Robotaxis are rapidly expanding, and the industry’s most frequently repeated claim is that they are 'safer than human drivers.' Yet this year’s publicly released figures contradict one another: one report claimed a company’s collision rate halved in a year, but after recalculating under a unified methodology, the drop was only 16%. The problem may not lie in whether the vehicles improved or deteriorated—often, the measuring stick changed first. Five pieces of evidence, released progressively this year and intensively from October onward, examine this issue from four distinct angles: the most rigorous methodology, traps hidden in federal data, mid-course changes to reporting rules, and how the industry itself views the matter.
## The Most Robust Positive Evidence—and Its Cost
First, consider the evidence supporting the 'safer' claim. This year, the U.S. Insurance Institute for Highway Safety (IIHS) published a research report that undertook painstaking work: manually reviewing every federally mandated autonomous vehicle collision report filed between July 2021 and December 2024, filtering out incidents that did not constitute collisions or occurred when autonomous mode was inactive, leaving 736 cases; then assessing whether each incident met the threshold for police reporting—because the federal reporting threshold is far lower than the police-reporting threshold, many minor scrapes are included. The result: only about 22% were definitively or possibly serious enough to warrant a police report.
Under this stricter definition, **Waymo’s robotaxi police-reportable collision rate is 68% lower than that of human drivers in the same regions and years**; its rear-end collision rate (as the striking party) is 91% lower, and its rate of being rear-ended is 40% lower; autonomous vehicles rarely initiate collisions and are seldom the primary contributing factor to accidents. Within the time window covered by this study, none of the analyzed autonomous vehicle collisions involved fatalities.
This is positive evidence grounded in methodological rigor. Yet IIHS itself acknowledges its limitations: manual coding of accident narratives becomes unsustainable as fleet size and incident volume grow; and the conclusions apply only to the specific set of autonomous vehicles studied.
## Three Traps Hidden in Federal Data
Is there an alternative approach—one that avoids manual coding and does not require companies to disclose mileage? A preprint posted on arXiv at the end of October and updated in November adopted a different strategy: instead of calculating collision rates, it examined **composition**—given that an accident occurred, does the distribution of accident types involving autonomous vehicles match that of human drivers? Because this question does not require mileage as a denominator, it can cover all operators. Researchers used a language model to code 2,450 federally reported incidents (from 2021 to 2026) into the accident categories used in nationally sampled police reports, comparing the 828 incidents overlapping with the human-driver sample period.
The compositional differences are clear: **accidents involving turning across traffic paths occur significantly less often in autonomous vehicle incidents** (only 0.23–0.46 times the human rate), and **single-vehicle accidents also occur significantly less often** (0.19–0.56 times the human rate); in rear-end collisions exceeding the researchers’ predefined severity threshold, autonomous vehicles’ relative likelihood of being the striking party is only 0.121 times the human baseline (95% confidence interval: 0.05–0.23). In other words, accident types more closely associated with active driving errors occur at lower proportions—consistent with the IIHS findings.
But the paper’s more valuable contribution lies in exposing three data traps: first, the determination of 'autonomous driving' is self-coded by operators, with inconsistent definitions across companies; second, severity markers such as 'towed away' are conflated with companies’ own fleet recovery practices; third—and most critical—the paper terms this 'occupant confounding': **about 42% of the apparent difference in accident severity between humans and autonomous vehicles can be explained by 'who is sitting in the vehicle'**: robotaxis have no driver in the driver’s seat, inherently eliminating one potential injury victim. Using such data directly to compare 'who is more dangerous' means the ruler itself is skewed.
## The Ruler Changed Midway
The third angle is the most concrete. On June 16, 2025, the U.S. National Highway Traffic Safety Administration (NHTSA) implemented its third revision of the crash reporting rule: companies no longer need to report incidents where an autonomous vehicle is struck, property damage does not exceed $1,000, and no injuries occur.
