How this calculator works, and what it refuses to claim
First, plainly: nobody has a statistically settled crash rate for autonomous semis. Not Tesla, not Aurora, not Kodiak, not any regulator, and not this page. The driverless trucks hauling freight in Texas today have logged hundreds of thousands of miles; a human-driven fleet expects its first fatal crash only after about 61 million. At that ratio, a clean record proves care, not safety. So this page does not claim autonomous trucks are safe, and it does not claim they are not. It does three honest things instead. It anchors the human baseline in verified federal crash data, the one side of this argument that is actually settled. It hands you the assumption: the crash-reduction dial, with each preset labeled by where its number comes from. And it shows every result across the whole range of assumptions, including the skeptic's zero, so you can watch the assumption do the work. That design is the whole page. A reader who leaves believing the answer depends on an assumption nobody has proven yet has understood the subject better than most of the internet arguing about it.
One more thing before the arithmetic, because the subject deserves it. Every number below stands for someone's worst day: 5,472 of them in 2023 alone. We will run the expected-value math, because it is the only honest way to compare a technology against a baseline, but we will not pretend a fraction of an expected death is anything other than a real death arriving on an unknown date. Where the numbers get small, this page switches from decimals to plain frequency, one fatal crash in so many years, because that is how the risk is actually lived.
The human baseline: what the federal numbers count
In 2023, the most recent complete federal data year, crashes involving large trucks (gross vehicle weight over 10,000 pounds) killed 5,472 people in the United States and injured an estimated 153,452 more (NHTSA, from the FARS fatality census and the CRSS crash sample). Large trucks drove 329.9 billion miles that year (Federal Highway Administration), which yields the baseline this page runs on: about 1.66 deaths, 46.5 people injured, and 160 police-reported crash involvements per 100 million truck miles. Our derived involvement rates, 1.63 fatal and 34.7 injury per 100 million miles, match NHTSA's own published rates of 1.63 and 35, and the test suite behind this page fails if they ever stop matching. The preliminary 2024 count, 5,340 deaths, says 2023 was no fluke.
And one fact in that data reframes the whole question of who autonomous-truck safety is for: of the 5,472 people killed, 70 percent were in the other vehicle (3,837 people), 18 percent were in the truck (961), and 12 percent were pedestrians, cyclists, or others outside any vehicle (674). When an 80,000 pound combination meets a passenger car, the car loses. Truck safety is mostly about the people the truck meets on its worst mile, which is why this argument belongs to everyone on the road, not just the freight industry.
The formula
expected events = scenario miles × federal 2023 rate per mile (deaths, injuries, crash involvements)
avoided events = expected events × assumed reduction
economic value = avoided crashes × FMCSA cost per crash (fatal $15,230,414; injury $326,810; property-damage $49,398, in 2023 dollars)
The rates are derived live from the federal counts (5,472 deaths, 153,452 injured, 528,177 crash involvements, 329,858 million miles) so the steps in every result show the division, and the assumed reduction is yours, never ours. The per-crash costs are FMCSA's comprehensive figures from its 2025 crash-cost methodology, which fold in medical costs, lost productivity, congestion, property damage, and the Department of Transportation's value of a statistical life: $13.2 million in 2023 dollars, $14.2 million in the current guidance. That last figure deserves its one respectful sentence: it is a statistical valuation used to compare safety investments in federal rulemaking, not a price on any person, and this page uses it only in that spirit.
Worked example
A 100-truck fleet, 100,000 miles per truck: 10 million miles a year. On the verified human baseline, those miles expect 16 crash involvements a year, 3.5 of them injury crashes, about 4.7 people injured, and one fatal crash about every 6.1 years.
Set the dial to 85 percent, the setting derived from Waymo's published passenger-car results (which is passenger-car ride-hail evidence, not truck evidence, and the page says so right on the dial). If that assumption held for semis, this fleet would see about 3 fewer injury crashes a year, and its expected fatal crash would move from once every 6.1 years out to once every 40.9 years. The economic line, priced at FMCSA's per-crash costs: 0.1385 avoided fatal involvements × $15,230,414, plus 2.952 injury × $326,810, plus 10.52 property-damage × $49,398, is about $3,593,877 a year of avoided crash costs.
Now the row that keeps this page honest: at 0 percent, the skeptic's setting, the fleet keeps its 3.5 injury crashes a year and its fatal crash every 6.1 years, and the avoided-cost line reads $0. Everything between those two rows is bought entirely by the assumption, which is exactly why the table shows them side by side.
The evidence, such as it is
The strongest safety numbers in autonomous driving belong to passenger cars, and it matters enormously that they do. Waymo's data, 82 percent fewer injury crashes, 83 percent fewer airbag deployments, and 92 percent fewer serious-or-fatal-injury crashes than matched human benchmarks over 170 million driverless miles (with the core findings peer reviewed in Traffic Injury Prevention at the 56.7 million mile mark), is the best evidence yet that machine driving can beat human driving somewhere. A separate study with reinsurer Swiss Re found 92 percent fewer bodily-injury liability claims and 88 percent fewer property-damage claims over 25.3 million miles against baselines built from half a million human claims. Those studies are why our 85 percent preset exists. But every one of those miles was driven by a passenger car doing ride-hail work on city streets. None of it was an 80,000 pound combination at highway speed, and borrowing the number across that gap is precisely the assumption this page makes you set by hand rather than making for you.
