2026-06-18 👁️ — views

AV Insurance and Liability — Who Pays When an Autonomous Vehicle Crashes

When a Waymo crashes, Waymo pays. When a Tesla FSD crashes, it depends — and that liability gap shapes every deployment decision both companies make.

Article 107 in the Physical AI Benchmark Series — AV Insurance and Liability: Who Pays When an Autonomous Vehicle Crashes, How Commercial AV Insurance Works, and Why the Liability Framework Shapes Tesla and Waymo’s Deployment Strategies Differently

When a Waymo driverless vehicle causes an accident on the streets of San Francisco, the liability question resolves quickly. Waymo is the operator. There is no human driver behind a steering wheel. Waymo bears liability — that is the commercial and legal arrangement that Waymo’s regulators, insurers, and the public expect, and it is the arrangement Waymo has explicitly accepted by deploying a fully driverless commercial service.

When a Tesla equipped with Full Self-Driving Supervised causes an accident, the same question opens into a web of competing claims. Was the human driver paying adequate attention? Did the system request a takeover that the driver failed to execute in time? Was there a software defect that caused the system to behave in a way no reasonable driver could have anticipated? Was the driver over-reliant on a product that Tesla’s own marketing had conditioned them to trust more than its Terms of Service suggested they should?

This distinction — unambiguous operator liability for Waymo’s driverless service, contested human-software split liability for Tesla’s supervised system — is not merely a legal detail. It is a fundamental structural difference that shapes how both companies design their products, how they write their Terms of Service, how they market their capabilities, which geographies they target, and what their long-run unit economics look like. Insurance is the price signal for liability: who pays the premium, how much, and under what conditions tells you everything about how the industry has assessed the risk of each deployment model.

Section 1 — The Liability Question by AV Mode

The starting point for any AV liability analysis is understanding how liability shifts as a function of automation level and system configuration.

The central tension in the table above is the Level 2+ row: FSD Supervised. The system is capable enough that drivers routinely disengage from the task of driving — studies of driver monitoring systems confirm that Level 2 users frequently stop watching the road even when instructed not to. Yet the legal framework assigns full liability to that disengaged driver. Tesla’s ToS reinforces this assignment contractually. The driver signed terms requiring them to maintain attention and take control when prompted. If they failed to do so, and the vehicle caused a crash, Tesla’s position is that the driver is responsible for that failure.

Whether courts consistently accept this position is an open question. Product liability claims have been brought against Tesla in multiple FSD-related crash cases, arguing that the system itself was defective — that it was the software’s failure to perceive a stopped vehicle, to recognize an intersection, or to respond correctly to a changed road condition that caused the crash, not the driver’s inattention. These cases have produced mixed outcomes and are ongoing. The NHTSA’s Special Crash Investigation program and Engineering Analysis investigations into FSD-related crashes are the regulatory parallel to this litigation — an effort to determine whether FSD contains safety defects that warrant recall, independent of how liability is ultimately assigned in any individual crash.

The future unsupervised FSD row is the forward-looking scenario that Tesla must navigate. When Tesla removes the supervision requirement — when FSD becomes Level 4 and no human monitoring is required — the liability shield disappears. Tesla cannot argue that a disengaged driver failed to take control of a vehicle that requires no driver. At that point, Tesla becomes the effective operator, and the liability model converges with Waymo’s.

Section 2 — How Waymo Insures Its Commercial Fleet

Waymo’s driverless commercial service operates under a clear liability model: Waymo is the operator and bears full liability for incidents involving its vehicles. This requires a commercial insurance structure that differs fundamentally from personal auto insurance.

Several aspects of this structure deserve emphasis. First, California’s minimum $5M per vehicle requirement for transportation network company operators sets a floor that is dramatically higher than personal auto insurance minimums (typically $15K–100K per incident in US states). Waymo operates above this floor (est.), but the exact coverage amounts are not publicly disclosed.

Second, the self-insurance layer is standard practice for large commercial fleet operators. Rather than paying premiums on the first layer of expected losses, Waymo absorbs them directly — effectively self-insuring the predictable, frequent minor incidents — and purchases excess coverage for catastrophic or high-severity events. This structure makes actuarial sense for a large fleet with a predictable incident rate, but it requires that Waymo have accurate data on its own incident frequency and severity. The California DMV’s mandatory incident reporting requirement provides exactly this data, both to Waymo internally and to regulators who can verify the claims Waymo makes about its safety record.

Third, the competitive moat argument is significant. AV insurance is currently priced largely on uncertainty — underwriters have limited actuarial history for driverless commercial vehicles, so premiums embed a risk premium for that uncertainty. As Waymo accumulates more driverless miles with documented incident rates that are lower than human-driven baselines, the actuarial uncertainty shrinks. Lower uncertainty means lower risk premiums, which means lower insurance costs per vehicle. The fleet operator that accumulates the most safety data the fastest — at scale, in commercial conditions — builds an insurance cost advantage that new entrants cannot replicate quickly.

