces another question: Who gets to explain the crisis?
The full findings are available through the African Humanitarian AI Citation Index.
