Of the 33 AI answers in my African Humanitarian AI Citation Index that cited an organisation based in the country they were discussing, 28 also cited a United Nations agency or another multilateral organisation. Only one answer cited sources solely from the country it discussed. The study covered humanitarian crises in ten African countries.
When an answer cites both kinds of organisation, readers have more than one place to look for evidence. A local report might cover one area of the country or a particular group affected by the crisis. If you want to quote a figure, the name of the organisation alone tells you very little about what the figure covers.
What the platforms showed readers
I began the Citation Index to find out whose work a reader could reach from an AI answer. I ran the same questions about humanitarian issues, including food insecurity and displacement, on ChatGPT, Gemini, DeepSeek and Perplexity. For each citation I could verify, I recorded where the organisation behind the source was based and what kind of organisation it was.
Of the 1,083 verifiable citations, 60, or 5.5 per cent, pointed to organisations based in the country being discussed. UN agencies and other multilateral organisations accounted for 745, or 68.8 per cent. Those percentages are shares of all the citations I could verify. To see how often local and multilateral sources appeared together, I counted answers instead. An answer with three local citations counted once, just as an answer with one local citation did.
In the study’s country breakdown, Kenyan sources accounted for 31 of the 60 country-local citations. In the records I could recover for the Democratic Republic of the Congo and Sudan, I found no citations to organisations based in the country being discussed. In this study, a Kenyan organisation counted as local in an answer about Kenya. The classification followed the organisation behind the cited document. A report from an international agency might still include a survey carried out by a local organisation. Finding out who collected the data would mean reading the report and tracing its sources. I also had no inventory of all the relevant local reports available for each question, so the citation share cannot tell us how often platforms passed over a source they could have cited.
A reader can check something more specific. In a report that covers several populations, which one does the quoted figure actually describe?
Two forecasts in one Kenyan report
In September 2026, after I ran the AI tests, the Integrated Food Security Phase Classification released its food-security assessment for Kenya with separate projections for the drylands and for refugees. The assessment estimated that 2.7 million people in Kenya’s arid and semi-arid lands faced crisis-level acute food insecurity or worse from July to October 2026. It projected that the figure would fall to 2.1 million between November 2026 and January 2027. For the refugee population it assessed, the numbers went in the opposite direction, from an estimated 351,000 to a projected 430,000 over the same periods.
An AI summary could link to the report and say, “Food insecurity in Kenya was expected to ease.” I wrote that example sentence for this article. The assessment projected improvement in the drylands as the short rains strengthened. It also projected that flooding could worsen conditions for refugees. Both projections appear on the same page.
A Kenyan diaspora association deciding where to send the proceeds of a fundraiser might read that summary and turn its attention elsewhere. If its members read the assessment, they could ask the organisations they support whether conditions among refugees were worsening as projected.
Before you quote the figure
Before you put an AI-supplied figure in a briefing, find it in the original report. If the answer lists its sources at the end, look for the one that leads to the passage or table containing the figure. Read the column heading and any notes beneath the table. A table might put the estimate for the assessment period in one column and the projection in the next. Once you lift a number out of the second column, the wording in your briefing has to carry the information the heading supplied. If your briefing is about current conditions, check for an assessment that describes conditions during that period. The latest report you find may still be projecting the period you need.
If you quote the 2.1 million figure, say that it was a projection for Kenya’s drylands from November 2026 to January 2027. If you’re summarising the whole report, include the outlook for the refugees it assessed. Otherwise, a reader could take the expected improvement in the drylands as an improvement for both groups.
Methodology note. On 23 August 2026, I ran 90 English-language questions on four platforms across ten African countries, for 360 question-and-platform observations. I analysed 1,083 verifiable citations. Some citation records could not be recovered. The September Kenya example comes from a separate published assessment.
Hezron Ochiel is a Nairobi-based researcher and the founder of Hezron Insights. He studies how AI search represents African evidence and expertise.