Section 3 — Tesla’s FSD Liability Model: The Supervised Driver Shield

Tesla’s liability exposure with FSD Supervised is structurally different from Waymo’s because the product is designed — and marketed — as a system that requires human supervision. This design choice is simultaneously a safety philosophy, a product positioning decision, and a liability management strategy.

The “supervised driver shield” metaphor is deliberate. As long as FSD requires human supervision, Tesla can argue — contractually and legally — that the human driver is in control of the vehicle and bears responsibility for its safe operation. The system is a driver assistance tool, not an autonomous operator. By this argument, Tesla’s role in a crash is analogous to a car manufacturer’s role when a driver causes an accident using a car with normal cruise control: the tool did not cause the crash; the human operating the tool did.

The risk to Tesla’s position comes from two directions. First, product liability law does not respect contracts in all circumstances: if a court finds that FSD contained a software defect that made it unreasonably dangerous, Tesla bears liability regardless of what the ToS says. The NHTSA’s Engineering Analysis investigations are the formal mechanism for determining whether such a defect exists. If NHTSA issues a recall finding, Tesla’s contractual liability shield against individual crash claims becomes harder to maintain.

Second, the supervised driver shield only works as long as FSD is actually supervised. If Tesla’s own product makes it too easy for drivers to disengage — if the driver monitoring system is insufficiently strict, or if Tesla’s marketing implicitly communicates that supervision is pro forma rather than essential — then Tesla’s legal position that the driver was in control becomes factually difficult to sustain. This is the pattern of concern that NHTSA’s SCI investigations have documented in multiple fatal crash cases.

The liability cliff at the unsupervised FSD transition is the strategic constraint that limits how aggressively Tesla can remove the supervision requirement, even if the underlying technology becomes capable enough. Removing supervision is not just a technical milestone — it is a legal and insurance milestone that fundamentally changes Tesla’s liability exposure.

Section 4 — State-by-State Liability Framework (US)

AV liability law in the United States is primarily a state matter. There is no comprehensive federal AV liability statute as of mid-2026 (est.). This creates a patchwork of frameworks that shapes where commercial AV operators deploy and how they structure their operations.

California and Arizona represent the two ends of the regulatory spectrum for mature AV commercial deployments. California has the most detailed framework — operators must obtain driverless deployment permits from the DMV, report all incidents, maintain minimum insurance coverage, and operate within CPUC’s transportation network company rules. The framework imposes real regulatory overhead but also provides clarity: an operator with a California driverless permit knows exactly what is required of it legally.

Arizona’s framework is deliberately minimal. The governor’s executive order established a permissive environment for AV testing and commercial deployment, leaving most questions to existing tort law. Waymo has operated in Phoenix under this framework since its earliest commercial deployments — Arizona’s light-touch approach was partly why Phoenix was selected as the first commercial market. But minimal regulation also means less legal clarity: when a novel incident occurs in Arizona, courts must work from general tort principles rather than AV-specific statutory guidance.

Texas’s framework is similarly permissive, and Austin was selected as Tesla’s robotaxi launch city partly because Texas does not require a human safety driver and does not impose a heavy pre-approval process for driverless commercial operations. The trade-off is that in the absence of specific AV liability statutes, Texas courts must reason from first principles about how liability attaches when no human is driving.

The absence of a federal framework is both a risk and an opportunity for AV operators. It is a risk because liability rules vary across state lines, complicating multi-state commercial deployments. It is an opportunity because the most permissive states — Arizona, Texas, Nevada, Florida — can serve as deployment laboratories without the overhead of the most rigorous frameworks.

Section 5 — How Insurance Shapes Deployment Strategy

The insurance and liability framework is not a compliance checkbox — it is a direct input to the economics of AV deployment that influences geography selection, product design, marketing strategy, and the pace of automation-level advancement.

Waymo’s geography selection is already shaped by insurance economics. Phoenix is a lower-risk operating environment than San Francisco in several dimensions that matter to commercial AV insurance: lower traffic density in suburban grid streets, fewer cyclists and pedestrians per mile, more predictable weather, lower average vehicle repair costs. The Phoenix deployment was not chosen purely on technical grounds — it was chosen partly because the unit economics of commercial AV service, including insurance, close more easily in Phoenix than in Manhattan or central Seattle.

The “safety as marketing” dynamic for Waymo is structurally different from consumer marketing. When Waymo publishes a safety report showing lower injury rates than human drivers, the primary audience is not just the public — it is the actuarial departments at Munich Re and Swiss Re who use that data to price Waymo’s fleet coverage. Every improvement in Waymo’s documented safety record has a direct monetary value in the form of lower insurance premiums. This creates a uniquely direct feedback loop between operational safety investment and insurance cost reduction that does not exist for consumer vehicle manufacturers.

