Every year, roughly 40,000 Americans die in traffic crashes — almost all of them caused by human error — and yet the technology that could meaningfully reduce that toll is being blocked, delayed, and legislated against, often by people who frame their opposition as a safety concern while the actual safety data points in the opposite direction.

At a Glance

• Waymo’s autonomous vehicles show 80% fewer injury-causing crashes and 92% fewer serious injury crashes than human drivers over 127 million miles of operation.
• Real documented failures — red-light running, flooded-street entries, school bus violations — are genuine and warrant regulatory scrutiny, but they don’t overturn the aggregate safety advantage.
• The core dispute is a methodological one: how you define and compare crash categories determines whether autonomous vehicles look better or worse than human drivers.
• Political opposition from labor groups and local governments is explicitly about job preservation, not safety — a distinction that matters enormously when lives are the cost of delay.
• The regulatory framework governing autonomous vehicles remains dangerously underdeveloped, with NHTSA relying on after-the-fact recall authority rather than prospective safety certification.

What the Safety Data Actually Shows

Waymo’s September 2025 Safety Impact Report — drawn from over 127 million fully autonomous miles, the equivalent of more than 150 human driving lifetimes — reports an 80% reduction in injury-causing crashes and a 92% reduction in serious injury or worse crashes compared to human drivers operating in the same cities. The numbers for vulnerable road users are similarly striking: 92% fewer pedestrian injury crashes, 83% fewer cyclist injury crashes, and 80% fewer motorcycle injury crashes. A Swiss Re insurance analysis covering 25 million Waymo miles found an 88–92% reduction in insurance claims relative to human-driven vehicles. These are not marginal improvements. They represent a categorical difference in outcomes.

The counter-data deserves honest engagement. APM Research Lab’s analysis of NHTSA Standing General Order data found that Waymo’s total crash rate — 4.5 crashes per million miles — exceeds the human driver benchmark of 1.94 per million miles. Damfirm.com’s aggregation of NHTSA crash reports tallied 1,429 accidents between July 2021 and November 2025, yielding 117 injuries and 2 fatalities. CNN’s review of public records documented specific alarming incidents: red lights run, flooded streets entered, emergency road closures ignored, intoxicated passengers requiring 911 welfare checks. NHTSA opened a formal investigation into 22 Waymo incidents, including 17 crashes and 5 unsafe non-crash behaviors such as driving in opposing lanes. These facts are real and they matter.

The methodological gap between these two pictures is the crux of the entire debate. Waymo’s figures measure injury-causing crashes — events where someone was actually hurt. The APM and NHTSA figures count all reported incidents, including minor contact events, property-damage-only fender-benders, and the kind of low-speed nudges that would never appear in human driving statistics because humans don’t face a mandatory federal reporting requirement for every incident. Autonomous vehicles operating under NHTSA’s Standing General Order must report any crash involving airbag deployment, injury, fatality, or a vulnerable road user — a reporting threshold that captures incidents human drivers routinely never disclose. The result is a comparison that is not apples to apples; it is apples to a fruit that has never been weighed on the same scale.

The Methodology Problem at the Heart of the Debate

This reporting asymmetry is not a trivial technicality — it is the single most important factor distorting the public’s understanding of autonomous vehicle safety. Human drivers self-report, or don’t. Police reports capture only a fraction of incidents. Insurance claims, the closest analog to comprehensive incident tracking, show Waymo performing dramatically better than human-driven vehicles. The Swiss Re data is particularly significant here because insurance actuaries have strong financial incentives to get the risk picture right, and they are not subject to the same promotional pressures as Waymo’s own safety team.

That said, the critics who flag Waymo’s per-mile fatality rate raise a legitimate concern that the company’s own framing tends to obscure. Grouping “serious injury or worse” crashes into a single category can mask how the fatality sub-component specifically compares to human drivers. An independent, standardized reanalysis using uniform injury and fatality definitions — the kind that NHTSA could conduct but has not yet required — would resolve this dispute definitively. The absence of that analysis is a regulatory failure, not evidence that Waymo’s claims are false.

A peer-reviewed matched case-control study published in Nature Communications found that vehicles equipped with Advanced Driving Systems generally show a lower probability of accidents than human-driven vehicles in most scenarios — but perform worse at dawn and dusk (5.25 times higher accident rate) and during turning maneuvers (1.98 times higher). This is exactly the kind of granular, scenario-specific analysis the public debate needs more of, and it illustrates that “autonomous vehicles are safer” and “autonomous vehicles have specific failure modes requiring attention” are both simultaneously true.

The Political Opposition and What It’s Really About

New York State Senator Luis Sepulveda has proposed legislation to block driverless vehicles from operating in New York. Toronto Mayor Olivia Chow has pushed back against autonomous taxi deployment in her city. The explicit rationale in both cases centers on protecting driving jobs — Sepulveda’s office has cited over 100,000 professional drivers in New York whose livelihoods would be disrupted. That is a legitimate policy concern about economic transition. It is not a safety argument, and presenting it as one is a category error with lethal consequences.

