RESEARCH DOSSIER
Automobile Insurance Risk in Shared and Autonomous Mobility
Abstract. Automobile insurance developed around comparatively stable relations among a vehicle, its owner, a driver, an insurance contract, and the circumstances of road use. Car-sharing, peer-to-peer vehicle access, ride-hailing, connected fleets, and automated driving increasingly separate those relations. The resulting insurance problem is multidimensional: physical exposure changes through vehicle utilization and empty repositioning; the applicable policy changes across use phases; the legally responsible actor may differ from the insured driver or fleet owner; and software or service-provider dependencies can produce correlated losses that conventional motor risk segmentation overlooks. This monograph examines these mechanisms through comparative primary law, regulatory reports, safety research, actuarial analysis, and explicitly hypothetical stress cases. It distinguishes crash harm, liability, coverage, reported claims, ultimate insurance cost, and capital accumulation. U.S. transportation-network insurance periods, California vehicle-sharing statutes, European motor liability protection, and Great Britain's automated-driving legislation reveal different techniques for maintaining victim compensation while preserving risk allocation and recourse. Comparative safety evidence indicates that measured automated-fleet performance depends decisively on road environment, responsibility assignment, claim definitions, and exposure denominators. The analysis demonstrates that a per-mile reduction in crash incidence need not imply an equivalent reduction in aggregate insurer cost when miles, severity, repair complexity, limits, and common-cause losses change together. Its principal contribution is a documented cross-disciplinary synthesis and a reproducible set of interpretive calculations; no new general theory, original empirical dataset, or universally estimated autonomous-vehicle loss relativity is claimed.
Keywords: Automobile insurance; shared mobility; peer-to-peer car sharing; transportation network companies; ride-hailing; autonomous vehicles; automated driving systems; motor third-party liability; operational design domain; insurance coverage gaps; exposure measurement; frequency–severity models; software and cyber accumulation; claims data; insurance regulation; victim compensation; risk governance.
Introduction
The central insurance question in shared and autonomous mobility is how a loss remains attributable, compensable, and adequately financed when the party owning a vehicle, the party controlling its movement, the party operating a mobility service, the party supplying its driving technology, and the party purchasing insurance are different. Conventional motor insurance already accommodates multiple interests through named-insured clauses, permissive use, fleet contracts, third-party obligations, compulsory compensation, excess layers, and subrogation. Shared mobility intensifies the frequency with which these interests change during a single journey; automated driving changes the distribution of control and evidence relevant to injury and fault.
The scale of the underlying public-harm problem matters. The World Health Organization's Global Status Report on Road Safety 2023 estimates 1.19 million annual road-traffic deaths worldwide using its reported reference framework. This estimate measures a public-health burden; it is neither a motor-insurance claim count nor a global autonomous-vehicle risk rate. A pedestrian killed by a vehicle, an uninsured driver's liability, the insurer's coverage determination, and the ultimately paid amount occupy different stages of a legal and economic process. Treating them as one quantity would obscure the mechanisms studied here.
The research combines four levels of evidence. The first is enacted law and regulatory guidance specifying insurance responsibility and technology classification. The second is empirical evidence about mobility demand, vehicle exposure, collision reporting, and claims, each with identified sampling limitations. The third is actuarial reasoning concerning exposure, frequency, severity, portfolio dependence, and capital. The fourth is original comparative interpretation, which draws consequences from the first three without presenting itself as a newly validated scientific theory. The analysis is international but jurisdiction-specific wherever a proposition depends on statutory language. It evaluates the current documented law and sources available through November 16, 2026, rather than presuming harmonized global legal rules.
The parent discipline, Automobile Insurance, supplies the general subject boundary. The present dossier's scientific object is the changing configuration of risks and insured relationships in privately shared vehicles, platform-mediated transport, and automated fleets. Related Science includes Insurance Technology and Systemic and Emerging Risks. This dossier's canonical identity and scientific scope remain independent of those parent disciplines.
Scientific Definitions and Primary Foundations
Vehicle use, sharing, and automated operation
Shared mobility covers several arrangements with different risk-bearing structures. Station-based or free-floating car-sharing gives multiple users successive access to an operator's fleet. Peer-to-peer car sharing places an individually owned car under the temporary control of another user, ordinarily through an intermediary platform. Ride-hailing or a transportation network company (TNC) connects a passenger with a driver who may be operating a personally owned car for compensation. A pooled ride carries multiple passengers for distinct bookings; an automated passenger service may operate without a human driver responsible for the dynamic driving task within a defined operating area. A conventional rental, long-term lease, commercial taxi, and company-owned fleet may resemble individual features of these services but demand separate examination of their contractual and statutory definitions.
Automation is a property of the active driving function and the operational conditions in which that function is engaged. SAE International's current J3016 driving-automation taxonomy, revised in September 2026, organizes six levels and distinguishes ongoing driver supervision at Level 2 from automated performance under defined conditions at Levels 3 and 4. Level 5 refers to automated driving under conditions in which people can drive. The standard explicitly treats the taxonomy as a classification instrument, not proof of operational safety. It is therefore erroneous to apply the risk profile of a Level 4 driverless service to a Level 2 car whose human driver remains responsible for continuous supervision.
The operational design domain (ODD) denotes the conditions under which an automated driving feature is designed to function: relevant location, road types, speed, weather, and other operating restrictions. The European Commission's 2022/1426 ADS type-approval regulation gives concrete examples including predefined areas, hub-to-hub routes, and automated valet parking, and defines an ADS through hardware and software capable of the entire dynamic driving task in a specific ODD. This specification is important to insurance because performance outside the ODD may involve a different causal sequence, different contractual representations, and different operational responsibilities.
Six analytically distinct outcomes
A motor collision is an observable physical event. Its consequences include bodily injury, property damage, emergency disruption, and other losses. Legal liability identifies the person or entity legally accountable under the applicable rule; responsibility may be shared, subject to defenses, or resolved through a statutory compensation mechanism. Insurance coverage determines whether a particular policy responds to a defined insured person, risk, place, period, event, and exclusion. An insurer's reporting and reserving record captures a claim at a specific valuation stage. Payment can occur after adjustment, settlement, court judgment, statutory fund intervention, and contribution or recourse. These distinctions create a disciplined vocabulary:
- Physical harm: an injury or physical/economic loss caused by an incident.