For Tesla, the insurance dynamics are inverted. Consumer auto insurance is the driver’s expense, not Tesla’s — so FSD does not directly lower Tesla’s operating costs the way commercial fleet insurance affects Waymo. Tesla’s insurance exposure is primarily through product liability: the risk of judgments in cases where courts find FSD contained a safety defect. This is a contingent liability rather than a recurring operational cost, but it is harder to quantify and manage than a commercial fleet insurance premium.

The Cybercab changes Tesla’s insurance equation. A purpose-built robotaxi without a steering wheel, operated as a commercial fleet rather than a consumer vehicle, brings Tesla into the same insurance territory as Waymo: commercial fleet coverage, operator liability, geographic premium variation, and the actuarial data accumulation dynamic. When the Cybercab fleet scales, Tesla will face the same incentive structure Waymo operates under today — every percentage point improvement in Cybercab incident rates has direct insurance cost value.

Section 6 — The Liability Cliff: Tesla’s Transition Problem

The liability cliff is the most significant strategic constraint that the insurance and liability framework imposes on Tesla’s path to fully autonomous commercial operation. Understanding it requires tracking what changes when supervision is removed.

Under FSD Supervised: the human driver is legally the operator. Tesla is the tool manufacturer. Liability for crashes flows primarily to the driver (subject to product liability claims in specific defect cases). Tesla’s direct insurance exposure is limited to litigation risk on individual cases.

Under unsupervised FSD (Level 4 operation with no human monitoring required): there is no human operator. Tesla — or whoever operates the robotaxi network — is the effective operator. Liability attaches to the operator. This requires commercial fleet insurance, not personal auto insurance. The cost of that insurance falls on the operator, not the passenger or a consumer vehicle owner.

The transition from supervised to unsupervised operation is therefore not only a technical milestone (has the software reached the reliability threshold?) and a regulatory milestone (have regulators approved driverless operation in the relevant jurisdictions?). It is also a financial milestone that increases Tesla’s insurance-related cost exposure substantially. Tesla must be confident that the revenue model — robotaxi fares, network fees, or consumer FSD subscription revenue — covers this increased cost before removing the supervision requirement commercially.

This is why the liability cliff is a strategic constraint rather than just a future compliance item. Moving too fast toward unsupervised operation — before the revenue model is established, before the actuarial data exists to price the coverage, and before the legal framework for driverless commercial operation is stable — exposes Tesla to open-ended liability costs with no established means of offsetting them. Waymo crossed this cliff deliberately, with regulatory approval in specific jurisdictions, commercial insurance in place, and a service model that charges per ride. Tesla’s path across the same cliff, with a consumer-scale fleet rather than a geographically limited commercial service, is a more complex transition.

Section 7 — The Physical AI Liability Benchmark

For the Physical AI benchmark series, liability and insurance represent a distinct dimension of commercial AV maturity — one that is often overlooked because it is less visible than miles per disengagement or fleet size. The following framework tracks liability readiness as a benchmark metric.

Waymo has the most mature liability posture of the three configurations. It has accepted operator liability, established commercial insurance, accumulated actuarial data through mandatory California DMV reporting, and operates in regulatory frameworks that have explicitly addressed driverless commercial service. The costs are higher — commercial fleet insurance at $3K–8K/vehicle/year (est.) versus the consumer auto insurance that Tesla FSD users pay themselves — but the risks are known and quantified.

Tesla FSD Supervised’s liability posture is in the most legally ambiguous position. The ToS-based driver liability shield may or may not hold in product liability cases; NHTSA’s investigations have not produced a completed recall determination on FSD as of mid-2026 (est.); and the actuarial data on FSD crash rates, while available in aggregate from Tesla’s quarterly safety reports, has not been independently audited at the same level as Waymo’s California DMV-reported incident data.

Tesla’s Cybercab is pre-cliff: it has not yet accumulated the regulatory approvals, commercial insurance arrangements, or actuarial track record that commercial fleet operation requires. When it does, it will face a liability and insurance ramp that closely mirrors what Waymo has already navigated — with the additional complexity of doing so at consumer fleet scale rather than in the limited geographic markets where Waymo currently operates.

Note: All cost estimates, fleet size estimates, coverage amounts, and regulatory interpretations in this article are directional estimates based on publicly available information as of mid-2026. Figures labeled “(est.)” should not be treated as confirmed data. This article does not constitute legal or investment advice.

Sources

• Waymo safety report — Waymo ↗
• NHTSA Special Crash Investigation program — NHTSA ↗
• California DMV AV regulations — CA DMV ↗
• Tesla safety report — Tesla ↗
• AV insurance market — Swiss Re Institute ↗

Tags