Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Liability Insurance for Autonomous Products and Connected Devices

Abstract. Liability insurance for autonomous products and connected devices is shaped by the separation of physical control, software authorship, operational supervision, evidentiary access, and the legal obligation to compensate injury. A connected medical device, industrial robot, vehicle, domestic controller, or remotely managed machine may cause a single bodily injury through interacting hardware, embedded software, cloud-dependent functions, compromised communications, environmental conditions, and human decisions. The insured event emerges from this causal system, while applicable tort law and insurance coverage assign responsibilities according to distinct rules. This monograph analyzes the resulting insurance problem across product-defect doctrine, the allocation of operator and manufacturer responsibility, digital causal reconstruction, contractual attachment and exclusions, jurisdictional variation, risk measurement, and correlated accumulation. Its principal documentary foundations are the European Union's revised product liability law and 2026 corrigendum, related EU technology regulations, the UK product liability and automated-vehicle statutes, selected United States product liability doctrine and federal vehicle reporting, the NIST and ISO/SAE governance literature, international insurance supervision, and peer-reviewed law-and-economics scholarship. The analytical method distinguishes legal force, effective date, covered damage, evidentiary probability, insurer obligation, and modeled loss. An explicit numerical illustration separates expected injury losses, the insured share, defense expenses, and contingent common-cause software losses. The synthesis finds that insurability depends on the interaction of legal attribution, observable product versions, contractual definitions, and the dependence structure of losses. Insurance can finance clearly specified liabilities and support prevention through underwriting conditions, while data rights, update governance, claimant access to evidence, and limits on recoverability must be separately established. The inquiry is international and comparative; hypothetical values are not market estimates, and no proprietary claims dataset or novel authorial theory is asserted.

Keywords: Autonomous products; connected devices; product liability; liability insurance; embedded software; Internet of Things; autonomous vehicles; cyber-physical systems; manufacturer responsibility; operator responsibility; product defects; causation; digital evidence; cybersecurity; insurance coverage; claims-made; occurrence; risk accumulation; reinsurance; European Union law; United Kingdom law; United States tort law.

Introduction

A digital command has consequences in physical space when a product can steer, dispense, heat, brake, lift, monitor, or otherwise act upon its surroundings. The commercial structure around that command often includes a device manufacturer, a component supplier, a software provider, a network intermediary, a system integrator, a customer organization, and an individual user. Product liability asks which legally relevant defect or act caused compensable harm. Liability insurance asks whether a defined insured has an obligation within the insuring agreement, whether that obligation attaches to a policy period, and how defense, indemnity, limits, and recovery are allocated. A scientifically adequate account must analyze those questions in sequence.

The expansion of software and connectivity alters the available causes of loss. A manufacturer may deploy a software update after sale; a remotely supplied component may become indispensable to a safety function; an operator may change a configuration; a shared library or signing infrastructure may link thousands of devices. The revised European Product Liability Directive formally addresses software, related services, manufacturer control, and post-market change, while Buiten (2024) analyzes how risk awareness and practical control affect the efficient allocation of responsibility in defective AI systems. Those different kinds of source must remain distinct: a legislative rule can impose duties, while a scholarly economic argument illuminates incentives.

The object of this Research Dossier is liability insurance exposure created by commercial and noncommercial use of autonomous products and connected devices in international markets. Its primary focus is injury and third-party property damage, with associated defense costs, cyber-physical losses, privacy/data-related claims where separately covered, and financial loss where the governing form responds. It examines accident liability rather than general intellectual-property insurance, ordinary first-party electronic equipment cover, or generalized actuarial reserving. The inquiry is not a proposed universal insurance contract or a statement that all connected products are subject to one mandatory standard.

The analysis uses a documentary comparative method: identification of operative and prospective primary instruments; direct inspection of official guidance and scholarly arguments; mapping of legally relevant facts to policy elements; sensitivity calculations with expressly hypothetical parameters; and source-to-claim audit. The legal cutoff is November 16, 2026. Changes in national transposition, commenced UK provisions, policy form editions, and future regulatory guidance must be checked against the incident date in any real claim. No external peer review or empirical causal inference from industry claims records is represented here.

Product Defect and Operator Responsibility

Defect as a legally specific safety judgment

A product's failure is evidence of a physical outcome; defect is a legal characterization that depends on statutory language, foreseeable use, safety expectations, risk-utility analysis, available alternatives, and other facts. In Great Britain, the Consumer Protection Act 1987, section 3 asks whether the safety of the product meets what persons generally are entitled to expect and refers to marketing, warnings, reasonably expected use, and time of supply. This provision provides an evidentiary starting point for UK product claims. It should not be recast as an autonomous-technology-specific statutory presumption.

Buiten (2024) argues in Product liability for defective AI that manufacturer and user knowledge and control are central to accident-prevention incentives. This analysis yields an important insurance inference: underwriters need a reasonably observable account of who can actually mitigate a given hazard, even when litigation may place responsibility differently. For example, a supplier may control remote model updates while an industrial operator controls site speeds and access barriers. The hazards are coupled, and neither party can be assessed solely from its legal label.