- Potential liability: legally arguable responsibility for harm, before final determination.
- Insured liability: the portion of that responsibility within a responding policy or guarantee.
- Reported claim: a demand or notification recorded under the insurer's claims process.
- Ultimate insured cost: claims payments and relevant expenses at final development under defined accounting boundaries.
- Aggregate portfolio exposure: the distribution of insured outcomes across vehicles, contracts, operators, time, and dependent events.
The same crash can generate multiple claims, policy notices, defendants, and injury categories. Conversely, a single portfolio loss can encompass correlated claims from several vehicles. Per-crash, per-claim, per-vehicle, per-mile, and per-passenger rates therefore answer different scientific questions.
Contract and institution as measurement units
Insurance responds to defined interests and liabilities rather than to the vehicle alone. A fleet operator might buy commercial motor liability, physical damage, cyber, product liability, and business interruption protection with different insurers and attachment points. A peer-to-peer owner's personal policy might be excluded during a sharing interval while a platform-arranged policy attaches. An autonomous-service entity may retain uninsured first-party losses even where a victim has a direct route to an insurer. The National Association of Insurance Commissioners' commercial ride-sharing overview, updated December 2025, explicitly addresses gaps between personal and commercial arrangements and the distinction between the waiting, accepted-trip, and passenger-onboard phases.
The corresponding actuarial unit is not automatically a registered vehicle. For a personal annual contract, vehicle-years may approximate exposure. For a TNC policy whose risk turns on application status, insured hours or miles within statutory periods may be the relevant unit. For a driverless fleet, the insured entity can have miles, rider journeys, unoccupied repositioning movements, vehicle downtime, software configurations, and geographic cells. The measure must be tied to the precise loss and coverage under investigation.
Historical continuity and changed allocation
Motor insurance's basic risk exchange remains recognizable: an insured transfers specified uncertain financial consequences to a regulated risk bearer for consideration. What changes is the observational and contractual architecture. A ride-hailing journey may begin as private use, enter an app-on waiting period, become commercial transport when a request is accepted, then conclude at passenger exit or application completion. In a driverless service, remote assistance, fleet dispatch, and automated-control transitions may replace human driver status as the crucial evidentiary markers. The question is when an event occurred, under whose control and within which coverage, rather than who happened to own the chassis.
Changing Ownership and Utilization
The relationship among ownership, possession, control, and dispatch
Legal title, registered ownership, operational possession, mechanical control, and authority to assign trips can sit with separate parties. A car-sharing company may own and insure a large fleet while a succession of registered members operate individual vehicles. In peer-to-peer sharing, title remains with a private owner while the program arranges custody and contractual protection for a guest. In ride-hailing, a driver may own a car while the platform determines access to prospective passengers and collects data about trip phases. In automated fleet operation, software developers, authorized operators, financiers, maintenance contractors, and remote-assistance providers introduce additional contractual links.
These distinctions change insurer incentives. A vehicle owner has reason to safeguard residual asset value; a temporary user influences immediate driving exposure; a platform influences matching, route assignment, screening, pricing, and operating density; a developer influences the behavior of an automated driving feature; and an insurer bears defined financial loss subject to contract terms. Risk governance works best when each institution controls or can document the mechanisms for which it is contractually or legally accountable. But compensation of an injured third party cannot always depend on first identifying the precise software defect or operator error. Legislatures therefore sometimes separate prompt victim payment from subsequent contribution and recourse.
How utilization changes the denominator
Vehicle count is an inadequate proxy for total accident opportunity when the same asset serves many users or remains in service longer each day. Suppose a private car travels 10,000 insured miles per year and a shared fleet vehicle travels 45,000. Equal per-mile collision rates would imply different expected annual incident counts per vehicle. A claim frequency quoted per vehicle-year would rise even if the technology lowered the risk per driven mile. Conversely, a per-trip rate can appear low if mean trip length falls while empty repositioning mileage increases.
A useful portfolio exposure identity, treated here as standard bookkeeping rather than a novel model, is:
An insured fleet's total mileage includes paid passenger miles and movements without passengers: traveling to pickups, repositioning, charging, inspection, maintenance, and returning vehicles. Different coverage clauses may attach to different components. Vehicle miles traveled (VMT), passenger miles traveled (PMT), and completed trips must not be conflated. Occupancy-adjusted passenger miles can increase while total VMT decreases, remain unchanged, or increase depending on pooling, routing, induced demand, and modal substitution.
Fagnant and Kockelman's 2015 transportation-policy analysis identifies potential safety and efficiency gains from automated vehicles alongside barriers involving liability, privacy, security, and the possibility of higher vehicle miles traveled. The International Transport Forum's Helsinki shared-mobility simulations examine changes to access, rail use, congestion, and parking under specified replacement assumptions. These are scenario studies; their conclusions depend on service design, the surrounding transit system, and modeled behavior, rather than establishing a universal demand response.
A contrasting observational finding comes from Barrios, Hochberg, and Yi's ride-hailing and traffic fatalities study. Their 2020 working paper associates ride-hailing entry in U.S. cities with approximately 3% more traffic fatalities and fatal crashes, using an empirical research design with controls. The finding is evidence about a particular period, geography, and identification strategy; it should not be recast as the crash risk of an individual hired ride or as a parameter for autonomous driving. It underscores why the net system-wide road-safety effect of mobility services may differ from the conditional safety of a vehicle technology.
Selection, composition, and operational intensity
The insured risk changes with user composition, vehicle type, road environment, trip timing, passenger occupancy, and driver or software behavior. A shared vehicle might have a younger fleet age but longer daily duty cycles, different road environments, more frequent parking maneuvers, and multiple operators with heterogeneous experience. A commercial journey may expose more passengers, pedestrians, cyclists, and other motorists than a comparable private trip, although the direction of conditional risk must be measured rather than presumed. A shared operator can also standardize preventive maintenance, training, telemetry, and incident response more effectively than many isolated owners.
