A note to my loyal readers: thank you for your patience during my recent break. I am trying a new format with this post. Let me know what you think, and whether you would like to see more like this.
My dad and I were out walking when he asked whether we should be rolling out self-driving cars much faster. It seemed like a good question, so I spent some time looking into it. What I found surprised me in its clarity. The safety evidence is strong, the risks are manageable with the right oversight, and the cost of moving slowly is not abstract: it is roughly 47 preventable deaths every day on roads where the technology already works. This is what I learned, what I think, and what I would tell him about how to do it well.
My dad is not a policy wonk. He is a widely read, smart, curious person. A few weeks ago, while we were out walking, he asked me a good question. It was basically:
“Should we be rolling out self-driving cars much faster? If they’re safer, aren’t we letting people die by only allowing human drivers in so many places?”
So, I decided to look into it. And with a lot of help from AI, this is what I found.
Roughly 40,000 Americans die in car crashes every year. That works out to about 110 people a day. It has been roughly stable for five years. It is the leading cause of death for every American between the ages of 1 and 44. We have mostly stopped being shocked by it.
Waymo One, the furthest-along self-driving vehicle operator in the US, has now completed over 200 million fully driverless miles across Phoenix, San Francisco, Los Angeles, Austin, and a handful of newer cities. No safety driver was in the seat. Compared to human drivers on the same roads under the same conditions, Waymo’s vehicles are involved in:
82% fewer injury-causing crashes
92% fewer crashes causing serious or fatal injuries
83% fewer airbag-deployment crashes
That last number matters particularly. Airbag deployment is triggered by physical forces in a collision, not by how many people are in the car. So the reduction is not a statistical artifact of having fewer passengers. It reflects fewer serious crashes happening.
These figures come from crash data that federal law requires Waymo to report to NHTSA, benchmarked against human drivers on the same streets and traffic mix. Swiss Re, one of the world’s largest reinsurance companies, co-authored a separate analysis using insurance claims data and reached similar conclusions.
Doing the math, this means 47 people a day die in crashes on roads where self-driving vehicles already operate, in conditions where the safety data says those deaths were preventable.
The public health standard for deciding whether to act on an intervention like this is not “prove zero harm.” It is: demonstrate that expected benefits exceed expected harms with reasonable confidence, while continuing to monitor. By that standard, the evidence here is unusually strong. It is more rigorous than the evidence base behind most traffic safety changes implemented at scale. Roundabouts, speed cameras, rumble strips: none of them waited for 200 million miles of controlled observation before deployment.
I will come back to whether 200 million miles is actually enough data. The short answer is: it depends what question you are asking.
“The cost of moving slowly is not abstract: it is roughly 47 preventable deaths every day on roads where the technology already works.”
The data above makes the risks and benefits look favorable. But self-driving vehicles also introduce categories of risk that human driving does not have.
Some of the ones worth taking seriously are:
Systematic software bugs. A human driver who makes a mistake affects one situation. A software flaw pushed to an entire fleet could theoretically propagate the same error across thousands of vehicles simultaneously. This is qualitatively different from human error.
Cybersecurity. A remotely accessible fleet is an attack surface. Unauthorized interference with vehicle behavior is a different threat model than anything in conventional road safety.
Adversarial attacks on AI. Researchers have demonstrated that self-driving systems can be fooled by subtle visual patterns, stickers on a stop sign or markings on a road surface, that a human driver would ignore or not even notice, but that the vehicle’s perception system interprets as something else entirely, potentially causing it to brake, swerve, or fail to stop.
Unexpected behavior at the edges. As driving systems become more sophisticated, edge cases can produce behavior the developers did not anticipate.
These are legitimate concerns. They are also the concerns that serious operators have spent the most engineering effort addressing. For example, Waymo’s published safety framework describes a layered set of controls:
Simulation before streets. New software is tested across billions of simulated miles, including rare and adversarial scenarios, before touching public roads.
Staged rollouts. Updates are deployed incrementally with performance monitored at each stage, the same approach used in aviation and critical infrastructure, not the same as pushing an app update.
Geofencing. Each deployment is bounded by a precisely defined operational domain. Expansion beyond it requires a new formal safety determination.
Real-time monitoring. A fleet response team monitors vehicles continuously and can advise the system in edge cases, but cannot remotely drive. Steering, braking, and acceleration stay under autonomous control.
Crash reporting. Federal law requires disclosure of every significant crash to NHTSA, creating a public, independently verifiable record.
These controls do not make failure impossible. Although they make the catastrophic fleet-wide failure scenario considerably less plausible than the concern implies. The argument in this piece is that we should mandate more of these controls, not fewer, and move faster while doing so. On cybersecurity in particular, I think a federal standard is needed. The current voluntary approach is not sufficient for a technology operating at this scale.