A safety case normally combines hazard identification, foreseeable misuse, validation under declared operating conditions, change control, monitoring, and incident response. It permits comparison between the intended and observed behavior of a device, but it does not eliminate residual uncertainty. An injury following an anomalous action may arise from software logic, sensor obstruction, training data, a malicious update, foreseeable operator misuse, or an external force. Expert evidence must test competing explanations before attributing a defect to an insured.

The revised EU product-liability architecture

Directive (EU) 2024/2853 includes software among products; it defines related digital services, manufacturer control, and substantial modification. Its definitions and liability provisions extend the analysis beyond a sealed physical object to authorized integration of components, updates, and services. Article 6 recognizes specified death or personal injury, eligible property damage, and destruction or corruption of nonprofessional data; it does not convert every business interruption or economic loss into a covered statutory claim. Article 7 assesses defectiveness through the safety that persons are entitled to expect or that applicable law requires, considering circumstances including product characteristics and subsequent matters where legally relevant.

The dates require unusually careful handling. The official May 2026 corrigendum changed Article 2(1) to products placed on the market or put into service after December 8, 2026. December 9, 2026 is consequently the first calendar day in the corrected prospective application window. Article 22 requires Member States to transpose by December 9, 2026, and Article 21 maintains the former directive's relevance to earlier products. The Commission's product liability implementation overview confirms the transposition deadline. As of this monograph's November 16, 2026 research cutoff, the revised framework's principal future product cohort has not yet begun. A claim discovered in 2028 concerning an older product cannot be allocated to the new instrument merely because the litigation is new.

The revised law also contains procedural provisions that influence insurer defense strategy. Articles 9 and 10 provide mechanisms for court-ordered disclosure and specified presumptions concerning defects or causation, including cases of technical complexity and failure to disclose relevant evidence. Those are conditional judicial mechanisms, subject to their stated tests and rebuttal, and not a legal presumption that every unexplained autonomous-system accident establishes manufacturer liability. Their prospective operation and national implementation can affect the evidentiary cost and probability of liability, matters that must be priced separately from technical incident frequency.

Software updates, service dependencies, and substantial modifications

The risk owner for a connected product can change as a result of updates and control arrangements. A manufacturer that authorizes third-party components or retains the ability to supply safety-critical updates has a different exposure from a user that installs unauthorized software outside manufacturer control. The revised EU manufacturer-control definition and substantial-modification provisions make that factual distinction legally material in their field of application. A cloud service that is necessary to a product function may meet a related-service definition under the statutory conditions; ordinary internet access is treated differently in the directive's recitals.

Underwriting should identify not merely whether automatic updates occur, but who signs releases, which devices receive them, how deployment is staged, whether rollback exists, how long support continues, and whether critical patches can be delayed by customers. Where a device operates offline, loss of connectivity may become a design risk if the safe state depends on remote instructions. The NIST IoT baseline supplies an official starting point for technical capabilities; the manufacturer lifecycle guidance supports evaluation of support and post-market practices. Neither source settles whether the contractual insured has a duty to maintain a particular patch in a given jurisdiction.

Operator negligence, supervision, and the human handoff

Operator responsibility enters where workers, clinicians, fleet supervisors, household owners, or integrators select a setting, disregard a warning, exceed an operating domain, or fail to maintain a machine. In partially automated vehicles, the monitoring duty differs from the expectations of a vehicle legally authorized to drive without continuous human supervision. Schellekens (2022) provides detailed control-transfer analysis under the earlier EU regime; its conclusions are scholarly reasoning rooted in its 2022 legal context, not statements of post-2026 national implementation.

Great Britain's Automated and Electric Vehicles Act 2018 establishes a distinctive route: under section 2, for specified accidents involving an insured automated vehicle driving itself, the vehicle insurer is liable for defined damage, subject to the Act. Section 4 addresses prohibited software alterations and certain safety-critical updates, with special limits on exclusions and recovery. Section 5 addresses recourse against other liable persons. That structure demonstrates that prompt victim compensation and later allocation of liability to a producer are institutionally separable functions. It cannot simply be generalized to non-road robots, all vehicles with driver assistance, or jurisdictions outside its scope.

The Automated Vehicles Act 2024 further develops the UK authorization and user-in-charge architecture, with rules subject to section-specific commencement. Distinct treatment of a user-in-charge, an authorized system, and an operator of a no-user-in-charge service complicates the attribution of conduct but can also provide clearer information about intended control. For legal and insurance analysis, the exact authorization status and operative statutory provisions at the accident date must be recorded.

Defect, cybersecurity, and regulatory conformity

Safety and cybersecurity intersect where a compromised connection can cause physical actuation, disable alarms, falsify sensor data, or prevent patch delivery. Regulation (EU) 2024/2847 establishes a phased Cyber Resilience Act framework for products with digital elements. Article 14 reporting obligations apply from September 11, 2026, while the regulation's main application is scheduled for December 11, 2027, subject to its detailed scope and exceptions. The existence of a reporting obligation does not by itself establish a civil tort claim or a policy's liability to pay.