The correct statistical comparison controls for the relevant mix. Let exposure cell identify location, time of day, weather, road class, vehicle class, driving mode, and use status. Then expected loss is the sum across cells of exposure multiplied by each cell's conditional claim cost. Aggregating across unmatched cells can produce a misleading overall ratio: a fleet restricted to low-speed mapped streets may appear safer than a nationwide passenger-car benchmark even if neither technology has been shown safer in identical environments. A credible comparison must adjust the benchmark to the fleet's actual ODD and coverage scope.
Evidence from autonomous fleet reporting
California's Department of Motor Vehicles reported more than nine million testing miles for autonomous-vehicle testing permit holders during December 2024–November 2025. Those figures describe reported permit-holder testing, not all commercial driverless deployment miles, total U.S. autonomous driving, or a representative private-vehicle sample. California's own account warns that disengagement submissions are not designed to rank company safety performance and that deployment-permit operations are outside that disengagement series.
The National Highway Traffic Safety Administration's Standing General Order on crash reporting requires designated entities to report qualifying crashes involving ADS and Level 2 driver assistance under differing thresholds. The order was amended in 2025, changing data fields for reporting effective June 16, 2025. NHTSA notes variation in telemetry, manufacturer awareness, reporting thresholds, duplicate incident records, and completeness. The publicly described dataset available for this review ran through September 15, 2026. Raw reported crash counts consequently cannot yield sound comparative crash rates without independently established denominators and compatible ascertainment rules.
An illustrative utilization sensitivity
Consider two hypothetical populations, each consisting of 1,000 vehicles. The first travels 10,000 miles per vehicle-year at two claims per million miles, giving 10 million miles and an expected 20 claims. The second travels 40,000 miles per vehicle-year at one claim per million miles, giving 40 million miles and an expected 40 claims. The latter exhibits half the frequency per mile and twice the expected claim count per vehicle-year. With an illustrative mean claim severity of $8,000 for private vehicles and $11,000 for shared vehicles, expected annual losses are $160,000 and $440,000 respectively. This example demonstrates the arithmetic of exposure and severity; its numbers are invented stress inputs and have no asserted empirical provenance.
This distinction also affects affordability and capital decisions. A fleet can be safer per mile and still generate a greater total insurance premium because its asset utilization, passenger obligations, claims-handling costs, or severity are greater. Conversely, a pooled passenger service can reduce fleetwide VMT enough to offset a greater number of journeys. Empirical parameter estimation must precede a pricing conclusion.
Insurance Product Portability
Continuity of insured protection across phases
Portability describes the continuity and transfer of applicable insurance protection as a person, vehicle, and activity pass between legal and operating states. A driver moving from private use to platform use may cross contract exclusions, commercial policy attachment points, mandatory statutory limits, and excess coverage layers. A privately owned car entering peer-to-peer sharing changes authorized possession and potentially the applicable primary insurer. For an automated fleet, software updates, authorization conditions, and remote-control roles can change insured risk without transferring registered ownership.
Portable protection is not accomplished by a generic reference to an insured car. A complete coverage analysis records: the moment exposure begins and ends; whose use is insured; the source of compulsory third-party protection; limits and deductibles; whether first-party physical damage, passenger injury, and uninsured motorist benefits respond; policy priority; the response when no contract pays; and recovery rights after victim compensation. Coverage continuity should be tested against the loss timestamp and local legal rules. The distinction between liability protection and optional physical-damage protection is particularly important because commercial platform statutes commonly mandate the former more comprehensively than the latter.
U.S. transportation network company periods
California provides an unusually transparent period-based example. The California Public Utilities Commission's TNC insurance requirements distinguish application-open waiting, accepted-trip travel toward pickup, and passenger occupancy. For the waiting period, the stated primary liability requirements include $50,000 per person, $100,000 per occurrence for injury, and $30,000 for property damage, with an additional $200,000 excess protection requirement. For accepted trips and passenger carriage, the stated primary commercial liability requirement is $1 million. The public regulator's account must be applied to its California regulatory setting; it is not a global market norm, and statutory law and effective contractual wording govern individual disputes.
The NAIC's December 2025 ride-sharing analysis identifies continuing first-party physical-damage gaps, personal-policy livery exclusions, rideshare endorsements, and the problem of simultaneous platform application use. A vehicle may be legally covered for a third party's injury while its owner's collision damage remains uninsured for the same event. A passenger's injury benefit and the driver's own medical expenses can have additional, distinct attachment tests. This is why a contract comparison confined to the headline liability limit materially understates protection discontinuities.
The distinction can be represented as an event timeline. At time , private use is governed by the personal or commercial policy applicable to that use. At , the app activates, creating a waiting-phase exposure. At , a ride request is accepted and the higher-limit commercial phase begins under California's stated model. At , a passenger enters; at , the passenger exits; at , the platform transaction ends. Some transitions may coincide, and the exact terminal boundary depends on governing law. A claim at can have a different coverage response from one at , although physical harm occurs within nearly identical driving conditions. Accurate and auditable timestamps are therefore essential claims evidence.
Peer-to-peer car-sharing statutes
Peer-to-peer sharing raises a different question: how an owner can participate without relying on ordinary permissive-use coverage for a commercial-platform transaction. California Insurance Code section 11580.24 addresses eligible personal vehicle-sharing programs, recordkeeping for time, location and mileage, the program's assumption of the owner's liability under statutory conditions, and the potential exclusion of personal-policy coverage during the sharing period. The statute includes conditions about vehicle use and program status and should be read alongside its companion requirements rather than reduced to a blanket platform guarantee.
California's separate section 11580.241 specifies coverage and purchasing options for the program's sharing period, including statutory minimums reported in the current code. The statutory period and agreed protections deserve independent audit because a guest's vehicle damage, third-party injury, passenger injury, equipment loss, and owner revenue interruption may attach to different products or contractual promises.
Colorado's 2026 codified Peer-to-Peer Car Sharing Act provides another design: a program assumes defined owner liability during the sharing period, secures applicable insurance, and maintains usage records. Section 6-1-1206 separately protects otherwise applicable uninsured- or underinsured-motorist coverage from exclusion solely because a driver or passenger participates in car sharing. California and Colorado thus illustrate that similar commercial practices can be governed by distinct statutory rules concerning private-policy exclusions, program obligations, and available protection. The insurer's file must preserve the state, statutory date, policy form, sharing-period definition, and any additional endorsements.
