ChatGPT is probably best known for helping students cheat on their term papers, helping internet comedians write knock-off Hallmark Christmas movies, and, yes, guzzling water. It turns out, though, that planners could use it and similar types of AI technology to make our cities more walkable, bikable, and just plan people friendly.
A pair of researchers at MIT and Virginia Tech recently found that two of the most prominent Large Language Learning Models – ChatGPT 4.0 and Gemini 1.5 — were able to perform many of the tasks of a typical built environment "audit" with a surprisingly high degree of accuracy, correctly identifying the presence of street trees, bikes, and benches more than 90 percent of the time within a given sample of Google Maps images.
Those programs didn't do as well at clocking things like sidewalks and steetlights, and both were significantly worse at evaluating rural areas than urban ones. The researchers argue, though, that even an inaccurate audit can help transportation officials do their jobs better — because right now, few U.S. cities regularly track the comprehensive condition of their streets at all. The ones that do, the researchers say, tend to rely on either painstaking and expensive manual audits, or virtual models that require planners to learn advanced algorithms and obtain pricy equipment.
If easy-to-use programs like ChatGPT can do a passable job, though, no city can claim that it simply has no idea where sidewalks, bus shelters, or trash cans are — and can get to work on actually replacing the ones that are missing.
"This kind of AI tool can increase the productivity of the planners, so they can spend less time in this type of work [and] more time holding town hall meetings," said Kee Moon Jang, a postdoctoral associate at MIT Senseable City Lab and co-author of the paper. "They can focus more on enhancing communication with their citizens. ... Or at least that's our hope."
The researchers argue that Large Language Learning Models can be particularly useful for small- and mid-sized cities that can't afford armies of auditors to roam their streets, or specialized experts to run Python codes on high-performance computers. And done right, said co-author Jungwhan Kim of Virginia Tech, they might even help "democratize" access to roadway information that poorer communities typically can't get at the granular pedestrian scale.
Still, Kim cautioned against cities over-relying on AI, which occasionally "hallucinates" objects that don't exist and makes assumptions about gaps in the data that generate inaccurate results. He suspects that has to do with the kind of data on which Large Language Learning Models are trained, which could pose a problem as the tool scales to more places.
"AI is a black box; we don't know what is actually going on inside," added Kim. "But one [area of] speculation is the disparity in training data. We have another research paper that focuses on the environmental justice issues, and it actually shows that AI models have more specific information for urban areas compared to the rural areas."
Kim acknowledged that it's not great that the small rural communities which can benefit most from Large Language Learning Model-assisted audits are also most likely to get inaccurate results from these tools. And he also acknowledged that even the best-informed city might not actually fix the walkability gaps that AI helps them find, whether because of lack of funding, lack of staff, or simple lack of political will.
Still, for cities who are willing to be put in the work, ChatGPT and Gemini could provide a critical starting point to make their streets more people friendly — even if it's not a silver bullet.
"As a geographer and an urban researcher myself, we're not trying to say this is the optimal answer," said Kim. "But we are trying, in a more 'user perspective' way, to understand how these tools are performing, and how this can still be very effective, regardless of the risks."