The revised EU Machinery Regulation addresses safety of machinery and related products, with main application from January 20, 2027. The EU Artificial Intelligence Act distinguishes, among other categories, high-risk safety components connected with product harmonization and conformity assessment; applicability and classification are governed by the actual consolidated text and phased rules. The law's product-safety classification is analytically relevant to a claim, yet insurance exposure still depends on the accident circumstances, plaintiff's cause of action, insured actor, and policy wording. A voluntary technical framework, a market-access obligation, and a civil compensation rule remain separate legal propositions.

Data and Causal Reconstruction

The event record as a legal and actuarial instrument

A claim involving an autonomous device normally depends on a time-ordered sequence: operating mode, environmental state, software version, configuration, sensor inputs, decisions, commands, physical response, injury, and subsequent preservation. A usable reconstruction needs calibrated clocks, event identifiers, integrity protection, versioned documentation, and clear distinction between observations recorded by the device and conclusions inferred by an investigator. A log that says a command was issued proves that a recorded command existed, subject to authenticity; it does not alone prove that the commanded movement caused an injury, that the command was unsafe, or that the legally responsible party was the software supplier.

Relevant evidence may reside on a device, a local gateway, a manufacturer cloud service, a fleet controller, or a hospital's operational records. The inquiry should identify evidence custodians, retention periods, time synchronization, data format, custody transfers, and possible conflicts between a claimant's access rights and the security of other users. A device that overwrites logs after seven days creates a different litigation and reserving profile from an equivalent device with authenticated historical traces. The cost of preserving and interpreting the evidence may become part of a policy's defense expenditure, depending on the wording and procedural law.

The NHTSA crash-reporting program illustrates the distinction between administrative incident information and an unbiased measure of insurable loss. Its 2025 amendment changed report structure, and the agency expressly cautions that awareness of incidents, telemetry coverage, reporting obligations, and duplicate reports differ between automated driving systems and Level 2 driver assistance. As of the reporting page examined in November 2026, displayed data extend through September 15, 2026. Report totals cannot be treated as comparable manufacturer accident rates without exposure denominators, system-engagement hours or miles, reporting probabilities, and deduplication.

Causal inference amid interacting systems

Multiple plausible sequences may generate the same observed accident. Consider a robot that unexpectedly changes direction. A sensor may have failed, the perception algorithm may have misclassified an object, a remote update may have changed braking thresholds, a floor map may have been modified, a human may have stepped into a restricted zone, or network latency may have affected an external control function. A defensible expert analysis identifies each hypothesis, the physical pathway linking it to injury, what evidence would discriminate among alternatives, and which assumptions remain untested. Temporal correlation between an update and an incident is informative, but it is not a complete causal account.

A practical investigation can compare event data with independent sensors, maintenance records, release histories, tests under equivalent configurations, and known fault modes. Counterfactual questions must be precise: would a timely patch have prevented the particular incident; would a warning have changed the operator's conduct; would a manufacturer-provided safe state have limited severity; and would the system have acted differently under the same observed inputs? These counterfactuals may be unidentifiable when logs are missing or when a model's behavior depends on unrecorded environmental conditions. Insurers should explicitly represent the resulting range of liability outcomes rather than turning a technical uncertainty into a fictitious measured probability.

The OECD AI incidents program recognizes that risk and harm emerge across the AI value chain. Its incident framework supplies governance context, not a population-level rate of legally compensable product defects. For insurance research, an incident database is best treated as a hypothesis generator until inclusion rules, reporting incentives, geography, product population, loss development, and observer duplication are specified.

A documented chain of preservation and access

A scientific claims file can preserve the original device image, firmware hash, configuration export, safety logs, cloud access records, software release notes, vulnerability disclosures, maintenance tickets, and incident scene documentation with verifiable collection timestamps. The dataset should retain distinction between source evidence and altered working copies, record investigator interventions, and document any known gaps. Security controls on digital evidence protect it from both adversarial modification and accidental destruction. These practices strengthen an insurer's ability to investigate coverage as well as a defendant's ability to respond to a claimant.

The revised EU directive's Article 9 evidence disclosure authorizes proportionate judicial disclosure in defined circumstances and protects legitimate interests under its conditions. Article 10's rebuttable presumptions can change the litigation consequences of missing records. A supplier's recordkeeping policy is therefore economically relevant to prospective defense costs and legal uncertainty, even when it does not alter the physical hazard frequency. For products placed on the market before the new directive's corrected prospective date, the applicable earlier regime and national procedural law must be used instead.

Privacy, trade secrets, and connected-product data rights

Access to operational records involves competing obligations. The EU Data Act overview records the main application of Regulation (EU) 2023/2854 from September 12, 2025 and its purpose of improving users' access to data generated through connected products. Those access arrangements are not automatically equivalent to litigation discovery, a third-party victim's data right, or an insurer's unilateral right to receive unrestricted personal information. The General Data Protection Regulation requires lawful processing, purpose limitation, minimization, appropriate retention, and security where personal data are involved.

