Start with the scoreboard, not the science fiction.
In 2024, U.S. roads killed 39,254 people – a fatality rate of 1.19 deaths per 100 million vehicle miles traveled, down from 1.26 in 2023, with about 2.42 million people injured (NHTSA Overview of Motor Vehicle Traffic Crashes in 2024). That is the human baseline. The case for autonomy is whether published crash rates beat it in comparable conditions – and by how much.
They do, for the systems with the strongest public data. The gap is largest for fully driverless robotaxis with peer-reviewed comparisons. Tesla’s Full Self-Driving (Supervised) publishes large fleet collision rates that look better than national averages on Tesla’s own chart – but it is a different product class (supervised Level 2), uses different crash definitions. The rest of this piece is those numbers, named sources, and the caveats that keep the argument honest.
Travelers already live inside that baseline every time they close a car door or walk a crosswalk. The useful question is not whether machines will ever err. It is whether the best-measured systems cut the crashes that fill hospitals – late recognition, bad decisions, fatigue, distraction – more often than ordinary human driving does on the same streets.
The human baseline in numbers
- 39,254 U.S. traffic deaths in 2024; 1.19 per 100 million VMT (NHTSA)
- ~2.42 million people injured; injury rate ~74 per 100 million VMT (same NHTSA overview)
- In NHTSA’s National Motor Vehicle Crash Causation Survey, the “critical reason” (last failure in the crash chain) was assigned to the driver in an estimated 94% (±2.2%) of studied crashes – recognition, decision, and performance errors dominating. NHTSA stresses that is not the same as sole legal blame; it is where the final failure usually sits.
That 94% figure is easy to misuse. It does not mean “drivers are reckless.” It means the last failure in the chain is usually human: looking but not seeing, picking a gap that is not there, or reacting late. Roads, weather, and other users still matter. The scalable lever is still the driver task – how often the system that steers and brakes misses the same cues people miss when tired or distracted.
Fatigue, drowsiness, and attention
Tired drivers are not a niche edge case. They are already inside the human baseline those NHTSA fatality and injury rates measure. Fatigue rarely announces itself. It slows reaction time, narrows what you notice, and turns ordinary recognition errors into crashes: the late brake, the drifted lane, the gap that was never there.
Official tallies undercount sleepiness because officers do not have a roadside test for drowsiness the way they do for alcohol. NHTSA’s drowsy-driving overview still reports hundreds of deaths a year in police-coded drowsy crashes (on the order of 600+ recently; the agency page has listed about 644 for 2024) and stresses that precise totals are hard to pin down. CDC survey work found roughly 1 in 25 adult drivers saying they had fallen asleep while driving in the prior 30 days (CDC MMWR BRFSS, 2011–2012). IIHS summarizes the same gap: police reports put drowsiness in only about 1–2% of crashes, while naturalistic studies that watch alertness find closer to 9–10%. Separate AAA Foundation modeling has put drowsy involvement near 17.6% of fatal crashes in 2017–2021 (AAA Foundation) – far above the raw police-coded share.
In plain terms, a short night of sleep raises the odds of the same failure modes that fill the Crash Causation Survey: looking but not seeing, deciding late, performing poorly when a sudden stop or merge arrives. Those mistakes sit inside the same U.S. baseline of roughly 39,000 traffic deaths a year and about 2.42 million injured. A supervised system that keeps tracking the road when the person in the seat is fading does not erase human risk, and it does not excuse zoning out. It targets a failure mode willpower does not reliably fix on a Tuesday night commute or a long highway run after work.
Rates beat headlines. A large connected fleet can post scary absolute crash counts because telemetry reports every hard event; a quieter city can hide worse rates if miles are thin. Compare injuries and police-reportable events per mile, in similar places, with readable definitions. That is the discipline behind the sections below.
Scoreboard: published rates side by side
Different programs measure different things. Read the columns, not just the green percentages.