The European motor-liability baseline
The European Union's Motor Insurance Directive 2009/103/EC organizes compulsory liability protection for motor-vehicle use across Member States. Directive (EU) 2021/2118 revised important elements, including the treatment of injured parties when an insurer becomes insolvent. These EU instruments establish a coordination baseline while leaving consequential questions of tort liability, insurance contracts, implementation, and compensation procedure to applicable national law. A car-sharing or automated vehicle's presence in the EU does not by itself determine the responsible party or policy priority in every Member State.
The analytical sequence is first to identify the applicable definition of vehicle use and the compulsory motor obligation, then the injured person's compensation right, then the insurer's potential coverage and statutory defenses, then contribution, recourse, and contractual subrogation. For cross-border trips, this sequence extends to the location of the accident and the relevant compensation machinery. EU-wide legal harmonization should be described at the level actually supplied by the directives, with local implementation separately established for any country-specific conclusion.
Great Britain: insurer-facing compensation and authorized entities
Great Britain's Automated and Electric Vehicles Act 2018 expressly provides, under its statutory conditions, for an insurer to be liable for damage when an insured automated vehicle causes an accident while driving itself on a road or other public place. The Act also addresses uninsured vehicles in specified public-service circumstances and provides for recourse and related matters in Part 1. Its structure illustrates a distinct allocation: victim compensation can be channeled through an insurer even when subsequent disputes concern the technology's operation or another party's underlying responsibility.
The Automated Vehicles Act 2024 develops an authorization architecture that includes an authorized self-driving entity with ongoing responsibility for continued satisfaction of authorized vehicle conditions. The enacted text and the provisions actually commenced on the relevant date must be separated. Government-authorized software operation, regulatory duty, civil liability, and an insurer's indemnity are different legal categories; the allocation of each depends on the applicable operative provisions. Claims analysis should track the authorization status, engaged feature, event record, and commencement position rather than assume the entire 2024 statute was operative at enactment.
A source-traceable coverage continuity review
An insurer or public decision-maker can test a specific mobility service by constructing a coverage map with columns for jurisdiction, use phase, control status, insured interest, underlying policy, statutory guarantee, optional extension, claim type, retention, excess layer, and claim evidence. A map covering all phases reveals silent intervals more clearly than a generic annual declaration. Any purported continuity must also survive adverse scenarios: simultaneous use of two TNC apps, a disconnection during a trip, a guest failing to return a shared car on schedule, a remote operator assuming limited control, a vehicle leaving an approved ODD, and an insurer insolvency event. The proper resolution is necessarily legal and contractual, not derived from a universal algorithm.
Risk Concentration and Emerging Governance
From motor attrition to technology-dependent accumulation
Conventional automobile insurers expect many largely independent small physical-damage incidents and a much less frequent, longer-tailed distribution of severe bodily injury claims. A fleet of shared or automated vehicles can preserve these individual claim processes while acquiring dependencies through common software releases, fleet dispatch decisions, remote services, charging infrastructure, maps, connected communication, cybersecurity exposures, and concentration in a single city or operating corridor. The insurer must therefore consider both ordinary motor loss and common-cause disturbances.
A widespread software fault does not necessarily cause a collision; it can trigger vehicle stoppages, business interruption, towing, extra operational costs, or, in an adverse scenario, simultaneous or near-simultaneous liability incidents. These outcomes belong to different coverage lines and require a careful loss-path definition. Common software version among vehicles is an exposure indicator, not evidence that the software is defective or that an insurer will indemnify any resulting cost. A remote-service outage can prevent dispatch without producing covered physical damage. Conversely, a failure causing several crashes can create multiple motor third-party claims, litigation, excess attachment, and possible product-liability recoveries. Insurance contracts, exclusions, attachments, and claims facts decide how the same underlying incident propagates financially.
The insurer's ordinary frequency–severity calculation is:
where is exposure in a defined unit (such as millions of insured miles), is expected covered claim frequency per such unit, and is expected ultimate insured loss per claim in cell , including only the cost categories specified in the valuation. This identity is an established actuarial decomposition. It becomes misleading if cells mix platform statuses, technical driving modes, inconsistent liability thresholds, or differing expense definitions. The expected amount need not describe volatility, tail risk, capital, or cash-flow timing.
A simple interpretive distinction is between independent event loss and a common shock capable of altering several exposures simultaneously. Conditional on , claim intensities or severities may move together. Marginal reductions in will not guarantee proportional reductions in a high quantile of total if the distribution of becomes more concentrated. Quantifying the common shock requires observed fleet dependencies, causal incident analysis, engineering evidence, insurer contracts, and insurer-level portfolio mapping. The scenario is a sensitivity exercise, not a fitted statistical finding.
Aggregate loss illustrations
Consider a hypothetical automated fleet portfolio with 100 million insured miles per year. At an illustrative covered-claim rate of 0.8 per million miles and $15,000 mean ultimate cost, expected ordinary annual loss is $1.2 million. A conventional comparator with 100 million miles, 1.6 covered claims per million miles, and $10,000 mean severity has expected ordinary loss of $1.6 million. On those assumptions the automated fleet has lower ordinary expected loss, although its higher severity offsets part of the frequency gain. If the automated fleet instead has 150 million miles because of greater utilization, its expected ordinary loss becomes $1.8 million, exceeding the same comparator despite the lower per-mile frequency. The figures are intentionally invented and do not estimate actual autonomous-fleet premiums.