An insurer's useful data request should be purpose-specific and access-controlled. The minimum viable incident record may include version, fault status, event time, actuation history, and relevant location, while sensitive audio, video, patient information, or unrelated user behavior may require additional justification. Contracts should specify lawful preservation and cooperation without promising disclosure beyond the parties' legal power. Technical confidentiality arrangements may allow expert inspection and courts to protect trade secrets. The appropriate design is a balanced evidence procedure with explicit permissions, not an indiscriminate data collection obligation.

Security incidents and the physical-damage boundary

Cybersecurity claims are particularly difficult when a vulnerability produces two classes of harm. Unauthorized access may cause loss of private data or a business interruption; it may also cause a machine to move unexpectedly and injure a person. The first class may engage privacy and cyber policies; the second may engage product, general, motor, or specialized liability cover. The actual damages, insured status, exclusions, and definitions decide which policies respond. No single use of the word cyber establishes that a cyber policy covers bodily injury or that a product policy excludes a remotely triggered accident.

The Cyber Resilience Act reporting rules and the NHTSA vehicle cybersecurity guidance supply distinct regulatory and voluntary evidence about vulnerability handling. The NHTSA guidance explicitly describes industry best practices as nonbinding. The ISO/SAE road-vehicle cybersecurity standard has a published lifecycle remit, but an insurer needs independent information about actual processes, testing and deployed safeguards. Underwriting questionnaires should seek demonstrable activities, disclosure timelines, patch release controls, and testing records rather than relying solely on a certification label.

Multi-Jurisdiction Contract Coordination

From civil obligation to an insured loss

An insured's legal responsibility and an insurer's contractual promise are separate. A claimant may allege negligence against an operator, strict product liability against a manufacturer, breach of warranty against a seller, and an information-security violation against a service provider. For every defendant, the insurer must evaluate status as a named or additional insured, relevant coverage part, allegations and actual facts under applicable defense law, covered damage, trigger, limits, exclusions, and any applicable conditions. A claim against four defendants is not necessarily four covered losses; conversely, a single physical occurrence may generate claims under different policy periods and forms.

Commercial general liability, product liability, products-completed operations, errors-and-omissions, technology liability, product recall, cyber, motor, professional indemnity, and first-party property policies protect different interests. Policy labels describe markets, not sufficient contractual terms. A general liability insuring agreement may respond to specified bodily injury or property damage caused by an occurrence, while the cost of repairing the insured's own defective device or conducting a preventive product recall may require separate coverage. Pure financial loss, impaired-product provisions, electronic-data definitions, contractual assumptions, and cyber exclusions can materially change the result. The analysis must refer to an authenticated, dated policy and endorsements before deciding coverage.

Defense and indemnity occupy different decision stages

The duty to defend, where present, may be triggered by allegations that potentially fall within an insurance grant, subject to governing law. The duty to indemnify depends on the covered obligation finally established or resolved. Some contracts require reimbursement of defense expenses instead of a direct duty to conduct the defense. Policies also differ on whether defense costs erode the liability limit and whether the insurer controls counsel, settlement, or expert selection. These differences matter especially in technically complex cases where forensic expense precedes a reliable causal conclusion.

A product may have a low probability of liability but a high probability of an expensive defense because the explanation for the failure is proprietary and contested. The apparent paradox is resolved by separating the probability of filing a claim, the probability of defense attachment, the conditional cost of defense, and the probability and severity of indemnification. Reserving models that treat defense as a fixed percentage of final indemnity can materially misstate this profile when claims are dismissed after costly expert work or when one common software defect generates extensive pre-settlement litigation.

Occurrence, claims-made, retroactive dates, and updates

An occurrence-form policy and a claims-made policy can attach to different periods for identical conduct. In a stylized occurrence-based product-liability case, injury during the effective period may matter even when the design and software release were earlier and litigation occurs years later, subject to actual policy definition and local law. In a claims-made technology errors-and-omissions form, the time a claim is first made, any reporting requirements, retroactive date, and related-claims wording may control. One must avoid equating the patch installation date, physical accident date, data discovery date, first written demand, insurer notice date, and court filing date.

Consider a device delivered in 2025, remotely updated in February 2027, involved in bodily injury in March 2027, and first the subject of a written demand in October 2028. The sequence identifies four different policy years and a potential transition between old and new EU product law only if the legal product-placement test, update or substantial-modification rules, and actual facts so require. The applicable legal cohort and the insurance contract's trigger do not automatically coincide. When two claims share one defective update, related-claims and aggregation clauses may deem them one event or claim for defined purposes; different wording may produce separate attachments.

The insurer-first motor example

The UK Automated and Electric Vehicles Act demonstrates an explicit statutory compensation path for listed automated vehicles driving themselves in specified circumstances. When section 2 applies, the vehicle insurer is directly liable for described damage, preserving other parties' liabilities and allowing relevant recovery. Section 4 treats certain software alteration and safety-critical update circumstances, while section 5 creates recourse. For the insurance economics of autonomy, this divides prompt injury compensation from ultimate defect attribution and recovery. Recoverability, litigation cost, and time lag then influence the insurer's net risk.

