On 30th July 2026 the US State Department gave a presentation at an international development conference in Brazil. It included a map showing various African countries that it had the potential to invest in. This map quickly attracted worldwide attention as it was photographed by some delegates at the conference and sent to various media outlets. The map appeared to be AI-generated (reportedly by an OpenAI tool based on the watermark) and was grotesquely error-strewn, with the countries in completely the wrong locations. I don’t have copyright to the original image, but this can be seen in this Guardian article. The map went viral on social media partly due to the sheer number of obvious errors, and because the US State Department officials who had presented it clearly had no idea that it was wrong. This does not inspire confidence in their knowledge of geography, despite the fact that developing foreign policy is literally their job.
LLMs are fluent at language but often struggle with seemingly simple tasks outside of the context of language. For example, they often struggle with arithmetic, despite being able to handle complex mathematics (much maths theory involves pattern recognition, which LLMs are very good at). They also are terrible at chess. It seems that spatial recognition is also a limitation. Rather than start with a published map of Africa and draw arrows to marked countries, LLMs work differently. An LLM image generator synthesises a new image based on patterns learned during training. Their ability to reproduce precise geographical relationships appears to be much weaker than their language capabilities. This illustrates that image generation and spatial reasoning remain weaker than language generation in LLMs.
I was intrigued as to whether this map incident was representative of the state of the art in AI in August 2026, so I developed a simple experiment. I wrote a prompt and ran this through a number of AIs. The prompt was:
“Could you draw a map of Africa with a text panel to one side, with six specific countries named in the panel, and arrows pointing to their location on the map? The countries are Nigeria, Cameroon, Uganda, Mozambique, Malawi and Cote d’Ivoire.”
I then tried this in a number of AIs. The results were mixed.
To begin with, this was the attempt of ChatGPT, which was the AI used by the US State Department. As can be seen, there are many errors, with countries in incorrect positions, and even the arrows in strange places.
Produced by ChatGPT
Next I tried Grok. This was quite a mess. The countries are in wildly incorrect positions, and even the arrows are inconsistent, or even missing entirely in the case of Malawi.
Produced by Grok
Claude’s attempt was also fairly basic, with the map of Africa misshapen, and Mozambique shown as being somewhere in the Indian Ocean.
Produced by Claude
Kimi K3 took quite a long time, and then produced this image. The outline of Africa is crude, and the borders are not marked, but at least the countries are roughly in the right relative places.
Produced by Kimi K3
DeepSeek cannot produce images, but did produce a lengthy and detailed prompt that could be loaded into a specialist image generating AI like Dall-E or MidJourney. This was what was produced by Dall-E in response to the DeepSeek prompt.
Produced by Dall-E
This map also has major errors, with the countries badly misplaced, and arrows also oddly placed.
Mistral failed to produce anything at all. Easily the best was that from Gemini.
Produced by Gemini
This map has countries in the right place, and arrows pointing correctly. The borders are correct, and indeed I could find little wrong with it.
It is interesting just how much the various models vary in quality. In most general AI benchmarks, ChatGPT and Claude battle it out at the top, but in this specific case the easy winner was Google’s Gemini.
This illustrates that the outputs of the different models can, even today, vary drastically. Whether this reflects differences in training data, image-generation architecture or Google’s long experience with geospatial information is unclear, but it demonstrates that individual AI models can possess surprisingly different strengths. LLMs do not possess an internal map of Africa in the way humans do. Instead, they have linguistic knowledge of the countries, approximate statistical relationships, but no precise geometric representation. It also reinforces the message that you should not blindly trust LLM output, especially if you are about to present it in public. As AI systems increasingly move beyond text into images, video and robotics, spatial reasoning is likely to become a much more important capability than it has been for today’s predominantly language-based models.