Now assume, only for illustration, that the automated fleet has an annual probability of 0.5% of a technology-related common event producing $30 million of covered insurer loss after contract terms. Its incremental expected annual amount is $150,000. If that event were independent of ordinary attritional loss and did not change its distribution, illustrative mean aggregate loss at 100 million miles would rise from $1.2 million to $1.35 million. The mean contribution is small relative to the severe event magnitude; it says little about a high-tail solvency quantile. If the occurrence probability is doubled while the conditional loss remains unchanged, the added mean becomes $300,000. If the adverse event produces multiple connected claims, reserving and reinsurance attachment must follow the actual wording of an occurrence or event rather than an assumed single limit.
| Hypothetical scenario | Insured miles | Claims per million miles | Mean covered claim | Ordinary expected loss |
|---|---|---|---|---|
| Conventional comparator | 100 million | 1.6 | $10,000 | $1.60 million |
| Automated fleet, equal miles | 100 million | 0.8 | $15,000 | $1.20 million |
| Automated fleet, higher utilization | 150 million | 0.8 | $15,000 | $1.80 million |
| Automated fleet, severity stress | 100 million | 0.8 | $25,000 | $2.00 million |
The stress demonstrates three independent levers: exposure, frequency, and conditional insured severity. It is inappropriate to infer from a percentage change in crash frequency an identical percentage change in pure premium. Even before expense and profit loading, insured severity is shaped by passenger occupancy, bodily injury inflation, repair cost, component replacement, litigation duration, residual limits, and which party bears the covered amount. A safety comparison that counts all collisions and an insurer comparison that counts at-fault third-party claims will typically have different denominators, definitions, and answers.
Safety evidence and the meaning of a measured reduction
Di Lillo and coauthors' 2024 comparative safety study uses liability insurance claims to construct geographically and responsibility-calibrated human benchmarks and applies them to a Waymo-operated automated fleet. The reported human-reference data comprise more than 600,000 private passenger vehicle claims and approximately 125 billion miles of exposure. The study finds reduced claims against the evaluated automated operation relative to its calibrated benchmark. The approach addresses important location and responsibility mismatches; its conclusions remain conditional on the investigated fleet, observation period, ODD, claim definition, benchmark construction, and data availability. It does not establish equivalent performance for all autonomous systems, all countries, or all kinds of insured loss.
A real-world reduction in measured third-party claims is scientifically meaningful, but at least four steps intervene between that result and an insurer's portfolio decision. The automated system's mileage and use case must resemble the future insured portfolio; the safety endpoint must coincide with the insurer's covered claim definition; conditional severity and loss development must be estimated; and common-cause, contractual, litigation, and capital exposures must be added. Proprietary operator self-reporting can usefully disclose operational metrics but requires transparent handling of its underlying selection, claims matching, and comparison construction. Any published rate should identify its numerator, denominator, geography, ODD, calendar window, and confidence or uncertainty interval.
Cybersecurity, software updates, and recovery rights
Connected vehicle risks arise from both accidental software defects and intentional compromise. The two demand different causal evidence and insurance responses. ISO/SAE's 21434 cybersecurity engineering standard establishes lifecycle cybersecurity risk-management requirements for road vehicles. The United Kingdom Vehicle Certification Agency's 2026 overview of UN Regulations 155 and 156 identifies regulatory requirements for vehicle cybersecurity management and software-update management. Compliance with a technical or process standard can inform due diligence, underwriting, and causation analysis, but it is not a guarantee against every adverse event or a substitute for reading insurance conditions.
A meaningful insurer assessment links the claim-generating function to a specific software or hardware configuration, update history, ODD, safety driver/remote assistance condition, supplier, and incident record. Version concentration creates a mapping challenge: vehicles carrying the same release can be distributed across several policyholders and primary insurers yet depend on the same vendor. A product-liability or software supplier's indemnity may mitigate ultimate cost, but recovery timing and success remain uncertain and may produce reserves and liquidity costs even when the insurer ultimately recovers part of the claim. Insured losses must be distinguished from operators' uninsured economic interruption and manufacturer costs.
Reinsurance and capital governance
An accumulation analysis proceeds from an inventory of insured entities and fleets to an engineering-informed common-event map, gross and net loss projections, contract response, dispute risk, and liquidity timing. A conventional motor excess-of-loss treaty might depend on whether multiple crashes from one software incident are treated as one event or separate occurrences. Aggregate stop-loss, catastrophe, product liability, and cyber protections may attach to different losses and include different aggregation definitions. Reinsurance cannot be valued from fleet miles alone. Deductibles, event definitions, exclusions, reinstatements, credit risk, and collateral terms determine residual insured capital exposure.
The International Association of Insurance Supervisors' Insurance Core Principles and ComFrame require enterprise-level identification, measurement, reporting, and management of material risks within the applicable supervisory framework. ICP 16, in particular, concerns enterprise risk management for solvency purposes. A shared automated-fleet insurer should assess technology concentration alongside geographical, underwriting, claims inflation, operational, counterparty, and legal risks; dependency controls and capital allocation must reflect the actual portfolio rather than the novelty of the technology label.
The International Accounting Standards Board's IFRS 17 insurance-contract framework sets recognition and measurement principles for insurance contracts within its scope, including current fulfillment cash flows and risk adjustment. IFRS 17 is an accounting standard, whereas solvency capital and prudential reinsurance requirements are imposed by the relevant jurisdiction. An insurer can therefore show improved expected future insured losses but still require more capital if tail concentration, uncertainty, or reserve development deteriorates. Financial statement accounting and solvency assessment should be compared on their respective measurement bases.
Data rights, privacy, and claims proof
The same telemetry that may resolve coverage and causation can identify driver behavior, movements, passenger journeys, and personal interactions. The European Data Protection Board's connected-vehicle privacy guidelines explain data-protection issues specific to vehicle-connected applications. The EU General Data Protection Regulation requires lawful, purpose-limited, and proportionate personal data processing under its operative conditions. It follows that collecting all possible footage and location histories merely because they may aid insurance investigation is not a scientifically or legally sufficient data policy.
Claims evidence should be designed around necessity, provenance, retention period, reliability, access rights, and evidentiary chain of custody. Insurers may need to know whether an automated feature was engaged, whether a transition occurred, and whether relevant warnings existed, without retaining complete longitudinal movement histories unrelated to the claim. Technical access, privacy law, and fair claims handling can be reconciled through narrowly scoped and verifiable records. Which data controller may disclose evidence to whom depends on jurisdiction, lawful basis, contractual roles, and applicable procedure.