A policy response under the UK automated-vehicle regime differs from the manufacturer-liability assessment of a robot, an autonomous ship, a connected medical device, or a consumer appliance. In the United States, state-specific product liability doctrine interacts with state motor insurance and other statutory systems; generalized claims that every autonomous vehicle accident has shifted automatically to a manufacturer's product liability policy are legally unsound. The Automated Vehicles Act 2024 adds authorization-related categories, but section commencement and the vehicle's actual authorization status remain essential. Comparative work must preserve institutional details rather than transfer a national motor rule across product classes.

Manufacturers, software providers, and additional insureds

A complex product often crosses multiple insured organizations. The manufacturer may purchase products liability cover; the integrator may buy professional or operations liability; a cloud provider may buy technology errors-and-omissions and cyber cover; a commercial customer may insure premises and operations. Contracts among them may include indemnities, insurance procurement requirements, waivers, defense provisions, liability caps, and choice of law. These arrangements influence the financial allocation but cannot eliminate a nonparty claimant's statutory rights merely by private agreement.

The insurance analysis should distinguish direct statutory liability, legal indemnity under a supply contract, liability assumed through a contractual promise, and separately insured expenses. An indemnitee's defense may be reimbursable only under specified insured-contract or additional-insured language. One supplier's contractual liability cap may have exceptions for bodily injury, gross negligence, statutory product claims, or cybersecurity violations. A claimant can sue a manufacturer in one jurisdiction while manufacturer and supplier litigate their indemnity rights under another governing law. The policy must be interpreted alongside the actual claims and the law that governs each obligation.

Coverage mapping across legal systems

A comparative mapping begins with the accident location, the claimant's connection to each defendant, product distribution and service locations, the parties' contracts, relevant choice-of-law rules, mandatory insurance and consumer statutes, and service of process and judgment-enforcement paths. It then identifies every applicable policy's insured persons, territory, jurisdiction, loss definition, timing, notice, limits, deductibles, and exclusions. Only after this mapping does financial modeling meaningfully allocate expected cost among insurers, insureds, reinsurers, and uninsured claimants.

EU product liability rules under Directive (EU) 2024/2853 apply through Member State transposition and the corrected product-placement date; the UK product safety entitlement test remains a different statutory mechanism; and US product liability is broadly state-based, with multiple defect tests and procedural differences. The comparison concerns independent systems with partially overlapping goals. An insurance contract written in one market can respond to foreign litigation only within its terms, compulsory-law constraints, and enforceability of any jurisdiction clause. The location of a cloud server alone does not determine the law of all resulting physical injuries.

Primary, excess, reinsurance, and the net loss

A single event can penetrate a primary liability limit, an umbrella or excess layer, and one or more facultative or treaty reinsurance arrangements. Terms governing occurrences, related claims, batch defects, catastrophe events, attachment points, reinstatements, exclusions, and foreign-currency conversion determine the net retained loss. The insurer should not assume that a manufacturer policy's aggregation language and its reinsurance treaty's event wording are identical. A correlated firmware defect across several countries is especially likely to generate disputes over event count, cause, time period, and aggregation.

The IAIS risk-management principles call for effective controls and risk identification, including concentration and reinsurance risk in the supervisory framework. Their function is prudential, rather than resolving individual policy coverage. A defensible net-loss analysis therefore contains separate gross claim, allocated defense expense, insurer-share, salvage or contribution, subrogation, reinsurance recoverable, and retained cash-flow projections, with recoveries modeled according to enforceable rights and counterparty risk.

Evidence, Uncertainty, and Model Validation

What constitutes an exposure denominator

A physical device count is an initial portfolio measure, but claims frequency depends on the hours or cycles in which a hazardous function operates, environmental severity, deployment sector, product generation, software version, repair history, and monitoring intensity. Two fleets with 10,000 devices each can have radically different hazard exposure if one runs continuously in public spaces and the other runs for short periods in restricted factories. For connected devices, common software components and patch schedules must be recorded because products sold by different distributors may share the same vulnerability.

The observational unit should follow the underwriting question. For bodily injury from autonomous industrial movement, device operating-hours and proximity encounters may be useful; for driverless transport, relevant engaged-system miles, operational design domains, passenger journeys, and road conditions may matter; for battery thermal events, installed units, charge cycles and age may be necessary. A transition from device count to actual use entails measurement error and exposure lag. The NHTSA's reporting limitations warn that detection and telemetry differences can bias incident comparisons, so raw report counts provide a poor premium rate without a measured denominator.