What the truck operators have actually done, as of August 2026: Aurora launched commercial driverless service between Dallas and Houston in May 2025, expanded to Fort Worth, El Paso, and Phoenix routes, passed 250,000 driverless miles in January 2026 with no collisions attributed to its system, and plans over 200 driverless trucks by the end of 2026. Kodiak has run driverless trucks hauling frac sand in the Permian Basin since 2024, mostly on private routes, with public-road operation planned. Tesla's Semi entered volume production in 2026 as a driver-operated electric truck; its autonomy remains a stated aim, about a year away by Elon Musk's July 2026 estimate, and any Tesla Semi safety-by-autonomy figure is a target, not a measurement. Respect what that record is: real freight, real highways, zero attributed collisions. And respect what it is not: at 250,000 miles against a baseline of one fatal crash per 61 million miles, the data cannot yet distinguish excellent from average, let alone prove the revolution.
Fatigue: the strongest honest argument for the technology
Here is the argument for autonomous trucks that survives every skeptical filter, and it is not a marketing number. FMCSA's Large Truck Crash Causation Study coded 13 percent of truck drivers in serious crashes as fatigued at the moment of the crash, and researchers treat that as a floor, because fatigue leaves no breathalyzer. At the grim end of the scale, a 1990 NTSB study of crashes that killed the truck driver found fatigue the most frequently cited probable cause, at 31 percent. Somewhere between those numbers lives the most preventable slice of truck-crash deaths, and it is the one slice a computer deletes entirely rather than merely improving: software does not get drowsy at hour ten of a Texas interstate, does not push through the last hundred miles, and does not fall asleep at 65 mph. The same holds for the impaired and the distracted. A machine that merely drove like a median sober, rested, attentive human, all day, every day, would already be better than the fleet average, because the fleet average includes the tired. That is a real argument, honestly stated. It is still an argument, not a measurement.
Why trucks might be the strongest case, and the honest counters
The optimist's case is about where truck miles happen. Long-haul trucking concentrates its miles on limited-access interstates: no pedestrians, no cyclists, no intersections, traffic all moving one direction at similar speeds. That is the most structured, most machine-legible driving environment in America, and it is exactly the environment where current autonomy performs best. Waymo's hardest problems, the darting child, the ambiguous four-way stop, the wrong-way cyclist, are mostly problems trucks meeting their miles on I-45 do not face. It is a genuinely strong argument, and it is the reason serious people believe trucks, not robotaxis, will be autonomy's first profitable scale.
Now the other side of the same coin, with no thumb on the scale. Trucks do not live only on interstates: they start and end at docks, fuel islands, and surface streets, and those first-and-last miles contain most of the complexity the interstate argument waves away. Weather that a Phoenix robotaxi never meets, crosswinds, black ice, blinding snow, is routine for a national fleet. And the physics are unforgiving in a way passenger-car evidence cannot speak to: a loaded combination needs roughly half again the stopping distance of a car, so a rare failure that a robotaxi walks away from can be catastrophic in a semi. Rare-but-severe is exactly the failure mode small datasets are worst at measuring. Both arguments are true at once. The interstate makes the everyday miles easier; the mass makes the bad mile worse. Which effect dominates is an empirical question that only published miles will settle, which is why the dial on this page is yours.
Insurance and the liability shift
This page keeps insurance out of the results on purpose, because premium predictions for driverless fleets would be invented numbers. The context, though, is worth knowing. Commercial truck insurance commonly runs $8,000 to $16,000 or more per truck per year, and it has been dragged upward less by fender-benders than by verdict severity: ATRI's research found the mean verdict in truck cases above $1 million grew from about $2.3 million to $22.3 million between 2010 and 2018, and awards above $10 million are common enough to have their own industry nickname, nuclear verdicts. If autonomy actually cuts crash frequency, claims frequency falls with it; the Swiss Re result on Waymo's passenger cars is the template for what that looks like in an insurer's ledger. But autonomy also moves the defendant. A tired-driver case is negligence against a driver and a motor carrier; a driverless crash is a product-liability case against a manufacturer with deep pockets and a jury meeting a headline technology. Fewer claims, scarier claims, and nobody has enough verdicts yet to price the trade. Qualitative, honestly labeled, and kept out of the math.
What would settle the question
One thing, and only one: published per-mile crash rates from truck operators at scale, injury crashes and worse per million miles, methodology stated, benchmarks matched to the same roads, the way Waymo chose to publish and be checked. The company that does this first will not just win an argument; it will define how the public learns whether this technology keeps its central promise. Until then, the honest state of knowledge fits in one sentence: the human baseline is verified, the improvement is an assumption, and this page prices the assumption without ever mistaking it for a fact.