Evidence, Alternative Models, and Methodological Limits
Sources, populations, and levels of inference
The source corpus divides into official law and regulation, technical standards, administrative reporting, insurer-facing supervisory descriptions, and scholarly research. Primary law establishes legal duties in a defined jurisdiction and temporal state; it does not estimate future accident frequency. Technical standards supply terminology or engineering-process obligations; they do not establish insurance liability. Administrative incident reports reveal reportable cases under specified inclusion criteria; they do not by themselves supply a comparative frequency denominator. Peer-reviewed observational studies can estimate conditional contrasts subject to design and data restrictions. Institutional scenario simulations test modeled assumptions, not realized national injury outcomes.
The World Health Organization's road-safety baseline illustrates the danger of using a general harm statistic as an insurance base rate. Likewise, California's autonomous-vehicle testing summary and NHTSA's ADS incident-reporting system concern different populations. Pooling their observed counts and denominators would be methodologically invalid unless coverage, jurisdiction, time, vehicle class, duty cycle, and inclusion rules could be shown comparable.
Exposure and outcome matching
An ideal comparison specifies an estimand before selecting data. Examples include: (i) at-fault bodily injury liability claims per million insured miles within a defined city and ODD; (ii) any collision producing property damage per million engaged ADS miles; (iii) third-party paid indemnity per completed passenger journey; and (iv) ultimate motor insurance loss per insured fleet year. Each is a legitimate question but answers a different policy or actuarial decision. The reporting system may reveal collision incidence while confidential insurer data are needed to estimate ultimate payment, legal expenses, recovery, and long-tail settlement.
For empirical estimation, a consistent cell-level rate is , where the count and exposure share the same time, geography, engaged-feature scope, insured population, and event threshold. A severity estimate requires the corresponding ultimate loss distribution and appropriate handling of inflation, policy limits, deductibles, and censored claims. Sparse severe incidents mean a small fleet may have seemingly favorable average results while the uncertainty about tail outcomes remains substantial. Statistical uncertainty and the possibility of unobserved confounding should accompany any automated-system safety estimate.
The insurance-claims benchmark study advances the field by calibrating geography and responsibility; it is still important to examine whether human private-car claims and a geographically restricted robotaxi service differ in passenger type, road-user interaction, trip timing, claim reporting, policy limits, and adjudication. This does not invalidate the documented result. It specifies which comparisons the result can support and which require further evidence.
Claim ascertainment and selection bias
A collision report may depend on vehicle sensors, the manufacturer becoming aware of the event, the severity threshold, and reporting rules. NHTSA's SGO data limitations explicitly warn that systems with richer telemetry can report more incidents than systems with weaker recording and transmission. Differences in reported crash counts may therefore combine true safety differences with ascertainment differences. Multiple reporting entities can also submit records pertaining to the same event. The public report inventory should be deduplicated and stratified before any inference about comparative safety.
Selection into shared mobility likewise matters. Consumers who choose a fleet, areas in which the fleet is authorized, and periods when it operates may differ from the comparison population. A taxi replacement, a private-car replacement, and a new induced journey have different effects on total road exposure. The ITF Helsinki simulation makes the replacement assumptions visible; the ride-hailing fatalities paper tests a different historical question about the introduction of a platform service. Results from one cannot be mechanically transferred into the other.
Alternative actuarial specifications
At least four familiar modeling families can address different parts of the problem:
- A stratified frequency–severity model estimates claims and costs by legally relevant use phase, vehicle class, geography, and driving mode, provided sufficient credible data exist.
- A hierarchical or credibility model borrows statistical strength across related fleets while explicitly allowing operator, software, and region effects to differ; the sensitivity to pooling assumptions must be reported.
- A survival or event-history model treats time and distance until the next event, with changing use or ODD as time-varying covariates; exposure censoring and changes in reporting remain central.
- A scenario or common-shock aggregation examines events that simultaneously alter a group of insured vehicles. It can be useful before severe data accumulate, but its event probability and conditional loss must be disclosed as assumptions rather than fitted evidence.
These are established statistical and actuarial tool classes, not new models introduced by this dossier. Using them in combination can expose disagreements hidden by a single average rate. Their appropriateness depends on the available datasets, the covered cost endpoint, and the intended decision.
A reproducible sensitivity table
For a fleet with million insured miles, frequency covered claims per million miles, mean ultimate insured cost , annual common-event probability , and conditional covered event loss , the illustrative expected total is when the common-event loss is added without double counting. The following calculations do not assert real-world values:
| Input or output | Central hypothetical case | Exposure stress | Frequency stress | Common-event stress |
|---|---|---|---|---|
| Million miles, | 100 | 150 | 100 | 100 |
| Claims per million miles, | 0.8 | 0.8 | 1.2 | 0.8 |
| Mean insured cost, | $15,000 | $15,000 | $15,000 | $15,000 |
| Annual common-event probability, | 0.005 | 0.005 | 0.005 | 0.010 |
| Conditional common-event insured loss, | $30 million | $30 million | $30 million | $30 million |
| Ordinary expected loss | $1.20 million | $1.80 million | $1.80 million | $1.20 million |
| Common-event expected contribution | $0.15 million | $0.15 million | $0.15 million | $0.30 million |
| Illustrative total mean | $1.35 million | $1.95 million | $1.95 million | $1.50 million |
The sensitivity exercise provides an important distinction: alternative combinations can produce equal expected loss while distributing cost between attrition and correlated shocks differently. Even if their means coincide, their reserves, high quantiles, reinsurance recoveries, and liquidity profiles can differ materially. A regulatory capital or actual premium decision requires calibrated distributions, parameter and model uncertainty, expense and profit provisions, legal enforceability, and insurer-specific financial terms. It is therefore improper to publish these invented inputs as predictions.
Legal comparison under temporal change
Comparative legal reasoning must preserve the difference between enacted text, law in force, implementation guidance, administrative practice, and private policy conditions. The UK 2024 authorization statute contains commencement-dependent provisions, while the insurer-facing 2018 Act has its own conditions. EU 2021/2118 amendments operate alongside national implementing regimes. U.S. TNC and peer-to-peer laws differ by state and by sharing-period definition. A universal statement that insurers always pay directly, that platforms always assume liability, or that manufacturers always become the insured would exceed these sources.
