Ashok has noted that Robotaxi maintains an impeccable safety record across the 380,000 unsupervised miles driven so far. Yet many observers remain concerned that a serious crash is only a matter of time.
That concern is not irrational. As the service scales and cumulative mileage grows, the first major incidents become statistically inevitable. The relevant question is not whether they will occur, but when, and at what severity.
In a previous analysis, we modeled the likely Robotaxi mileage trajectory using Ashok’s stated week-over-week growth rate and the installed Cybercab production capacity. Combining that projection with publicly available NHTSA data allows us to estimate the probability of Robotaxi involvement in crashes of different severities over time.
One clarification is important: the figures that follow are involvement rates, not pure at-fault rates. They describe the likelihood that a Robotaxi will be present in a crash that produces an injury or fatality, regardless of which party is primarily responsible. This provides a realistic baseline against which future safety performance can be judged.
Part 1 – The Human Baseline: Road Safety Today
We begin with the most recent detailed national statistics from the National Highway Traffic Safety Administration (NHTSA).
In 2023, passenger vehicles accounted for roughly 89% of all vehicle miles traveled in the United States. For occupants of these vehicles, the rates were:
In absolute terms, this corresponds to approximately 24,000 occupant fatalities and 1.95 million police-reported injuries across nearly 2.9 trillion passenger-vehicle miles.
Police-reported injuries are classified on the KABCO scale:
For the remainder of this analysis, we adopt the “all passenger vehicles” rates as the human baseline (0.83 fatalities and 68 injuries per 100 million miles). We then assume that a mature Robotaxi system will be 10 times safer than this human average — an ambitious but commonly referenced target in the industry (and one that Waymo has also used as a benchmark in its own safety reporting).
This 10× assumption becomes the foundation for the statistical projections that follow.
Part 2 – Why Incidents Follow a Poisson Process
Traffic crashes are rare events relative to the enormous number of miles driven. When events are infrequent, independent, and occur at a roughly constant average rate, the number of occurrences in a given exposure (here, miles) is well described by a Poisson distribution.
Under a Poisson process, the probability of observing exactly k events after M miles is:
where λ is the expected number of events per mile.
We are primarily interested in the probability of experiencing at least one event. That is simply the complement of the probability of zero events:
How to obtain λ
From the NHTSA data in Part 1, we already know the average number of miles between events for human-driven passenger vehicles. The event rate per mile is therefore the reciprocal:
Equivalently, if the published rate is given per 100 million miles:
For our baseline (all passenger vehicles):
- Fatality: λ_human = 0.83 / 10⁸ = 8.3 × 10⁻⁹ per mile
- Injury (all severities): λ_human = 68 / 10⁸ = 6.8 × 10⁻⁷ per mile
Because we assume Robotaxi is 10 times safer, we simply divide these rates by 10:
Substituting into the earlier expression then yields the probability of at least one incident after any cumulative mileage M:
This is the formula used for all projections in the next section.
Part 3 – When the First Incidents Become Probable
We can now apply the Poisson formula from Part 2 to the four outcome categories of interest:
- Fatality (K)
- Serious / Incapacitating injury (A)
- Minor / Non-incapacitating injury (B)
- Possible injury (C)
Using the human baseline rates from Part 1, scaled by the 10× safety assumption, we obtain a separate event rate λ for each category. The first view plots the probability of at least one event against cumulative unsupervised Robotaxi miles:
From mileage to calendar time
In our previous article, we constructed a forward model of weekly unsupervised miles using Ashok’s stated double-digit week-over-week growth rates, constrained by the Cybercab production capacity of ~2,400 vehicles per week (Tesla’s 125k annual guidance) and a high-utilization assumption of 200 miles per vehicle per day.
Taking the 14% growth case as the central scenario and accumulating the capped weekly miles produces a cumulative-mileage trajectory over calendar time. Substituting that cumulative mileage into the Poisson formula yields the probability of each incident type as a function of time:
Under this trajectory, the 50% probability thresholds are reached approximately as follows:
Waymo’s current cumulative mileage
For scale, we have marked Waymo’s approximate cumulative fully driverless miles on the mileage chart.
Waymo’s official Safety Impact dashboard reports 220.6 million rider-only miles through March 2026. At a recent run-rate of roughly 4 million rider-only miles per week, extrapolation to mid/late 2026 yields approximately 280–300 million paid miles. Accounting for empty repositioning and staging miles, we conservatively double that figure to arrive at an estimated ~600 million total autonomous miles.
Waymo’s real-world track record
Public SGO / NHTSA reporting on Waymo incidents, mapped onto the KABCO scale, gives the following approximate picture:
Key observations:
- The large majority of reported injury cases fall into the lower-severity categories (C and B).
- Waymo’s own safety research treats only confirmed police-reported A or K cases as “serious injury or worse.”
- Across more than 220 million rider-only miles, Waymo reports approximately 94% fewer serious-injury-or-worse crashes and ~82% fewer injury crashes of any severity than the human benchmark in the same operating areas.
- These comparisons are made regardless of fault. In the small number of serious (A) and fatal (K) incidents, available public information indicates that the Waymo vehicle was typically not the primary at-fault party.
This real-world record illustrates the distinction that matters for Robotaxi as well: at sufficient scale, involvement in crashes becomes statistically inevitable. The decisive question is whether the system is at fault less often — and produces fewer severe outcomes — than human drivers covering the same miles. The Poisson framework developed here simply calibrates when Tesla Robotaxi, under a 10× safety assumption, can be expected to begin encountering the same categories of events.
As the cumulative-mileage graph shows, once a driverless fleet reaches several hundred million miles, involvement in serious and even fatal incidents becomes statistically inevitable — even under a 10× safety assumption. Waymo’s record is therefore consistent with what the Poisson model predicts at that scale: the relevant question is no longer whether such events occur, but how often the system is at fault and how its severity profile compares with human drivers.
Conclusion
We have modeled the probability of Robotaxi involvement in crashes of different severities, assuming the system ultimately achieves a 10× safety improvement over the average human-driven passenger vehicle. Combining that assumption with the mileage trajectory developed in our previous analysis yields clear, quantitative expectations: lower-severity injuries become probable in the first half of 2027, serious injuries later that year, and a fatality toward the end of 2027 under the central growth case.
Waymo’s real-world record is consistent with what the same model predicts at several hundred million cumulative miles. Even a system that is substantially safer than human drivers will, at sufficient scale, be involved in serious and fatal crashes. The statistics make this inevitable.
Two distinctions remain essential. First, these projections concern involvement, not fault. Second, the first at-fault fatality is also statistically inevitable eventually; the only open variable is when it occurs relative to the public’s lived experience of the technology.
This is the core strategic challenge for Tesla. The first at-fault death must arrive late enough that a large number of people have already used Robotaxi, trust its ordinary performance, and understand its safety relative to human driving. The analogy is commercial aviation. Modern airliners still kill people, yet the fatality rate is on the order of 0.01 deaths per 100 million passenger-miles — roughly an order of magnitude lower than even a 10×-safer Robotaxi (≈ 0.08 deaths per 100 million miles). Society accepts residual aviation risk because the benefit is large and the relative safety is well established. Robotaxi will need the same foundation of demonstrated, large-scale performance by the time the first clearly at-fault fatality occurs.
That is the threshold Tesla must cross.