The pattern here is historically familiar. Longshore workers fought container shipping. Typesetters fought desktop publishing. Taxi medallion holders fought Uber. In each case, the incumbents were right that their jobs were at risk, and wrong that the technology should therefore be suppressed. The difference with autonomous vehicles is that the technology in question doesn’t just create economic disruption — it has a measurable, documented capacity to prevent deaths. Delaying deployment to protect jobs is, in effect, a policy choice to accept preventable fatalities. That trade-off should be named honestly rather than laundered through safety rhetoric.

The union pressure around Waymo has taken forms that go beyond conventional lobbying. Reports of organized groups demanding “job training compensation” from Waymo as a condition of political non-opposition — described by some observers as a protection racket dynamic — illustrate how economic interests can distort the regulatory environment in ways that have nothing to do with the public interest in road safety.

Genuine Safety Gaps That Demand Real Answers

Acknowledging Waymo’s aggregate safety advantage does not require dismissing its documented failures. The school bus incidents are serious: Waymo vehicles failed to yield to loading or unloading school buses more than 24 documented times, triggering a formal investigation. The May 2025 recall of 1,212 vehicles for a software bug causing collisions with roadside barriers demonstrates that over-the-air software updates — one of autonomous vehicles’ theoretical advantages — also introduce the possibility of fleet-wide simultaneous failure modes that have no analog in human driving. A single software flaw can degrade every vehicle in a network simultaneously; a drunk human driver degrades only one.

The teleoperator question is similarly unresolved. Waymo vehicles rely on remote human operators who can intervene when the system encounters situations it cannot handle autonomously. These operators are largely unregulated: their certification requirements, response-time standards, geographic location, and the frequency and nature of their interventions are not publicly disclosed in any standardized form. This is a genuine gap. If a teleoperator in an overseas call center is making safety-critical decisions about vehicles operating on American streets, the public has a right to know the competency standards governing that role.

The operating-environment limitation is also real, if frequently overstated. Waymo currently operates predominantly on urban streets at lower speeds, not on highways where the bulk of American vehicle miles are driven. The safety comparison with human drivers is therefore not a full-spectrum comparison — it is a comparison within a specific, relatively controlled operating domain. Waymo’s performance in that domain is excellent; what happens when the operational design domain expands to higher-speed, higher-complexity environments remains an open empirical question.

The Regulatory Vacuum and What It Costs

The Brookings Institution has documented what is perhaps the most important structural problem in this entire landscape: there is currently no federal requirement for autonomous vehicle companies to demonstrate that their vehicles are reasonably safe before deploying them on public roads. NHTSA’s posture has been reactive — issuing recalls and investigations after incidents occur — rather than prospective certification. This is not how aviation safety works, not how pharmaceutical approval works, and not how any other safety-critical technology with direct public exposure is regulated.

The irony is that this regulatory vacuum actually undermines public trust in autonomous vehicles more than any specific incident does. When people cannot point to a credible, independent authority that has reviewed and certified the technology, they fall back on fear and anecdote. The incidents that make headlines — the flooded-street entries, the red lights run, the ambulance blockages — fill the credibility vacuum that a robust regulatory framework would otherwise occupy. Building that framework is not an argument against autonomous vehicles; it is the prerequisite for their legitimate, trusted, and ultimately life-saving deployment at scale.

The AAA found that public fear of self-driving cars rose from 55% to 68% between 2021 and 2023, while trust in the technology fell from 15% to just 9% over the same period. That collapse in public confidence happened alongside the accumulation of genuine incident data — but also alongside a media environment that amplifies individual failures while the aggregate safety improvement, distributed across millions of uneventful rides, generates no headlines at all. Forty thousand Americans dying in human-caused crashes every year is not news. One Waymo running a red light is.

The Actual Stakes of Getting This Wrong

The moral arithmetic here is uncomfortable but inescapable. If Waymo’s injury-crash rate of 0.6 per million miles holds at scale against the human driver rate of 2.8, and if autonomous vehicles were to displace even a fraction of human-driven miles, the lives saved would number in the thousands annually — not as a projection or a hope, but as a straightforward extrapolation of the existing data. Every year of delayed deployment, every legislative block, every regulatory moratorium imposed for reasons that turn out to be economic rather than safety-driven, has a body count attached to it. That body count is invisible because the deaths occur in the counterfactual — in the crashes that would have been prevented by a technology that wasn’t allowed to operate.

This does not mean autonomous vehicles should be deployed without oversight, without transparency, without independent audit of crash data and teleoperator logs, and without standardized safety certification. It means those things should be built urgently and in good faith, not used as indefinite delaying mechanisms by parties whose primary interest is protecting incumbent industries. The difference between a rigorous safety regulator and a captured one is whether it demands evidence before permitting deployment — or demands impossibly high standards of proof specifically to prevent deployment from ever occurring.

The technology is not perfect. The data is not complete. The regulatory framework is inadequate. All of that is true, and all of it is fixable. What is not fixable is the harm done by decades of delay while the perfect becomes the enemy of the demonstrably better.

Sources:

reason.com, aitkenlaw.com, ark-invest.com, waymo.com, reddit.com, damfirm.com, wshblaw.com