FSD Is Trying to Replace the Safest Drivers in America

NHTSA opened engineering analysis EA26002 on March 18, targeting Tesla’s Full Self-Driving system across 3,203,754 vehicles. Nine crashes. One fatality. The defect: FSD’s cameras can’t tell when they can’t see. Glare, fog, debris—the system either missed the visibility degradation entirely or fired warnings milliseconds before impact.^[1]

That investigation is warranted. But the FARS database adds a dimension NHTSA’s filing doesn’t mention.

The Model 3’s fatality rate is 0.05 deaths per 100 million VMT—the lowest of any vehicle in FARS. Honda Accord: 3.07. Toyota Camry: 2.03. Chevy Silverado: 1.25. Even the Mazda CX-5, a perennial IIHS darling, sits at 0.12.^[2] The Model 3’s fatal crash frequency is 13 per 100,000 fleet vehicles. Dataset median: 174. The Cobalt hits 726.

FSD isn’t trying to improve upon Cobalt-level driving. It’s trying to improve upon Model 3-level driving—the lowest-risk baseline in the entire dataset.

Now, that 0.05 deserves an engineering autopsy before anyone awards it a trophy. Three confounders compress the number. First, fleet age: the oldest Model 3 in FARS is from 2017, while Accords carry death counts from 1990s-era cars lacking ESC, side curtain airbags, and AEB.^[3] Second, demographics: Model 3 buyers skew affluent, suburban, and sober—all independent predictors of not dying on the highway. Third, active safety: every Model 3 ships with collision avoidance systems that prevent crashes from reaching FARS in the first place.^[4]

Strip away confounders and the Model 3 still beats the field. But the 61x gap over the Accord is partially a measurement artifact, not a 61x engineering advantage.

The strongest case against FSD is Tesla’s own data. Their January 2026 robotaxi filings showed an autonomous crash rate roughly three times worse than human drivers under comparable conditions—even with a safety monitor present.^[5] If FSD degrades an already exceptional human baseline, nine crashes are a rounding error. The real stakes: 3.2 million of America’s safest-driven vehicles potentially made less safe by the software claiming to protect them.

FARS can’t separate FSD-engaged crashes from human-driven ones within the Model 3’s records. We don’t know whether the 0.05 rate is despite FSD or because of it. VMT estimates for EVs carry ±15–20% uncertainty. And FARS only captures the ~40,000 annual fatalities atop 6.7 million total crashes—the Model 3 could be generating plenty of non-fatal incidents this dataset will never see.

But the paradox holds: the fleet NHTSA is investigating for automation safety failures is, by every available fatality metric, the safest fleet on American roads. The real question is whether FSD is trying to improve a record that the humans behind its wheel have already nearly perfected.