Self-Driving Car Accidents

The State of Self-Driving Car Safety

Since NHTSA began requiring manufacturers to report autonomous vehicle incidents under Standing General Order 2021-01, the database has grown to 7,082 reported incidents across 61 manufacturers. These incidents span 52 U.S. states and territories, painting the most complete picture of autonomous driving safety available to the public.

The data reveals two fundamentally different approaches to vehicle automation. ADS (Automated Driving Systems) like Waymo and Cruise operate vehicles without a human driver, accounting for 2,977 incidents. ADAS (Advanced Driver Assistance Systems) like Tesla Autopilot and GM Super Cruise require a human driver, with 4,105 reported incidents. For a detailed comparison of these approaches, see our Tesla vs Waymo analysis.

Severity Distribution

Of all reported incidents, 71 involved at least one fatality and 3,491 involved injuries. The severity distribution shows that 71 were fatal (1.0%), 3,491 involved injuries (49.3%), and 3,520 resulted in no injuries (49.7%).

The vast majority of AV incidents result in no injuries — a pattern that holds across both ADS and ADAS categories. However, when fatalities do occur, they tend to involve ADAS vehicles at highway speeds, where crash energy is dramatically higher. For detailed fatality data, see our self-driving car deaths database.

Year-Over-Year Trends

Reported AV incidents have been trending upward, growing 51% from 2024 to 2025. This growth reflects several factors:

• Expanding fleets: More vehicles with autonomous features are on the road each year, especially Tesla with its growing ADAS-equipped fleet.
• Improved reporting: NHTSA's SGO reporting requirements have become better understood and more consistently followed by manufacturers.
• New entrants: Additional ADS companies have begun commercial operations, adding to the incident count.
• Urban expansion: ADS robotaxis are expanding to new cities, introducing new operating environments and new incident scenarios.

It's critical to understand that increasing incident counts don't necessarily indicate worsening safety. Without per-mile exposure data, we can't determine whether the per-mile crash rate is rising, falling, or stable.

Speed and Weather Patterns

The average speed at the time of AV incidents is 36 mph, reflecting a mix of highway ADAS crashes (higher speeds) and urban ADS crashes (lower speeds). Weather conditions at the time of incidents were predominantly clear (3968 incidents), followed by cloudy conditions and rain.

The clear-weather dominance likely reflects both the geographic concentration of AV operations in Sun Belt states and the tendency for autonomous systems to be disengaged in severe weather. This also means the data tells us little about how AVs perform in rain, snow, or fog — conditions that human drivers find most challenging.

ADS vs ADAS: Understanding the Split

The 2,977 ADS incidents and 4,105 ADAS incidents represent fundamentally different safety challenges. ADS incidents tend to be low-speed urban events where the vehicle is fully in control. ADAS incidents span a wider range of scenarios, from highway lane departures to intersection collisions, with a human driver who may or may not have been attentive.

The distinction matters for policy, insurance, and consumer decision-making. When someone asks "are self-driving cars safe?", the answer depends heavily on which type of system they're asking about. A Waymo robotaxi and a Tesla with Autopilot are both "self-driving" in common parlance, but they're very different in capability, responsibility, and risk profile.

What This Data Doesn't Tell Us

While the NHTSA database is the most comprehensive public source on AV safety, it has important limitations:

• No per-mile rates: Without knowing total miles driven with automation engaged, raw incident counts can't tell us whether AVs are safer than human driving.
• Inconsistent reporting: Different manufacturers have different interpretations of what constitutes a reportable incident, creating comparison challenges.
• System engagement unclear: Not all incidents occurred while automation was actively engaged — some involved vehicles merely equipped with the technology.
• Fault determination missing: The data doesn't consistently indicate whether the AV or another party was at fault.
• Near-miss data absent: Only crashes are reported — near-misses and successful interventions that prevented crashes aren't captured, making it impossible to assess how often the system prevents incidents.

For a consumer-focused interpretation of what this data means for safety, see our guide on how safe self-driving cars really are.