| System / source | Exposure | Key rate or reduction | Compared with | Automation level |
|---|---|---|---|---|
| U.S. human driving (NHTSA 2024) | ~3.29 trillion VMT | 1.19 fatalities / 100M VMT; ~74 injuries / 100M VMT | – | Human |
| Waymo Rider-Only (Safety Impact, through June 2026) | 271.3M rider-only miles | ~82% fewer any-injury crashes; ~95% fewer serious injury+; ~82% fewer any-vehicle airbag deployments (combined locations) | Local human police benchmarks on comparable surface streets | Driverless (SAE L4 robotaxi) |
| Waymo RO peer-reviewed (Kusano et al. 2024, TIP) | 7.1M miles | ~80% lower any-injury involvement; ~55% lower police-reported involvement | Same style of local human benchmarks | Driverless |
| Waymo RO, IIHS cleaned police-reportable (IIHS) | ~50M driverless miles in study window | ~68% lower police-reportable crash involvement / VMT; ~81% fewer injury crashes / VMT | Human police crashes, same cities/years | Driverless |
| Tesla FSD (Supervised) – major collisions (Tesla FSD Vehicle Safety Report, live counters / rolling 12-mo chart) | ~15.0 billion cumulative FSD miles; ~5.8 billion city miles (counters tick live) | 5.7 million miles per major collision (N. America, all roads) | U.S. average 699,000 miles per major collision on Tesla’s chart; Tesla also claims 7× fewer major and minor collisions vs estimated U.S. average | Supervised ADAS (SAE L2); driver required |
| Tesla FSD (Supervised) – minor collisions (same source) | Same rolling FSD exposure | 1.6 million miles per minor collision (N. America, all roads) | U.S. average 237,000 miles per minor collision on Tesla’s chart | Supervised; Tesla collision definition (see caveats) |
Waymo: the strongest published driverless case
Waymo’s Safety Impact hub reports 271.3 million rider-only miles through June 2026. Against human benchmarks for passenger vehicles on the surface streets where Waymo operates, the combined-location tables show roughly:
- ~0.67 vs ~3.77 any-injury-reported incidents per million miles (~82% lower)
- ~0.01 vs ~0.21 serious injury or worse IPMM (~95% lower)
- ~0.29 vs ~1.62 airbag-deployment-in-any-vehicle IPMM (~82% lower)
- Vulnerable-road-user injury crashes down on the order of ~90%+ for pedestrians and high double digits for cyclists and motorcyclists (same hub)
Those methods appear in peer-reviewed Traffic Injury Prevention papers: Kusano et al. (2024) at 7.1 million miles (~80% any-injury reduction; ~55% police-reported), and Kusano et al. (2025) extending crash-type analysis to 56.7 million miles. IIHS’s independent cleaning of federal reports to police-reportable events still found Waymo’s driverless involvement rate about 68% lower per mile than humans in the same cities (IIHS).
“Rider-only” is the important phrase: miles with no safety driver – the product people mean by robotaxi. The benchmarks are local too, drawn from the same cities and surface streets Waymo serves, not a national blend of rural Interstates and downtown grids. IIHS’s cleanup still matters: when soft or non-comparable reports are stripped and police-reportable events per mile remain, the advantage shrinks from the flashiest hub figures and stays large. Outside Waymo’s mapped operating domain, the honest answer is “not yet measured the same way,” not a blank check either way.
Tesla FSD (Supervised): huge miles, different product
Tesla is the other large public dataset – and the one most people mean when they say “self-driving car.” It is also the easiest place to mix up categories. Tesla’s live FSD (Supervised) Vehicle Safety Report (the old VehicleSafetyReport URL redirects here) reports FSD only – not a separate Autopilot miles-per-crash chart on that page.
What Tesla publishes (primary page). Live counters show on the order of 15.0 billion cumulative FSD (Supervised) miles and 5.8 billion city miles (they increment continuously; Tesla says the series runs from 2020 onward). On the rolling twelve-month “miles driven before a collision” chart for North America, all road classes, Tesla shows roughly:
- Major collisions: FSD 5.7 million miles per collision vs U.S. average 699,000 (manual Teslas with active safety: 2.1M; without: 859K)
- Minor collisions: FSD 1.6 million vs U.S. average 237,000
- Headline multipliers on the same page: 7× fewer major and 7× fewer minor collisions vs the estimated U.S. average; Tesla also states FSD “improves U.S. road safety by over 80%” against human-error collisions
Highway miles look even longer between major collisions on Tesla’s chart (9.7 million FSD vs ~1.4 million U.S. average); non-highway is shorter but still ahead (3.5 million vs 526,000). Those are company telematics rates – not police-reportable injury rates like Waymo’s peer-reviewed tables.
How Tesla defines a crash (important). Collisions follow 49 C.F.R. § 563.5-style triggers: airbag/pyrotechnic “major” deployments, or Delta-V ≥ 8 km/h within 150 ms for “minor” collisions. If FSD was active at any point in the five seconds before impact, Tesla counts it as FSD-engaged. Fault is not assigned. The U.S. average on the chart is Tesla’s construct from FHWA VMT and NHTSA CISS involvement counts (with highway/non-highway and minor rates scaled using manual-Tesla ratios). Tesla itself notes police underreporting (NHTSA/Blincoe: ~60% of property-damage and ~32% of injury crashes unreported) and database mismatch limits.
