AV vs Human Driver: The Real Comparison

The Promise: AVs Are Safer

Both Waymo and Tesla publish safety data that paints an optimistic picture. Waymo's Safety Impact Dashboard, updated through December 2025 and covering 170.7 million rider-only miles, shows a 92% reduction in serious injury crashes compared to human benchmarks in its operating cities. Tesla's quarterly Vehicle Safety Report claims vehicles with Autopilot engaged experience roughly one crash per 6.4 million miles — compared to one per 702,000 miles without it.

The human baseline from NHTSA's FARS and FHWA data tells us what "normal" looks like: 1.26 fatalities per 100 million vehicle-miles traveled, and 2.07 crashes per million VMT across all road types and conditions. Approximately 40,990 Americans die in traffic each year, with 6.7 million total crashes.

Waymo: The Strongest Safety Case

Waymo's data is the most compelling in the industry. Across all operating locations, Waymo's serious injury rate is 0.02 incidents per million miles compared to a human benchmark of 0.22 IPMM — a 91% reduction. For any-injury crashes, the gap is similarly dramatic: 0.71 IPMM for Waymo vs 3.90 for human drivers.

The city-level data adds nuance. San Francisco — Waymo's densest and most challenging market — shows a serious injury rate of 0.04 IPMM against a benchmark of 0.43. Phoenix, with its wider roads and simpler traffic patterns, shows an even more dramatic 0.01 vs 0.10. Los Angeles and Austin both show 0.00 serious injuries from Waymo so far, though miles driven are lower.

Waymo also excels with vulnerable road users: 92% fewer pedestrian crashes, 85% fewer cyclist crashes, and 81% fewer motorcycle-involved crashes. These are the road users most at risk from human error, and Waymo's multi-sensor stack (lidar + radar + cameras) appears to detect them more reliably.

Tesla: Impressive But Caveated

Tesla's safety claims are harder to evaluate. The company reports crash rates per mile — 0.157 IPMM with Autopilot vs 1.425 without — but these numbers carry a critical asterisk: Autopilot is used predominantly on highways and divided freeways, which are inherently the safest road types. The human fatality rate on interstates is just 0.64 per 100M VMT, compared to 2.42 on rural roads.

This "highway bias" means Tesla's Autopilot isn't necessarily outperforming the human average — it may simply be operating on roads where everyone crashes less. When Tesla compares its Autopilot crash rate against the national average (which includes urban streets, rural roads, and impaired drivers at 2 AM), the comparison flatters the technology.

The quarterly trend does show improvement. Tesla's Autopilot crash rate has declined from 0.232 IPMM in Q3 2023 to 0.157 in Q3 2025 — a 32% improvement over two years. Whether that reflects better software or simply more highway miles is debatable.

The Human Baseline: Context Matters

The "average human driver" is a fiction. A sober, alert 40-year-old commuting on an interstate in daylight has a vastly different risk profile than a 19-year-old driving a rural road at midnight after drinking. The national averages — 2.07 crashes per million VMT, 1.26 fatalities per 100M VMT — blend all of these scenarios together.

For a fair comparison, you'd need to compare each AV system against the human performance on the same road types, at the same times, in the same conditions. Waymo does this to some degree with its city-level benchmarks. Tesla does not.

The Reporting Gap Problem

How each entity counts "crashes" matters enormously. Waymo reports every contact event, including minor scrapes where another vehicle rear-ends the stopped robotaxi. Tesla's crash definition in its safety reports isn't identical to NHTSA's. The national FARS data only counts crashes where a police report was filed. These methodological differences mean the numbers aren't truly apples-to-apples, even when they appear to be.

The Bottom Line

Waymo's fully autonomous system appears genuinely safer than human drivers in its operating cities, with strong data across multiple severity levels and road user types. Tesla's Autopilot data shows promise but is compromised by highway bias — comparing mostly-highway miles against all-road averages overstates the safety benefit. And the human baseline itself is an average that includes the worst drivers in the worst conditions, making any technology look good by comparison.

The real answer to "are AVs safer?" is: some are, in some places, compared to some human drivers. The data supports cautious optimism for fully autonomous systems like Waymo. For ADAS systems like Tesla Autopilot, the jury is still deliberating — and the 56 fatalities in the NHTSA database make that deliberation urgent.