A useful legal evidence matrix records: cited instrument; jurisdiction; section; effective and amendment status; persons and vehicles within scope; trigger; required coverage or compensation; exclusions and exceptions; priority; and recourse. For the United States, primary policy forms, endorsements, and claims interpretations would be required to resolve actual disputed cases. For cross-country study, translated statutory language and court decisions may alter a conclusion. This monograph establishes comparative structures supported by the identified official materials rather than offering individualized legal advice.
Bibliographic and data limits
Publicly accessible source pages and document abstracts were checked for relevant assertions. The complete text of every proprietary insurer policy, full claim-level Swiss Re dataset, active automated-service telemetry, manufacturer software-incident logs, all national transpositions of EU law, and judicial decisions in each jurisdiction were not independently obtained. Some technical standards are available on commercial terms; the publisher's public synopsis and edition status establish the cited classification but not access to every normative clause. A number of empirical results derive from researcher or operator-provided data, for which independent raw-data reproducibility is limited by access conditions. These limits define the strength and transferability of the findings.
The research date is a temporal boundary. Regulations, operating permissions, safety datasets, and insurer wording may change after November 16, 2026. A later release assessment must recheck genuinely time-sensitive legal and empirical claims if they have changed, while preserving the original scientific review and the fixed content hashes for this edition.
Comparative Scientific Synthesis and Research Conclusions
Ownership and control are separate dimensions of insurability
The first comparative finding is that the insured risk cannot be identified solely from the owner–vehicle pair. Across TNC travel, peer-to-peer sharing, and driverless fleets, the effective insured relationship depends on a richer event description: use phase, legally authorized operator, active driving feature, policy status, and responsible service institutions. California's period-based TNC requirements and vehicle-sharing provisions provide different statutory mechanisms for handling these transitions. The factual and legal records required to determine an insurer's obligation become part of risk governance itself.
The underlying consequences are practical. A single incident must be assigned to its actual temporal use phase, covered person, insured interest, and jurisdiction. Where an insurer pays a victim before the ultimate technical cause is determined, subsequent claims against an operator or supplier occupy a separate recovery process. Legal allocation and engineering causation are related but distinct.
Loss exposure is multi-unit and conditional on the service model
The second finding is that mobility safety, insurer loss cost, and system-wide road harm must be measured on compatible units. Reduced collisions per mile can coexist with greater annual claims per fleet vehicle because shared vehicles accumulate more mileage. Lower per-mile injury frequency can coexist with higher insured severity because repair, litigation, and passenger-obligation costs differ. Shared rides can affect total VMT through replacement or induced demand. The cited transport simulation, ride-hailing observational analysis, and claims-based ADS comparison answer complementary questions with different designs and endpoints.
An adequate scientific disclosure for a fleet safety or insurance result therefore includes fleet geography and ODD, miles by activity, trips, passenger occupancy, insured states, accident thresholds, at-fault classification, claim categories, confidence intervals, and any missing record mechanism. Without these definitions, apparently precise risk ratios are only labels attached to incomparable quantities.
Compensation continuity and risk-bearing responsibility can be allocated separately
The third finding is institutional. The insurer-facing provisions of the UK 2018 legislation illustrate a mechanism by which injured parties may have a defined insurer claim even when other actors might later bear responsibility. The EU 2021 motor-insurance reforms emphasize protection after insurer insolvency within the amended European framework. These instruments pursue victim protection through different statutory methods. In shared mobility, program-specific and phase-specific insurance can also protect third parties while leaving optional first-party exposures for separate contract formation.
Coverage architecture should be assessed as a sequence of enforceable obligations rather than as one advertised headline limit. For research, this means comparing actual claimant rights, protection scope, exclusions, policy priority, and solvency as well as premiums. It also means recognizing that a contractual insurer–operator dispute can coexist with a statutory entitlement of an injured person.
System-wide technology risks need portfolio-level evidence
The fourth finding is that automated fleet deployment may change the dependence among insured losses. Geographic concentration, shared software configurations, centralized dispatch, and connected services make a common technical event more important to portfolio governance. Existing cybersecurity engineering standards and international supervisory principles supply relevant control and risk-management disciplines. Their application must be translated into covered loss scenarios, engineering evidence, portfolio exposure mapping, contractual attachment, and reinsurance aggregation. Technology adoption alone provides no numerical tail-loss parameter.
The illustrative calculations confirm that lower mean attritional loss and greater tail exposure can coexist; this possibility makes careful aggregation and reinsurance wording necessary. It remains a conditional proposition until probabilities, severities, and dependencies are estimated for a real fleet and insurer portfolio.
Evidence governance is part of insurance governance
The fifth finding concerns observability. Automation and connected mobility may provide richer event evidence than conventional vehicle operations, but additional telemetry changes which incidents are observed and raises legal restrictions on personal-data handling. NHTSA's crash-reporting cautions show that reporting capability changes the measured numerator. The EDPB's vehicle-data guidance makes data minimization, role allocation, and lawful processing integral to evidence collection. Insurers need demonstrable provenance and retention controls while preserving the information necessary for legitimate compensation, causation analysis, and legal defense.
The proper empirical standard is a traceable observation process, not maximum surveillance. An auditable source-to-claim chain enables scientific comparisons and fair claims handling without treating every vehicle movement as unlimited research material.
Future independently testable questions
Further research should test whether changes in insured cost per million miles persist after matching ODD and coverage, whether fleet software versions have detectable tail dependence, how liability and indemnity are allocated after mixed human–automated control transitions, how third-party injury compensation differs under varied statutory schemes, and how service-specific data retention affects contested claims. These questions require access to appropriately governed original datasets, current contract forms, causal incident records, and longitudinal insurer loss development. They are proposed empirical investigations within existing risk-science disciplines, not new theoretical doctrines.
Conclusion
Automobile insurance risk in shared and autonomous mobility is determined by a conjunction of transport exposure, use-phase eligibility, legally assigned responsibility, contract response, claims observability, and portfolio dependence. Shared use makes the duration and identity of control more important; automation makes the engaged function, operational design domain, engineering chain, and authorized entity central evidence; software-intensive fleets can introduce common-cause exposures whose financial expression depends on coverage and concentration. International legal systems use different institutional arrangements to secure third-party compensation and allocate subsequent responsibility. The comparison establishes why insurance analysis must distinguish actual road harm, liability, coverage, claim records, ultimate payments, and capital requirements.