Supervised means the human is still the fallible backup. SAE Level 2 can steer and accelerate within design limits, but the driver must monitor and intervene. That is a different problem than a Level 4 robotaxi built to finish the trip without a ready human. Long miles between telematics collisions can still coexist with sharp edge cases – especially when cameras are impaired – because the backup driver also gets tired or overconfident. Tesla’s own labeling keeps saying Supervised for a reason.
Older adults and supervised workload relief. The same product class can matter for another practical reason as people age: not as an unsupervised robotaxi, and not as medical advice about who should keep driving, but as a way to reduce continuous steering and lane-keeping load on highway and other routine trips. Tesla FSD (Supervised) remains SAE Level 2. The driver must stay attentive and ready to intervene. Used correctly – eyes up, hands ready, takeover when the car hesitates or the scene turns messy – it can simplify stretches of driving that get harder when attention and reaction speed are taxed. It does not mean older drivers can stop supervising, leave the seat, or treat the car as a finished robotaxi. It is assistance inside Level 2 limits, not a substitute for judgment about whether someone should still be behind the wheel.
Tesla vs Waymo in one sentence. Waymo’s headline case is driverless crash rates versus local humans, peer-reviewed and cross-checked by IIHS. Tesla’s headline case is supervised FSD telematics (~5.7M miles per major collision vs ~699K U.S. average on its chart) – while still requiring a responsible driver. Treat a robotaxi without a driver and a consumer car with FSD engaged as different technologies that happen to share marketing vocabulary.
What the statistics do – and don’t – prove
- Do: Show large, repeated reductions in injury and serious outcomes for Waymo rider-only miles versus local human benchmarks.
- Do: Show Tesla’s published FSD (Supervised) miles-per-collision figures (e.g. 5.7M vs 699K major, N. America) beating the U.S. average on Tesla’s chart – at multi-billion-mile scale.
- Don’t: Prove every branded “self-driving” feature is equally safe. Supervised ≠ driverless.
- Don’t: Erase operating-domain limits (cities, weather, road type) or reporting-definition mismatches.
- Don’t: Let raw SGO crash counts, without miles, settle the debate.
Edge cases, weather, and the regulatory gap
Public fear often pictures a robotaxi frozen in an intersection or a camera blinded by sun. Those are real failure modes – and where the data story gets careful. Waymo’s published advantage is measured inside cities it is allowed to serve. Tesla’s FSD (Supervised) chart is a telematics collision rate across a huge fleet that still needs a human watching. Neither answers every question about the next storm, software push, or unmarked work-zone merge.
The seatbelt parallel
Seat-belt use sat around 11-14% in the late 1970s/early 1980s before mandates; New York passed the first state law in 1984 (CDC history). NHTSA estimates belts saved about 14,955 lives in 2017 and roughly 374,276 lives from 1975-2017 (NHTSA). Autonomy is earlier on that arc – contested, geofenced, imperfect – but the direction of the best rate data matches what seat belts looked like after the first serious evaluations: messy politics, clearer harm reduction when the safer system is actually used.
Belts did not make cars crash-proof; they cut the chance a crash killed you. Stay disciplined about rates and autonomy looks similar: best-measured driverless fleets show fewer injury and police-reportable crashes per mile than local humans; supervised FSD publishes longer miles between telematics collisions than Tesla’s U.S.-average construct, with the usual caveats that supervised systems still need an attentive driver. Adoption and measurement still have to catch up the way buckle-up campaigns did – publicly, with numbers that survive skepticism.
Bottom line
Human driving still kills on the order of 40,000 Americans a year at roughly 1.19 fatalities per 100 million VMT. The strongest published driverless data (Waymo Safety Impact) show large reductions vs local human benchmarks on rider-only miles. Tesla’s FSD (Supervised) Safety Report publishes roughly 5.7 million miles per major collision vs about 699,000 on Tesla’s U.S.-average chart, at roughly 15 billion cumulative FSD miles – but remains a supervised Level 2 system that still needs an attentive driver. The statistical case for autonomy is already strong where the data are cleanest. The job now is to keep measuring rates – not vibes – as more cities and more supervised miles come online.
One distinction to keep: safer than the human baseline on published rates is not unsupervised everywhere, and it is not finished. It is still the strongest evidence that replacing – or carefully assisting – the driver task can cut real harm, if we keep product classes straight and keep reading the denominators.