Study Finds Human Drivers Are Still Better Than ‘AI Drivers’

A comprehensive study recently analysed over 37,000 vehicle collisions, revealing that human drivers are less likely to crash during turns or in low-light (twilight) conditions compared to automated systems. The disparity in performance is significant.

Published in Nature Communications, the study found that vehicles using automated systems had five times as many crashes during sunrise or sunset.

In corners, the ratio was twice as many compared to human-driven vehicles. The data encompasses approximately 35,000 crashes involving human drivers and 2,100 with automated systems active.

The study also highlighted several limitations with current automated technology, particularly in low-light situations.

Cameras and sensors, crucial for these systems, struggle to adapt to varying light conditions.  Early-morning or late-day shadows can be misinterpreted as objects, and fluctuating light can confuse algorithms, causing the system to fail. 

Additionally, objects in shadows might not be detected at all. This is supported by crash tests where vehicles often brake late or fail to stop for simulated pedestrians or animals.

When it comes to turns, the study suggests that situational awareness is a significant challenge for current automated systems. 

Sensors and cameras may not detect all obstacles in dynamic environments like intersections. Human drivers can see further and take precautionary actions, such as slowing down for a fog bank a half-mile away, while automated systems may not recognise such distant hazards and continue at a constant speed.

The study even provides evidence to back these claims. Before a collision, vehicles under autonomous control were often driving straight and maintaining a constant speed until emergency maneuvers were necessary. In contrast, human-driven cars showed more instances of slowing down and changing lanes before impact. KR