At the AIDS 2026 conference in Rio de Janeiro, the U.S. Department of State projected a slide titled Current APS Funding Opportunities, intended to show available funding for six African countries: Nigeria, Mozambique, Uganda, Côte d’Ivoire, Malawi and Cameroon. The problem was that none of the six countries was represented correctly. Nigeria, a country with an Atlantic coastline, had become a large landlocked shape in the Sahara; Mozambique had been moved from southeastern Africa to the Horn of Africa; Côte d’Ivoire had crossed from the west coast to the other side of the continent; Uganda and Malawi appeared approximately in the right region but had invented shapes and borders, while Cameroon appeared in the legend without its line pointing to any country. Reuters analysed the image and identified an AI watermark indicating the use of OpenAI tools. The State Department apologised and explained that a team member had modified the slide in haste immediately before the presentation. AI had therefore achieved a geographical feat that would be difficult to match: it was given six countries and failed to place a single one correctly.

How can AI know where Nigeria is and still draw it incorrectly? Because an image generator does not operate by default like a GIS, or Geographic Information System, in which Nigeria is a polygon defined through exact coordinates, borders and neighbouring countries. The model tries to generate a visually plausible representation of the concept “map of Africa,” and this creates two distinct technical problems. The first is spatial or topological consistency: the system must simultaneously preserve the shape of the continent, the position of each country, its neighbours, coastline, proportions and relative distances. The second is text-image binding: the model must correctly connect the label “NIGERIA” to the exact shape and position of Nigeria, “MOZAMBIQUE” to another polygon, and so on. OpenAI itself lists binding problems and precise graphic representation among the limitations of image models, while visual systems can struggle with exact spatial localisation. The professional solution is therefore not to ask AI to “invent” the map. Its geometry should be produced from verified GIS or vector data, after which AI can be used for design, styling and explanations.