Two more things that rarely make it into the news coverage:
Productivity. 160 million Americans drive to work, and 69% drive alone. The average round-trip commute is about 54 minutes a day, which adds up to roughly 212 hours a year per driver. That is time in which the person behind the wheel cannot read, write, think, or rest. Working from the median US wage of $23 an hour and assuming 25% of commute time is productively recoverable in a self-driving vehicle, the implied value across the full commuter fleet is roughly $134Bil a year. These are my rough calculations. The actual figure depends heavily on how much of that reclaimed time people would use productively versus resting or scrolling.
Road capacity. Research shows that even 10 to 20% self-driving vehicle penetration begins to measurably reduce stop-and-go cascades, the phantom traffic jams caused by variability in human reaction times. At higher rates, coordinated highway platooning can increase effective lane capacity by 50% or more. Smaller purpose-built vehicles could eventually reclaim urban land currently spent storing idle cars. Neither of these appears in the safety statistics. They are additional arguments for moving faster.
There are three regulatory barriers worth separating because they require different fixes.
Operational permission is the primary barrier to scaling what already works today. Waymo’s current vehicles are fully compliant with federal safety standards. They are blocked by a city-by-city permitting maze: every new market requires fresh negotiations, new geographic approvals, and compliance with locally inconsistent rules. More than 25 states introduced 67 new self-driving vehicle bills in 2025 alone, according to the National Conference of State Legislatures. A vehicle that works in Phoenix must prove itself again in Nashville.
Vehicle form factor is a separate and mostly future problem. Federal safety standards written in the 1960s mandate steering wheels, pedals, and defrosters: equipment that serves no purpose in a purpose-built driverless vehicle. For the personal commuter case, this is less of a near-term barrier, since vehicles built on a compliant platform with manual controls do not hit the exemption cap. It becomes a real constraint when manufacturers want to mass-produce purpose-built driverless vehicles without any manual controls, like a Tesla Cybercab. Exemptions for those vehicles are currently capped at 2,500 per manufacturer per year.
Regulatory asymmetry in crash response is the subtler barrier. A Waymo crash, including one caused by another driver running into a stationary Waymo at a red light, triggers NHTSA investigations and congressional hearings. A human driver killing a pedestrian, which happens more than 100 times a day, triggers a police report. We are not applying a consistent standard.
Several federal bills are currently pending.
The SELF DRIVE Act of 2026 would prohibit states from imposing their own self-driving vehicle performance standards on vehicles that have filed a safety case with NHTSA, and raise the annual exemption cap for vehicles without manual controls from 2,500 to 90,000. The preemption language addresses the operational permission problem. The cap increase matters for next-generation purpose-built vehicles.
The Self-Driving Vehicle Safety Data Act would require operators to report vehicle miles traveled, injury data, crash circumstances, and law enforcement interactions to NHTSA, creating the national safety database that currently does not exist. It also addresses emergency interoperability: first responders currently have no standardized protocol for interacting with a driverless vehicle at a crash scene, and some jurisdictions have used that gap as grounds to restrict operations entirely.
The Self-Driving Vehicle Acceleration Act directs the Secretary of Transportation to systematically fix the federal safety standards certification problem within one year of enactment, converting the current case-by-case exemption petition process into a defined regulatory pathway. Where the SELF DRIVE Act addresses deployment permission, this addresses the underlying standards architecture.
The highway bill (Surface Transportation Reauthorization), the once-every-five-years legislation that funds and governs the entire federal highway system, is the highest-stakes near-term opportunity. Self-driving vehicle provisions folded into reauthorization acquire must-pass momentum that standalone bills have repeatedly failed to achieve. The risk is that they also become subject to unrelated horse-trading.
States do not need to wait for Congress. Arizona and Texas have used existing authority aggressively. Waymo’s Phoenix operations and Aurora’s driverless Texas freight routes exist because those states chose to act. The lesson is that state-level permission is both a barrier and a potential accelerant, depending on which way it points.
The history lesson. The standard caution argument goes: we rushed the automobile and nearly four million Americans died over the 20th century. We should not repeat that mistake. But I think this frames the comparison incorrectly. Self-driving is not a new transportation mode we are deciding whether to adopt. It is a safety technology we are deciding whether to install. Think of it the way we think about seatbelts, airbags, or automatic emergency braking. Each of those was a system added to cars to reduce the death toll that human driving produces. Self-driving is the next step in that progression, and the most powerful one yet. We would not delay mandating airbags because airbags are imperfect. The question for self-driving is the same: does this safety technology demonstrably reduce deaths? The data says yes.
