AI models still can't read a map as well as a cartographer would like.
Researchers built a benchmark of 5,760 choropleth maps and 28,800 questions, testing 21 open-source and proprietary multimodal models on tasks like identifying attributes, ranking regions, comparing values, and spotting patterns. They varied three design choices: hue palette, whether colors followed a proper light-to-dark sequence, and lightness contrast. Breaking the sequential color order hurt performance the most, especially on comparison and ranking questions. Low contrast also consistently made models worse, though pushing contrast higher past a certain point barely helped further.
Most AI cartography guidance borrows straight from human perception research, and this suggests that instinct is only partly right - hue itself mattered far less than expected. LoRA fine-tuning raised overall accuracy but didn't fix the underlying sensitivity to broken color logic; the models still stumbled on the same categories of maps. That's worth noting for anyone building products that expect a model to read a chart or map instead of a person.
The paper doesn't test whether newer training methods could close the gap, so treat AI-friendly cartography as a to-do list, not a solved problem.