Human–AI Interaction and Traffic Safety in Semi-Autonomous Driving: Evidence from Tesla’s Autopilot

Abstract

While AI-powered autonomous driving holds the promise to confer economic and societal benefits, there has been a persistent concern on the safety of autonomous vehicles, often necessitating human oversight. Drawing on a theoretical framework of distributed cognition, this study investigates the impact of human-AI systems on safety within the context of semi-automated driving systems. We design a natural experiment by combining Tesla’s rollout of its semi-autonomous driving feature, Navigate on Autopilot, in the U.S. with regional variations in the intensity of Tesla vehicles at the county level. Using granular data on traffic accidents and Tesla registrations in Washington State during the period of 2011-2022, we find that a higher concentration of Tesla vehicles is significantly associated with a reduction in traffic accidents following the rollout of Tesla’s advanced autopilot feature, particularly in collisions between Tesla and non-Tesla vehicles. We further identify the conditions under which semi-autonomous driving is more effective. Results show that the effect of semi-autonomous driving on safety is stronger when driver uncertainty is low and the road environment is less complex. Our findings suggest that the safety of semi-autonomous driving as a shared cognition system of human and AI is shaped by driver-related and situational factors as well as the limits of its technological capabilities, offering important practical and policy implications for the development and regulation of human-AI systems.