Study finds country-local sources accounted for only 4.9 percent of ChatGPT citations on Ethiopia-related questions
Ethiopian institutions and media organisations remain largely invisible in generative artificial-intelligence answers about the country’s humanitarian crises, according to a new study examining how AI platforms select and cite evidence.
The study, Citation Localisation in Generative AI Answers: Whose Evidence Is Visible in African Humanitarian Crises?, analysed answers generated by Gemini, DeepSeek, ChatGPT and Perplexity in response to humanitarian questions covering 10 African countries, including Ethiopia.
In a complete 90-question ChatGPT analysis, country-local sources accounted for 7.2 percent of all citations. However, the share for Ethiopia was lower: local Ethiopian sources provided four of 81 citations, equivalent to 4.9 percent.
The findings raise questions about whose knowledge becomes visible when journalists, policymakers, donors, researchers and members of the public use AI systems to understand humanitarian conditions in Ethiopia.
The study’s broader cross-platform sample found that country-local producers supplied 5.5 percent of 1,083 answer-linked citations, while African-based organisations accounted for 6.6 percent. UN and multilateral organisations dominated the citation landscape, providing 68.8 percent of the citations.
The complete ChatGPT analysis showed a similar pattern, with UN and multilateral organisations accounting for 70.9 percent of citations.
The study found that AI-generated humanitarian answers typically rely on a small group of major international institutions. Across the full citation corpus, UNICEF received 136 citations, followed by UNHCR with 100, OCHA and Humanitarian Country Team products with 73, OCHA with 64 and IOM with 63.
The five most-cited producers accounted for 40.3 percent of all answer-linked citations, while the top 10 accounted for 58.8 percent.
For Ethiopia, this means that information from international organisations is more likely to appear in AI-generated responses than evidence produced by national ministries, local humanitarian organisations, Ethiopian research institutions or domestic media outlets.
The study cautions that the results do not prove that AI systems are intentionally biased against Ethiopian sources. It also does not claim that local evidence was unavailable. Instead, it measures what sources were visibly cited under a standardised research protocol.
A formal measure of under-citation would require a complete list of all relevant, available and retrievable Ethiopian sources for each question, the study said.
The research found that local sources rarely replaced international sources entirely. Instead, they generally appeared alongside UN or multilateral evidence.
Of the 186 cited observations in the recoverable cross-platform sample, 33 included at least one country-local source. Twenty-eight of those 33 responses—84.8 percent—also cited a UN or multilateral organisation.
Only one response relied exclusively on country-local sources.
The pattern was similar in the ChatGPT analysis. Local-citing answers contained a lower average share of UN citations than responses without local sources, suggesting that local evidence can diversify the information presented by AI systems even when international organisations remain central.
For Ethiopia, the inclusion of local sources could provide additional context on displacement, conflict, food insecurity, humanitarian access, public services and the work of national and community-based responders.
The study points to several reasons why Ethiopian sources may struggle to appear in AI-generated answers.
International organisations often have dedicated information-management teams, stable websites, standardised publication formats, clear metadata and global syndication networks. Ethiopian institutions and local organisations may possess detailed field knowledge but publish less frequently, use less searchable formats or distribute information through channels that AI systems retrieve less consistently.
Language is another factor. All questions in the study were submitted in English. This may favour Ethiopian organisations that publish in English while disadvantaging authoritative material available mainly in Amharic or other Ethiopian languages.
The study also found that intermediary platforms played a significant role in how humanitarian evidence became visible. ReliefWeb hosted 217 of the answer-linked citations, accounting for 20 percent of the entire corpus and 93.9 percent of intermediary-hosted citations.
A domain-only analysis could therefore incorrectly identify ReliefWeb as the original producer of information created by OCHA, UN agencies, Ethiopian institutions, NGOs or other humanitarian actors.
The findings suggest that Ethiopian humanitarian organisations need to invest not only in collecting information, but also in making that information discoverable and attributable.
The study recommends publishing evidence through durable web addresses, accessible HTML pages, consistent organisational names, clear dates, identifiable authorship and machine-readable formats. Such measures could help AI systems and other digital platforms retrieve and attribute Ethiopian evidence more accurately.
International organisations and repositories also have a role to play. When reports are produced with Ethiopian partners or based on data collected by national institutions, metadata and publication pages should clearly preserve the original contributors’ identity.
