Seventeen inches of rain fell on metro Atlanta in a single week last September. Overpasses became rivers. The kind of flooding that insurance actuaries classify as a once-in-fifty-years event arrived, by some local meteorologists’ estimates, for the third time in a decade. Most cars stopped. A handful of Waymo vehicles, operating in a geofenced corner of the city, stopped too — pulled from service by remote operators watching weather feeds from an office in San Francisco. The fare requests went unanswered. Nobody was hurt. And the insurance industry barely noticed, which is precisely the problem.
The Actuarial Bet That Feels Safe Until the Water Rises
The working assumption among insurers who have begun writing policies for autonomous vehicle fleets is that superior sensor data will eventually produce superior risk models. More data in, better pricing out. Waymo and its competitors generate millions of miles of operational telemetry — lidar returns, braking events, near-miss logs — that a human-driven fleet could never match. The pitch to underwriters is essentially this: autonomous vehicles don’t text, don’t drink, don’t glance at the passenger seat. Their failure modes are knowable, finite, and correctable through software. On that premise, a nascent but real market for autonomous vehicle liability coverage has begun to form, with specialized carriers and Lloyd’s syndicates cautiously writing policies against robotaxi fleets in Phoenix, San Francisco, and now Atlanta.
The assumption is seductive. It is also, in one specific and underappreciated way, likely wrong.
Weather Is Not a Sensor Problem. It Is a Coverage Problem.
The fragility isn’t in the vehicle’s ability to navigate rain. Waymo’s hardware can detect standing water; its operations team can withdraw vehicles from service when conditions exceed operational design domain. The fragility is in what happens to the insurance contract when the car isn’t there. Roughly 10 percent of all U.S. auto accidents are weather-related, a share that has grown alongside the frequency of extreme precipitation events. Traditional auto insurers price that risk into premiums and absorb claims when accidents happen. But autonomous vehicle liability operates on a different structural logic: the vehicle’s refusal to operate in dangerous conditions is supposed to be a safety feature, not a coverage gap. What no one has cleanly resolved is who bears the cost when a robotaxi withdrawal during a flood leaves a passenger stranded, causes a missed medical appointment, or pushes a rider into a less safe alternative.
Markel’s assessment of the top insurance trends heading into 2026 identifies autonomous and emerging vehicle technology as among the most structurally disruptive forces facing specialty insurers — not because the technology fails catastrophically, but because the liability boundaries remain genuinely unresolved. The question of whether an AV operator assumes duty of care the moment a passenger is matched to a vehicle, or only when the passenger is physically inside it, has not been litigated at scale. Atlanta’s floods are producing the conditions under which that question will be forced.
“The hardest thing to price in any new liability class isn’t the loss you can model. It’s the loss that emerges from a gap between what the product promised and what it delivered.”
— Senior underwriter, specialty transportation lines
Phoenix Gave Everyone the Wrong Confidence
There is a geography problem embedded in how autonomous vehicle liability has been understood so far. Waymo’s longest operational history is in Phoenix — a city where it rains fewer than eight inches per year, where roads drain quickly, and where the edge cases that break weather-dependent systems simply don’t arrive with the frequency or severity they do in the American Southeast. The dataset that convinced insurers that AV risk was manageable was built overwhelmingly in favorable conditions. Geofenced desert operations are not a neutral baseline. They are, from an actuarial standpoint, a best-case scenario that was quietly treated as a representative sample.
Atlanta changes the inputs. The city averages 49 inches of rain annually, more than Seattle, with the added complication of inadequate storm drainage infrastructure and a highway system that floods predictably and repeatedly. A Ceres analysis of U.S. property and casualty insurance results documented that rising losses from extreme weather events were already straining insurer models more than a decade ago — before the current acceleration in storm frequency. The AV industry walked into Atlanta carrying actuarial models built in the Sonoran Desert. The mismatch is not subtle.
Deloitte’s 2026 global insurance outlook flags climate volatility as a systemic pressure on insurance pricing accuracy, particularly in lines where historical loss data is thin and where new risk categories are still being defined. Autonomous vehicle liability sits at exactly that intersection: new enough that historical data is sparse, exposed enough to weather variability that Phoenix-era assumptions may be structurally insufficient.