On November 13, an independent analysis report published on an open research platform re-calculated ten publicly cited figures from over a year following that change, standardizing them under a single consistent rule. The result: media reports claiming 'Waymo’s California collision rate dropped from 13.7 per million miles in 2024 to 6.7 in 2025'—nearly halved—shrank to just a 16% decline when restricted to only those severe incidents that must be reported regardless of rule changes (e.g., tow-away, injury, airbag deployment, or involvement of pedestrians or cyclists). **Roughly three-quarters of the 'improvement' stems from bureaucratic effects of the rule change.** Another widely circulated comparison of Tesla versus Waymo collision intervals, based on two separate reporting rules, yielded a ratio of 1.6; under a unified rule, the actual ratio is approximately 2.9. The report itself notes its limitations: third-party mileage figures cannot be independently verified, and the ten cited figures were identified via web search—not an exhaustive census.
What do federal filings look like post-rule-change? A statistical brief released by a research institute on November 1 shows a cross-section: from June 2025 to September 2026, 1,551 initial reports were submitted, ~83% from a single company; 1,295 involved property damage only, and 2 claimed fatalities. The brief repeatedly emphasizes one sentence: **'Reports are not liability determinations, nor can they be converted into collision rates'**—without mileage data, collision rates simply cannot be calculated.
## What the Industry Itself Says
The most interesting angle is the fourth: industry leaders actually agree on the most critical point—that the data infrastructure cannot support definitive conclusions. In January this year, Waymo’s safety team submitted a paper to the International Conference on Enhanced Vehicle Safety (ESV), stating plainly: public reporting data lacks mileage denominators, so **collision rates for autonomous vehicles cannot currently be calculated at all**; roughly twenty comparative studies over the past decade have all used vehicle mileage as the denominator, and the literature contains no viable alternatives. They recommend: developers voluntarily disclose mileage and supplementary collision data (aggregated and anonymized); studies should align temporal and geographic scopes; benchmarks should be pre-registered to prevent bias; and external oversight—including peer review and independent researchers’ access to data—is essential. The paper also acknowledges: to the authors’ knowledge, only Waymo itself has thus far disclosed such data (starting from late 2024).
## My Assessment
Synthesizing these four angles, my assessment is:
First, the directional conclusion that 'autonomous vehicles are safer than humans' is supported by two methodologically independent lines of evidence: strict-denominator rate comparisons (68% lower) and denominator-free compositional analysis (fewer accidents tied to active driving errors)—both pointing in the same direction. Yet all specific **multipliers** have been contaminated by measurement artifacts—reporting rule changes, occupant confounding, and inconsistent coding definitions—two of which have already been quantified as causing drifts of several tens of percentage points. Trust the direction; distrust the multipliers.
Second, the data pitfalls are not abstract—they are three concrete, nameable mechanisms: mid-course reporting rule changes, ~42% occupant confounding, and operator-specific definitions of 'autonomous driving.' Any comparison using federal data must first ask whether these three elements have been aligned.
Third, the true bottleneck is data infrastructure: the federal level lacks mandatory mileage reporting requirements, making collision-rate calculation impossible without voluntary corporate disclosure—which carries inherent commercial incentives. Manual coding is unsustainable; the agency projected receiving 5,425 reports annually per the independent report’s calculation but received only 1,336—both symptoms of the same underlying problem.
Fourth, what’s encouraging is emerging cross-sector consensus on 'how to evaluate': independent researchers, the insurance industry, and operators themselves all demand public disclosure of denominators, standardized definitions, and external oversight. Before resolving whether autonomous vehicles are 'safe,' agreeing on 'how to measure' is the most actionable step in this debate.
A question for readers: Should the federal government mandate mileage reporting to enable anyone to recalculate collision rates? And next time you see news headlined 'collision rate halved,' should you first ask—what version of the reporting rule was used?
Primary source: https://arxiv.org/abs/2610.42121