AI Cars Can't Quite Handle Busy City Streets Yet, New Test Shows

Researchers at a group of Chinese universities built a fake 3D city called VehicleArena to test how well AI can drive. Unlike most driving tests, this one does not tell the cars what to do. Each AI car is controlled by a large language model (the same kind of AI behind ChatGPT) and gets its own passenger requests, like a ride-hailing driver. The cars have to figure out routes, handle traffic, and drop people off — all in the same shared world at the same time.

The results are a reality check. Nine AI models were tested across 112 tasks, and the best one finished only 65% of trips. That means about one in three rides ended in failure. Perhaps more revealing: models that scored well on pleasant passenger interactions — friendly chat, comfortable cabin behavior — were not the ones that actually completed the drive. Being a good conversationalist is not the same as being a good driver.

The multi-car test uncovered something that matters for real roads. When AI cars drove alongside regular simulated traffic, every single model tested made the other cars slower. In other words, one clever AI driver created a traffic jam for everyone else. This is called an "externality" — a cost your decision pushes onto other people. It's the same reason one person slamming the brakes on a highway causes a ripple of delays miles back.

Why does this matter? Self-driving cars are already on some roads, and the whole pitch is that they'll be safer and smoother than humans. This paper suggests that current AI brains, when placed in complicated city conditions with other drivers, are neither reliably competent nor reliably polite to the traffic around them. Before robotaxis fill our streets, someone needs to solve both problems at once.