OpenAI has introduced an improved version of GPT-4o, its AI image generation system, which promises more lifelike visual outputs. Developed through extensive collaboration with human trainers over the past year, this new model will now serve as the primary image generation engine for ChatGPT, replacing DALL-E 3. The updated capability is accessible to users across all ChatGPT subscription tiers, including Free, Plus, Team, and Pro.
So while the world is busy turning their favorite family photos into scenes from South Park, I am asking ChatGPT to generate maps. Every few months I go through the exercise of generating AI images of relief maps, to see how close I am to losing my job to the robots. So far, I have been unimpressed by the outputs.
DALL·E 3 Image Generation - New York State Shaded Relief Map
The old model (sort of) knows what a relief map is, but the generated topography is random and the results don't make any geographic sense. The above map shows an attempt by DALL·E 3 to make a relief map of New York State. The state boundary is mostly correct, but the topography is completely made up.
Fast forward to GPT-4o, it looks like OpenAI has started to catch up on its cartography skills. When I read the news this morning about the new GPT-4o image generation, I immediately asked ChatGPT to create a shaded relief map of New York State. To my great surprise, it created an incredibly realistic image.
GPT‑4o Image Generation - New York State Shaded Relief Map
The resulting map was genuinely impressive. You can clearly see the correct placement of the Adirondack Mountains, Tug Hill Plateau, Mohawk Valley, Finger Lakes, Hudson River, and so on. And the continuity between the regions is pretty much seamless. If you look too closely, however, it starts to get a little sloppy. But broad strokes, this is a solid representation of the topography of New York State.
After this impressive result, I absolutely had to test the limits and see where this all breaks.
GPT-4o seems to be good at generating accurate(ish) topography for large areas, such as whole countries or continents:
GPT‑4o Image Generation - Conterminous United States Shaded Relief Map
GPT‑4o Image Generation - Africa Shaded Relief Map
GPT‑4o Image Generation - India Shaded Relief Map (note the boundaries - no input given on conflicted borders)
Again, not perfect, but absolutely recognizable as realistic relief maps with generally accurate topography. This is way beyond the capabilities of Dalle 3.
Where things start to break is when you ask for maps of smaller, less known regions:
GPT‑4o Image Generation - Onondaga County Shaded Relief Map
This one was a big miss. It sort of understands this county has a lake and highways and might be a little hilly. The bounding borders are also completely incorrect. Here is one of our maps of Onondaga County for comparison.
There is a lot more exploring to do with maps of topography, but how about other styles of maps? Here is an example of a vector style map, where accuracy and readability are the only important factors:
GPT‑4o Image Generation - NYC Subway Map
This one has the general feelings of a subway map, and if you squint, you can tell its New York City, but overall not useful. Unless you are trying to catch a train from Brooky to Hunterion Averias.
Here is an attempt at a demographic map that requires actual census data:
GPT‑4o Image Generation - Population Density of California.
I don’t know where it was going with that one. It seems to have identified 34°N, 118°W as a spot of high population density which decreases radially outwards. (34°N, 118°W is the center of Los Angeles, turns out). Not a recognizable map of California.
Maybe a hydrologic map will be an easy task:
GPT‑4o Image Generation - Hydrologic Map of Alberta.
Another big miss, looks cool though! The general hydrologic patterns in Alberta are dictated by the Rocky Mountains, which make a lot of the big rivers flow towards the Northeast.
Last one, how about satellite images?
GPT‑4o Image Generation - The UK from Space
This one is fairly recognizable as the United Kingdom, but for some reason, we can see stars through the planet and the radius of the Earth is off by a factor of 10.
What does it all mean?
As we know, AI models like this are trained on images from across the internet. What we’re seeing here is essentially a complex digital blend of raster maps that were created and posted online (a big oversimplification, but you get the idea). No actual GIS data is being scraped or used to improve the accuracy of the images.
To be fair to ChatGPT, there were a few moments where the AI tried to redirect me to a web search for a more accurate map, admitting it didn’t have access to the data needed to generate one. A noble attempt, which I bypassed by typing "do it anyway."
The most impressive results were the shaded relief maps. In most cases, the topography looked incredibly realistic. Again, the model doesn’t have access to elevation data, it’s just learned what a relief map should look like from millions of online examples. That explains why the map of Onondaga County wasn’t great. There just aren’t as many reference images of that area compared to more popular regions like entire states or countries.
So, are cartographers and GIS professionals out of a job anytime soon? I think we’re safe, at least until the next OpenAI update.
Do you use AI tools in your mapmaking career? Let us know your thoughts in the comments below.