Frequency, severity, and coverage eligibility

A transparent preliminary expected-loss decomposition is

E[Linsured]=N p c E[Ycovered∣event,eligible]+E[D]+E[O]−E[R],E[L_{insured}]=N\,p\,c\,E[Y_{covered}\mid event,eligible]+E[D]+E[O]-E[R],

where N is the exposed population, p is event probability per defined exposure period, c is the probability that a realized event produces a claim eligible under a specified policy, Y is indemnifiable severity after relevant deductibles and limits when modeled accordingly, D is defense cost, O is other contractually insured cost, and R is recoveries. This is a bookkeeping identity for an illustrative calculation when the terms are defined consistently. It is not a separately instituted insurance model. Correlation can leave the mean unchanged while greatly altering the distribution of aggregate loss and reinsurance attachment.

The incidence of all technical faults, the incidence of harmful accidents, the incidence of compensable harms, and the incidence of covered claims are different quantities. A device may fail safely and produce no loss. An injured person may have an actionable claim without any identified insured being covered. An insurer may fund a defense while later declining indemnity under the governing terms. Data used to estimate one stage must not silently be substituted for another. The difference between economic harm and insurer-paid loss is central to scientific inference and rate adequacy.

A reproducible hypothetical portfolio calculation

Consider an entirely hypothetical annual portfolio of 100,000 connected mobile products. Assume an annual bodily-injury event frequency of 0.0015 per device-year, mean gross compensable severity of $85,000, and a 60% probability that an event is eligible under one defined liability policy. Assume for simplicity that eligibility and severity are independent, that severity is already net of applicable deductibles and limits, and that each eligible claim entails $40,000 of defense spending in addition to indemnity. The expected technical event count is 100,000 × 0.0015 = 150; the expected eligible count is 150 × 0.60 = 90. Gross expected event severity across all incidents is 150 × $85,000 = $12,750,000. Expected insured indemnity is 90 × $85,000 = $7,650,000. Expected covered defense is 90 × $40,000 = $3,600,000. On these assumptions, expected indemnity plus defense is $11,250,000 before administrative expenses, capital load, investment effects, recoveries, and taxes.

The three dollar figures answer separate questions. The $12.75 million gross expected injury loss is a loss-cost measure before the stated policy-eligibility test; $7.65 million is indemnity in this hypothetical insurance contract; $3.60 million is expense. Their difference must not be labeled a measured insurance penetration gap, because the coverage probability is an assumption and some excluded losses may be insured elsewhere. A sensitivity change from 60% to 80% eligibility would increase expected covered counts from 90 to 120 and indemnity plus defense from $11.25 million to $15 million; the technical event count would remain 150. The comparison illustrates how contract interpretation and coverage mix can change insurer cost without changing device reliability.

A separate common-cause scenario demonstrates dependence. Assume a single hypothetical compromised software release reaches 0.8% of the same 100,000 devices, and 15% of the affected devices produce a $15,000 covered claim. Conditional on that release, 800 devices are affected, 120 claims arise, and covered claim amounts total $1.8 million before defense or limits. At an affected share of 8%, the conditional total rises to $18 million. If the first scenario has a 5% annual probability of occurring, its contribution to the unconditional mean is $90,000, but the $1.8 million conditional loss does not vanish from a tail scenario. These figures are invented to test analytical relationships; they describe no insurer, manufacturer, failure rate, or observed attack.

Correlation, tail dependence, and layer exhaustion

Independent-event calculations assume exposures contribute losses without a material shared driver. Connected products can violate that assumption through firmware, libraries, certificates, cloud orchestration, supplier dependence, industrial control platforms, or standardized sensors. A common cause can also yield several categories of claim: bodily injury, property damage, business interruption, forensic expense, product recall, and defense of multiple defendants. Which components are actually insured must be established before aggregating them. A common technical cause need not produce one legal occurrence or one reinsurance event.

Insurers should test the sensitivity of aggregate outcomes to the fraction of products sharing a component, the conditional probability of unsafe behavior once compromised, deployment lag, regulatory recall timing, downstream liability allocation, and the possibility of several simultaneous litigations. Portfolio diversification by manufacturer brand may provide little protection if multiple brands source the same critical control system. Correlation controls should thus examine software provenance and suppliers, not merely sector and geography. The IAIS cyber-risk resource collection and the IAIS risk-management framework identify systemic and concentration concerns in their respective supervisory contexts, but neither supplies an empirical distribution for this specific portfolio.

The probability of an excess-layer claim depends on the full distribution of aggregated loss and the exact layer terms. An increase in conditional severity can have a much larger proportional impact on the probability of piercing a high attachment point than on the mean of primary claims. Reinsurance analyses need stress assumptions for defense allocation, aggregate limits, event definitions, and collections risk. An insurer recording only expected annual loss may understate liquidity demand during a batch defect whose legal resolution is distributed over several years.

Reserving, legal development, and financial presentation

A long-tail product-liability claim may contain years between incident, notification, discovery of defect, litigation, adjudication, settlement, and ultimate recovery. A later safety recall may reveal additional affected cohorts, while a regulatory investigation may change evidence availability without determining civil liability. Case reserves and incurred-but-not-reported estimates therefore depend on claim development as well as coverage interpretation. In a portfolio with rapid technological version changes, the assumption that past accident years develop identically to future years is especially fragile.

The IFRS 17 measurement overview describes fulfilment cash flows based on current estimates and risk adjustment within the relevant accounting regime. This reporting framework does not determine whether a claim is covered under civil law or a particular liability contract. A financially coherent claim file tracks alleged damages, legally supportable settlement expectations, allocated and unallocated expenses, recoveries, currency, discounting where appropriate, and uncertainty associated with court outcomes. Where the insurer performs a legal reinterpretation or discovers a common defect, the financial estimate should be updated using the evidence and governing measurement rules.

Evidence quality and research design

The evidence base for this monograph includes official EU and UK legislative texts, US agency reporting and guidance, standards metadata, insurance-supervision material, official accounting overviews, university legal reference material, and scholarly articles. The strongest legal claims attach to directly inspected statute provisions; scholarly interpretations are attributed to their authors. Publisher metadata for a paywalled standard justify its stated title and scope but are insufficient evidence of detailed clauses or actual certification. An agency reporting dashboard documents reported incidents under administrative rules, not a randomized sample of global accidents.

A credible empirical follow-up would predefine device populations, engagement measures, type and timing of claims, severity definitions, claims inflation, country and policy form, product versions, claim reporting probability, common components, censoring, and exposure completeness. To compare two manufacturers, the analyst would need comparable reporting systems, deployment environments, and activity measures, then estimate uncertainty and conduct robustness tests for underreporting and missing data. Claims frequency might be modeled with Poisson or overdispersed counts conditional on exposure and covariates, while severity could be analyzed separately; dependence would need stress scenarios or data adequate to infer joint loss behavior. No such model is estimated or empirically validated in this documentary study.

Limits of inference and counterarguments

One argument expects greater producer responsibility because the manufacturer designs the system and controls safety updates. Another assigns a substantial role to operators who choose environments, configure equipment, and maintain barriers. Both identify genuine mechanisms. Their relative explanatory strength varies across the product lifecycle, accessible warnings, statutory regime, and practical capacity to prevent the particular harm. Buiten's producer–user control analysis illuminates this variation, while the UK automated insurer-first mechanism shows that the party required to pay promptly can differ from the ultimately responsible product supplier.

Some scholars argue that algorithms' unpredictability makes a conventional defect inquiry too burdensome. The revised EU proof and disclosure mechanisms respond to defined technical complexity but retain legal conditions and rebuttal. Conversely, it would be inaccurate to presume every autonomous failure is a defective product: safety performance is assessed under the applicable statute and the claimant's particular circumstances. A third view emphasizes that compulsory insurance can improve compensation even when causal reconstruction is costly; its institutional feasibility depends on jurisdictional legislation, premiums, recoveries, and the distribution of risk. The monograph draws no universal policy recommendation to mandate one product-liability insurance form across all connected devices.

Integrative Scientific Synthesis and Research Conclusions

Liability as a layered allocation of risk

The evidence yields a coherent descriptive finding. Control over physical operation, power to change software, access to incident records, legal responsibility for injury, and the insured obligation may sit with different organizations. Every transfer of responsibility must be justified by its own legal, technical, and contractual facts. An insurer can pay promptly yet have recovery rights; a manufacturer can owe compensation yet lack cover for a particular form of economic loss; an integrator can possess the decisive operational logs while a supplier owns the code. The economic importance of these separations grows as devices rely on post-sale updates and shared services.

The comparison also demonstrates a temporal mismatch. Safety obligations attach to design and deployment stages, product liability may apply by placement cohort, an occurrence policy may attach to injury, a claims-made policy may attach to demand and notice, and insurance accounting measures expected future cash flows at reporting dates. The corrected EU product liability commencement date is a concrete illustration: its prospective product-cohort test must be kept distinct from the claim and policy dates. This inference applies broadly to international research without creating a new authorial doctrine.

Causal evidence has both compensation and prevention value

Good logging and forensic cooperation improve the probability of distinguishing defect from misuse or malicious interference. Better evidence can reduce unnecessary dispute expense and inform targeted recalls, but it can also reveal previously unrecognized liabilities and increase valid claims. Therefore, a blanket assumption that more telemetry always lowers insured loss is unjustified. The magnitude and direction depend on the reliability of the data, the applicable burden of proof, claimant access rights, contract language, and prevention response. The NHTSA reporting caveats provide a real institutional reminder that richer data availability can raise observed incident counts through better detection.

The evidence-access problem is inseparable from privacy and security. A device manufacturer may require sensitive records for root-cause analysis, while its lawful processing and disclosure duties remain constrained by the GDPR's data principles and other applicable law. The Data Act access framework addresses specific connected-product data rights rather than overriding every separate restriction. Insurers should consequently evaluate technically demonstrable preservation procedures and contractual cooperation routes, alongside lawful limits on disclosure.

Insurance design follows the physical and legal loss pathway

Insurance cover for a cyber-induced physical accident cannot be inferred from the word connected. Coverage analysis begins with the injured person, type of damage, legally responsible defendant, and the policy's actual terms. It proceeds through temporal triggers, defense duties, insured status, exclusions, limits, allocation, and recoveries. The same defective update can produce a bodily-injury liability claim, direct damage to the device, a third party's business interruption, and an urgent recall expense. Such components may belong to different insurance products, or some may remain uninsured. The distinction should be explicit in underwriting materials and claims records.

This mapping also clarifies how engineering evidence informs pricing. A versioned component inventory, documented vulnerability management, measured device exposure, safe-failure design, and incident logs are risk characteristics. The NIST IoT device baseline, NIST AI RMF, and ISO/SAE cybersecurity engineering standard contribute different evidence about controls. They support structured inquiries and audits; they are not automatic underwriting guarantees. Contract language must determine whether cybersecurity incidents, data impairment, recalled products, and third-party physical injury are included, excluded, or subject to special limits.

The relevant international differences persist

The EU's prospective revised product liability framework explicitly incorporates software and related services; the UK's product regime retains a distinct entitlement-to-safety test and automated motor compensation structure; US tort law remains materially state-specific. The AI Act classification rules, Cyber Resilience Act, and Machinery Regulation add regulatory layers in the EU, with phased dates and overlapping subject matter. A global insurer must treat these differences as variables in the assessment, not noise to be averaged away. Contractual indemnities, coverage limits, and reinsurance agreements further diversify the ultimate payer.

A diversified insurance market can spread losses across many product users when accident probabilities and payout obligations are tractable. The most challenging exposures arise when the population of insured devices is incompletely observed, shared components create correlated failures, event definitions are disputed, and liability litigation develops after many software versions have changed. The hypothetical portfolio calculation illustrates that coverage decisions alter expected insurer payments independently of technical failure frequency, while a common-cause scenario creates conditional tail loss out of a small affected fraction. The calculations are an analytical demonstration, not evidence of observed premiums or claims.

Research priorities arising from the evidence

The immediate research priorities are documented populations of deployed devices and active-function exposure, standardized reporting of device and software versions, empirical comparison of accident and claim outcomes, forensic trace validity, and matched policy-form interpretation. Comparative legal work needs jurisdiction-specific implementation of Directive (EU) 2024/2853 after the corrected application date, detailed commencement of UK automation provisions, and practical treatment of evidence orders and rebuttable presumptions. Insurance analysis needs contracts with known vintage, exposure-based frequency, conditional legal-defense cost, aggregate software event stress, and recoveries including reinsurance. Those priorities follow directly from the gaps identified in current sources and remain testable research questions.

Conclusion

Liability insurance for autonomous products and connected devices covers an evolving intersection of product safety, operational responsibility, software maintenance, cyber-physical causation, civil compensation, and insurance contract design. Its central scientific requirement is accurate differentiation among an unsafe product, a legally proved liability, a contractually covered obligation, and an insurer's ultimate payment. In connected systems these stages may involve several institutions and different dates, and their proper sequence makes both litigation analysis and actuarial valuation more reliable.

The examined primary sources establish material divergence among legal regimes. The revised European Product Liability Directive recognizes software, related services, updates, and specified evidentiary mechanisms, with its prospective product application corrected to dates after December 8, 2026. The United Kingdom provides both its existing product safety test and a distinctive insurer-first mechanism for eligible automated road accidents. United States product-liability actions depend on state doctrine and applicable federal safety-reporting evidence. Technology risk standards and insurance supervisory principles inform prevention and governance while retaining their separate normative roles.

The original analytical synthesis is that the insurable risk is best understood through precisely documented product functions, legal and contractual attachment, evidentiary access, and loss dependence. Improvements in safety governance, data integrity, and policy drafting can make these exposures more intelligible; their effect on loss costs must be evaluated rather than assumed. The hypothetical calculations show how coverage eligibility changes expected indemnity and how shared software vulnerabilities alter aggregate-tail exposure. Continued comparative legal work and carefully sampled real claim and operating-exposure data are required before estimating population rates, market adequacy, or causal effects. This completed documentary monograph institutes no new authorial theory or doctrine.

References

  • European Parliament and Council of the European Union. (2024). Directive (EU) 2024/2853 on liability for defective products. Original source

  • European Parliament and Council of the European Union. (2026). Corrigendum to Directive (EU) 2024/2853, OJ L 2026/90364, May 7, 2026. Original source

  • European Commission. (2026). Liability for defective products: overview and transposition. Original source

  • European Parliament and Council of the European Union. (2024, consolidated 2026). Regulation (EU) 2024/1689 laying down harmonized rules on artificial intelligence. Original source

  • European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/2847 on horizontal cybersecurity requirements for products with digital elements (Cyber Resilience Act). Original source

  • European Parliament and Council of the European Union. (2023, consolidated 2026). Regulation (EU) 2023/1230 on machinery. Original source

  • European Parliament and Council of the European Union. (2023). Regulation (EU) 2023/2854 on harmonized rules on fair access to and use of data (Data Act): EUR-Lex summary. Original source

  • European Parliament and Council of the European Union. (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). Original source

  • Parliament of the United Kingdom. (1987). