What Happened
On the night of March 18, 2018, an Uber autonomous test vehicle operating in self-driving mode struck and killed 49-year-old Elaine Herzberg as she walked a bicycle across a road in Tempe, Arizona. The vehicle’s AI system detected Herzberg approximately 6 seconds before impact but failed to properly classify her or initiate emergency braking. The backup safety driver was looking at a phone and did not intervene.
The NTSB investigation later determined that the system had detected Herzberg but repeatedly reclassified her — as an unknown object, a vehicle, then a bicycle — never maintaining a consistent track long enough to trigger emergency braking.
Timeline
The fatal crash occurred at approximately 9:58 PM on March 18, 2018. Uber immediately suspended all autonomous testing across all cities. The NTSB launched a formal investigation. In December 2019, the NTSB released its final report attributing the crash to an inadequate safety culture at Uber, the failure of the automated driving system, and the inattention of the safety driver. In September 2020, Uber sold its self-driving division to Aurora Innovation.
Impact
The incident had profound effects on the autonomous vehicle industry. It demonstrated that self-driving technology could kill in real-world conditions and that the safety frameworks in place were insufficient. Uber’s autonomous testing program never recovered. The company eventually sold its entire self-driving division at a significant loss.
The fatality also triggered regulatory responses worldwide. Arizona suspended Uber’s autonomous testing permits. Other jurisdictions tightened requirements for safety drivers, disengagement protocols, and testing conditions. The incident set back public confidence in autonomous vehicles by years.
Response
Uber cooperated with the NTSB investigation and settled with Herzberg’s family for an undisclosed amount. The backup safety driver, Rafaela Vasquez, was charged with negligent homicide and ultimately pleaded guilty. Uber eventually sold its ATG division to Aurora Innovation in December 2020, effectively admitting it could not develop the technology safely or economically.
Lessons Learned
The Uber fatality demonstrated that AI decision-making systems in safety-critical applications require extensive redundancy, fail-safe mechanisms, and robust classification consistency. The system’s inability to maintain a consistent classification of Herzberg — cycling between “unknown,” “vehicle,” and “bicycle” — prevented appropriate emergency response.
The incident also revealed systemic organizational failures. Uber had previously disabled emergency braking on its test vehicles to prevent “uncomfortable” false stops, prioritizing ride comfort over safety. This decision directly contributed to the fatality and became a stark example of how organizational incentives can undermine AI safety.