AI Cars Can See Crashes But Can't Tell Who's at Fault

Self-driving cars are getting really good at noticing the world around them. They can spot pedestrians, other vehicles, and traffic lights. But when it comes to understanding an accident—like who caused it or what rule was broken—they're surprisingly bad. To test this, researchers built CAViAR, a collection of 2,249 real dashcam videos from crashes, each labeled by humans with details like weather, accident type, and which car was at fault.

They then ran the latest AI vision models on these videos. The results showed a clear gap. The models handled simple things like lighting almost perfectly. But they struggled with tough conditions like bad weather or slippery roads, often performing no better than a random guess. Worse, when asked to decide who was at fault or what traffic rule was violated, their accuracy dropped sharply. In other words, AI can see a crash but can't reason about why it happened.

This is a big deal for the future of self-driving cars. If an autonomous vehicle doesn't understand accidents, it can't learn to avoid them or explain its actions after a collision. It also raises questions about fairness: if an AI can't tell who's responsible, how can we trust its decisions in the real world?

The researchers hope CAViAR will help fix this. The dataset and tools are publicly available, so other teams can use it to train better AI. But until that wiring improves, fully self-driving cars will still lack the judgment needed to handle dangerous situations responsibly.