What the Contract Actually Says When the Car Doesn’t Come
Passengers rarely read the terms of service before requesting a Waymo. The app presents a clean interface, a wait time, and a pickup pin. What it does not prominently surface is the operational design domain — the technical specification that defines the weather, road, and geographic conditions under which the vehicle will operate. When conditions exceed that domain and the vehicle is pulled, the rider receives a cancellation. The legal status of that cancellation, relative to autonomous vehicle liability frameworks currently being developed at the state level, is genuinely ambiguous.
California, Georgia, and Texas have each begun drafting or revising AV operating regulations, but none has produced a clean answer to the withdrawal-in-adverse-conditions question. The closest analog in existing law is common carrier liability — the doctrine that governs when a bus company or airline owes a duty to passengers it has accepted but not yet transported. Courts have not consistently applied common carrier standards to ridehail, let alone to autonomous ridehail. That gap is where the litigation will concentrate, and it is where insurers trying to price autonomous vehicle liability are currently flying without instruments.
Robots make some decisions faster than humans. Choosing not to operate during a flood is one of them. Existing legal frameworks are poorly equipped to handle liability decisions made at machine speed, without human deliberation, in real time. The operator in San Francisco who pulled the Atlanta fleet didn’t hold a passenger relations meeting. An algorithm compared sensor data against geofence parameters and generated a service suspension. Somewhere in that automated judgment lives a liability exposure that no existing policy form describes cleanly.
Five Assumptions Enter. One Survives Atlanta.
Advocates for the AV insurance model make five distinct arguments for why autonomous vehicle liability is ultimately more manageable than traditional auto liability. AVs don’t impair. AVs don’t fatigue. AVs generate auditable data trails. AVs will improve continuously through software updates. And AVs, by refusing to operate in dangerous conditions, will prevent the accidents that would otherwise generate claims. Four of those arguments are probably correct. The fifth — that non-operation in dangerous conditions is a liability-reducing feature — depends entirely on a legal and contractual architecture that does not yet exist. When the car doesn’t come because it is raining too hard, the absence of an accident is not the same as the absence of harm. A missed dialysis appointment is a harm. A stranded passenger who then drives a personal vehicle into flooded water because no alternative was available is a harm. The claim that AV withdrawal is uniformly safety-positive has not survived contact with a Southeast American summer.
Numbers alone don’t settle this. Ten percent of accidents being weather-related tells you something about frequency; it tells you almost nothing about the secondary liability chains that emerge when an autonomous system declines to operate and a human makes a worse choice in its absence.
The Reinsurance Market Is Watching, Not Writing
Behind every specialty insurer writing autonomous vehicle liability coverage is a reinsurer pricing catastrophic exposure. Conversations with people who work in that market — not the primary carriers optimistic about telemetry data, but the firms pricing tail risk — suggest a more cautious posture. Reinsurers have watched property catastrophe models fail repeatedly when climate patterns shifted faster than actuarial tables could update. They are not eager to repeat the experience in a new liability class. The result is that primary AV insurers are retaining more risk than their pricing models may justify, and the cost of that retained exposure has not yet been discovered because the weather events that will trigger it have not yet produced a clean test case with sufficient scale to force a coverage dispute to litigation.
Atlanta is building that test case slowly, one flood at a time.
The next major precipitation event in a city where Waymo operates at meaningful scale will not just be a weather story.
FetchLogic Take
Within 36 months, at least one U.S. state will face a court ruling or regulatory determination that an AV operator bears partial liability for passenger harm following a weather-related service withdrawal — not because the withdrawal was wrong, but because the duty of care attached at the moment of acceptance, not at the moment of pickup. That ruling will force a rewrite of autonomous vehicle liability policy language industry-wide, trigger a measurable increase in AV insurance premiums in high-precipitation markets, and cause at least one major reinsurer to exclude weather-withdrawal events from standard treaty coverage. The carriers currently pricing AV risk on Phoenix data will discover, expensively, that they priced a climate that no longer exists.
Related Analysis
Memory Now Costs More Than the Processor in AI Chips – and That Changes Everything About Who Can Afford to CompeteMay 25, 2026
The Chip Inside the Chip: How Memory Became AI’s Most Expensive HabitMay 24, 2026
Waymo’s Water Problem Exposes the Last-Mile Weakness in Autonomous VehiclesMay 16, 2026
AMD’s MI455 and the Quiet Redrawing of the AI Chip MapApr 10, 2026