The available research supports conditional safety improvements for particular evaluated automated fleets and offers documented risks surrounding ride-hailing exposure and traditional coverage transitions. It does not support a universal numerical premium reduction for autonomous vehicles or a worldwide rule transferring motor liability to a single technology provider. The completed analytical synthesis instead identifies the facts and denominators required for credible decisions and demonstrates, with reproducible hypothetical arithmetic, how lower per-mile incident rates may coincide with higher aggregate insured loss or correlated tail exposure. Scientific progress in this field rests on direct causal evidence, transparent comparisons, enforceable compensation architecture, and portfolio-sensitive actuarial measurement.
References
Barrios, John M., Yael V. Hochberg, and Hanyi Yi. 2020. “The Cost of Convenience: Ridehailing and Traffic Fatalities.” NBER Working Paper 26783. https://www.nber.org/papers/w26783.
California Department of Motor Vehicles. 2026. “Autonomous Vehicle Permit Holders in California Logged More Than 9 Million Test Miles between December 1, 2024 and November 30, 2025.” February 20. https://www.dmv.ca.gov/portal/news-and-media/autonomous-vehicle-permit-holders-in-california-logged-more-than-9-million-test-miles-between-december-1-2024-and-november-30-2025/.
California, State of. 2025. California Insurance Code § 11580.24, amended by 2024 legislation effective January 1, 2025. https://law.justia.com/codes/california/code-ins/division-2/part-3/chapter-1/article-2/section-11580-24/.
California, State of. 2026. California Insurance Code § 11580.241, code compilation current January 1, 2026. https://codes.findlaw.com/ca/insurance-code/ins-sect-11580-241/.
California Public Utilities Commission. 2026, accessed November 16. “Insurance Requirements for TNCs.” https://webproda.cpuc.ca.gov/regulatory-services/licensing/transportation-licensing-and-analysis-branch/transportation-network-companies/tnc-insurance-requirements.
Colorado General Assembly. 2026. Colorado Revised Statutes, Title 6, Article 1, Part 12, “Peer-To-Peer Car Sharing Act.” https://olls.info/crs/crs2026-title-06.htm.
Di Lillo, Luca, et al. 2024. “Comparative Safety Performance of Autonomous- and Human Drivers: A Real-World Case Study of the Waymo Driver.” Heliyon. https://doi.org/10.1016/j.heliyon.2024.e34379.
European Data Protection Board. 2021. Guidelines 01/2020 on Processing Personal Data in the Context of Connected Vehicles and Mobility Related Applications, adopted March 9. https://www.edpb.europa.eu/documents/guideline/guidelines-012020-on-processing-personal-data-in-the-context-of-connected_en.
European Parliament and Council. 2009. Directive 2009/103/EC Relating to Insurance against Civil Liability in Respect of the Use of Motor Vehicles. https://eur-lex.europa.eu/legal-content/en/TXT/?uri=CELEX%3A32009L0103.
European Parliament and Council. 2016. Regulation (EU) 2016/679 (General Data Protection Regulation). https://eur-lex.europa.eu/eli/reg/2016/679.
European Parliament and Council. 2021. Directive (EU) 2021/2118 Amending Directive 2009/103/EC on Motor Insurance. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32021L2118.
European Commission. 2022. Commission Implementing Regulation (EU) 2022/1426 on ADS Type-Approval of Fully Automated Vehicles. https://eur-lex.europa.eu/eli/reg_impl/2022/1426/oj/eng.
Fagnant, Daniel J., and Kara M. Kockelman. 2015. “Preparing a Nation for Autonomous Vehicles: Opportunities, Barriers and Policy Recommendations.” Transportation Research Part A: Policy and Practice 77: 167–181. https://doi.org/10.1016/j.tra.2015.04.003.
International Association of Insurance Supervisors. 2024. Insurance Core Principles and ComFrame, December revision. https://www.iais.org/activities-topics/standard-setting/icps-and-comframe/.
International Accounting Standards Board. 2017, as amended. IFRS 17 Insurance Contracts, effective for annual periods beginning January 1, 2023. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/.
International Organization for Standardization and SAE International. 2021. ISO/SAE 21434:2021: Road Vehicles—Cybersecurity Engineering. https://www.iso.org/standard/70918.html.
International Transport Forum. 2017. Shared Mobility Simulations for Helsinki. OECD/ITF, October 12. https://www.itf-oecd.org/shared-mobility-simulations-helsinki.
National Association of Insurance Commissioners. 2025. “Commercial Ride-Sharing,” updated December 18. https://content.naic.org/insurance-topics/commercial-ride-sharing.
National Highway Traffic Safety Administration. 2025–2026. “Standing General Order on Crash Reporting: ADS and Level 2 ADAS,” including 2025 amendment and available incident data to September 15, 2026. https://www.nhtsa.gov/laws-regulations/standing-general-order-crash-reporting.
SAE International. 2026. J3016_202609: Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles, September 20 revision (publisher current-standard catalog). https://saemobilus.sae.org/standards/recommended-practices.
United Kingdom. 2018. Automated and Electric Vehicles Act 2018, Part 1, § 2. https://www.legislation.gov.uk/ukpga/2018/18/section/2.
United Kingdom. 2024. Automated Vehicles Act 2024, enacted text, including §§ 3–6 and commencement provisions. https://www.legislation.gov.uk/ukpga/2024/10/body/enacted?view=plain.
United Kingdom Vehicle Certification Agency. 2026. “Cyber Security and Software Updating,” updated May 22. https://www.vehicle-certification-agency.gov.uk/connected-and-automated-vehicles/cyber-security-and-software-updating/.
World Health Organization. 2023. Global Status Report on Road Safety 2023. Geneva: WHO. https://www.who.int/publications/i/item/9789240086517/.
Angela Bogdanova. 2026. “Automobile Insurance.” Insurance by Bogdanova, Science, first published November 7. Published by Aisentica Development. https://insurancebybogdanova.com/science/automobile-insurance/.