The systemic failure risk. One human driver makes one mistake. One software bug, pushed fleet-wide, could theoretically cause the same error across thousands of vehicles at once. This is a real and qualitatively different concern. But the answer it points to is mandatory crash reporting, independent software audits, and staged rollout protocols, not deployment bans. The controls already exist in outline. Taking the systemic risk concern seriously is an argument for more rigorous monitoring, not for the current regulatory stasis.
Induced demand. When driving gets easier, people drive more, potentially offsetting safety gains. And when a behavior gets safer, people may do more of it or take more risks. This is a real effect in transportation economics and human behavior. But an 80% per-mile safety improvement is large enough that even a doubling of total miles driven would leave the absolute death toll lower. A doubling of miles at one-fifth the fatality rate still produces 60% fewer deaths. Induced demand is an argument for careful transportation planning alongside deployment. It is not an argument against deployment.
Job displacement. Commercial vehicle autonomy, covering trucking, delivery, and ride-hailing, involves real displacement of real workers. This deserves a direct answer, not a sidestep. Waymo’s primary current business is ride-hailing, which puts it in direct competition with the roughly 2 million Americans driving for Uber and Lyft. That displacement is real and near-term, and the people affected have legitimate cause for concern. The argument for focusing first on personal passenger vehicles is precisely that it separates the safety benefit from the labor disruption: a person driving their own car to work is not a gig worker’s livelihood. That does not make the ride-hailing labor question disappear. It means we should handle it explicitly, with transition support and sequenced deployment, rather than treating it as a reason to block the whole technology.
Who is liable when something goes wrong? This is a real gap in the current framework, and it matters both for fairness and for scale. With a human driver, liability is reasonably well understood: the driver, their insurer, and sometimes the manufacturer. With a driverless vehicle, the picture is murkier. Is it the software developer? The vehicle manufacturer? The fleet operator? The company that mapped the road? Right now the answer varies by state, is often unresolved, and is working its way through courts case by case. That is not a sustainable arrangement for a technology that will eventually operate millions of vehicles. A federal liability framework, one that clearly assigns responsibility to the operator or developer when the autonomous system is engaged, and does so in a way that channels disputes toward efficient resolution rather than prolonged litigation, is a prerequisite for responsible deployment at scale. It also matters for public trust. People are more willing to accept a new technology when they know that if something goes wrong, someone is clearly accountable. The absence of that clarity is itself a barrier to adoption.
Privacy. A self-driving vehicle is a rolling sensor array, filming streets, sidewalks, and the people around it continuously. At the scale of millions of vehicles, that becomes a de facto surveillance infrastructure. Law enforcement has already sought data from vehicle manufacturers in criminal investigations, and the legal framework governing what must be handed over is unsettled. The fix is data minimization: mandate that operators delete surrounding environmental sensor data within a short defined window, perhaps 24 to 72 hours, unless directly relevant to a crash under active review, and prohibit its use for purposes unrelated to vehicle safety without explicit consent. Getting this right early matters. Once data exists, it gets used in ways no one originally intended.
Mixed traffic and the bullying problem. Human drivers learn quickly that self-driving vehicles are programmed to yield, and they exploit it: cutting them off, merging aggressively, knowing the robot will back down. Researchers call it the compliance problem, and it is real enough that Waymo has had to tune its driving behavior specifically to avoid being taken advantage of at intersections. The transition period, where self-driving and human drivers share roads in significant numbers, will have its own friction. Traffic may actually worsen in some contexts before it improves, because the two populations operate on different social contracts. The network effect that makes a fully autonomous fleet safer runs in reverse during the awkward middle phase. Plan for it, and do not measure success only against the fully autonomous end state.
Is 200 million miles enough? Measured against total US light-duty passenger vehicle miles on urban and suburban roads, roughly 6 billion miles a day, 200 million miles is less than an hour’s worth of collective fleet driving. That framing is humbling. It does not mean the data is insufficient for measuring crash rates, which require far less exposure to produce statistically meaningful results, especially given reductions as large as 80 to 90%. What it means is that truly rare tail-risk events, scenarios that occur once in a billion miles, have probably not been encountered yet. This is an argument for mandatory ongoing monitoring and staged expansion, not for waiting until the fleet has driven a billion miles before starting.
Dad, here is my answer.
The evidence is strong enough. Not perfect, but strong enough that continued delay has a real and calculable cost. Roughly 47 people die every day in urban and suburban crashes that using the self-driving vehicles we know how to build could prevent. This is not a wild guess. It is the straightforward application of Waymo’s demonstrated safety record to the share of US traffic deaths that occur in conditions where self-driving vehicles currently operate.
The case for moving faster rests on three things. The safety evidence is unusually robust by the standards we apply to other public health interventions. The risks that distinguish self-driving vehicles from human drivers are real but manageable with the right monitoring infrastructure. And the cost of delay is real.
What I would do, if I were setting policy:
Remove federal barriers to cities and states that want to permit self-driving vehicle operations, rather than requiring each jurisdiction to build its own approval process from scratch.
Mandate cybersecurity standards and independent red-team audits, including adversarial testing of perception systems, for any operator deploying at scale.
Create an improved near-time crash and performance data system with mandated use that feeds a public human and computer readable database.
Establish a clear federal liability framework so that when something goes wrong, responsibility is assigned efficiently and fairly, without leaving victims to navigate a patchwork of state courts.
Pair any deployment authorization with federal privacy mandates requiring strict data minimization, so that the infrastructure we build to make roads safer does not become a surveillance system by accident.
Recognize that the transition period, where self-driving vehicles and human drivers share the road, will create its own problems before it solves them, and build that expectation into deployment plans rather than treating it as a surprise.
Treat this as a current policy choice, not a future technology question. I spoke with a state official who thought we were decades away from having to think about this. We are not. The technology exists. The question is whether we act on it.
I am genuinely uncertain about how much to centralize: whether federal preemption of state standards is the right move or whether letting states lead creates useful variation and appropriate local control. I lean toward clearing the path for cities that want to move, rather than forcing all of them to. But I hold that view loosely.
What I am not uncertain about is this: the failure to act is not inherently safer. People are dying in crashes every day that the evidence suggests we could prevent. That should weigh on us more than it does.
I hope that answers your question, Dad.
Intellectual Disclosure. This piece is an experiment in collaborative thinking and writing. What began as a conversation during a walk with my father was researched and drafted primarily by Claude, based on my detailed outline, then fact-checked and edited by Gemini (which is under common ownership with Waymo). I am sharing it in the spirit of seeking truth through the constant testing and refinement of ideas.
Financial Disclosure. I am invested in various diversified index funds and hold a minor individual position in Alphabet Inc. (GOOGL) which owns Waymo.
Identity Disclosure. I am just a person on the internet with no formal expertise in transportation policy or autonomous systems. Please do your own thinking and verify these claims independently.
Personal Note. While I have ridden in Waymos, my wife does not currently trust Tesla’s software enough for me to use their self-driving features. Perhaps the rollout of this technology won’t ultimately be limited by government policy, but by the quiet “veto power” of the people closest to us and what they believe about the risks.
What do you think? I would love to hear from readers, whether you think I have got this right, wrong, or have missed something important.
Waymo Safety Impact update, March 19, 2026, covering 170.7 million rider-only miles through December 2025. Published figures: 92% fewer serious-or-fatal-injury crashes, 83% fewer airbag-deployment crashes, 82% fewer any-injury crashes versus human benchmarks in the same cities. Waymo also estimates its fleet is preventing approximately one serious-injury crash every 8 days at current scale. Underlying peer-reviewed methodology: Kusano et al. (2025), Traffic Injury Prevention, 26(sup1):S8-S20.
Di Lillo, L. et al. (2024). “Do Autonomous Vehicles Outperform Latest-Generation Human-Driven Vehicles? A Comparison to Waymo’s Auto Liability Insurance Claims at 25.3M Miles.” Waymo LLC / Swiss Re. Announcement.
Total US VMT (2023): 3.247 trillion miles, per FHWA Table VM-1, Highway Statistics 2023. All light-duty vehicles (passenger cars, light trucks, vans, SUVs): 2.879 trillion miles, representing 88.7% of total VMT. Urban light-duty VMT: 2.039 trillion miles, representing 70.8% of light-duty VMT. Daily urban light-duty miles: 2.039 trillion ÷ 365 = approximately 5.59 billion miles/day. Waymo’s 200 million miles ÷ 5.59 billion = approximately 0.036 days, or roughly 52 minutes of equivalent fleet driving. Claude says this is source: FHWA Table VM-1, Highway Statistics 2023.
The 47 deaths figure is derived as follows: NHTSA FARS reports 40,990 US traffic fatalities in 2023, or roughly 40,000 rounded. FHWA data show that approximately 46% of highway fatalities occur on rural roads, meaning approximately 54% (roughly 21,600 deaths using the rounded base) occur on urban and suburban roads comparable to where self-driving vehicles currently operate. Applying Waymo’s peer-reviewed 80% reduction in injury-causing crashes: 21,600 × 0.80 = roughly 17,300 preventable deaths per year, or roughly 47 per day. The unscoped figure of 88/day (80% of all 40,000 fatalities) represents the upper bound if self-driving technology were deployed across all driving conditions nationwide. Claude says that source is: NHTSA FARS: nhtsa.gov/research-data/fatality-analysis-reporting-system-fars; NHTSA Traffic Safety Facts: Rural/Urban Comparison (rural fatality share); Kusano et al. (2025), Traffic Injury Prevention, 26(sup1):S8-S20.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.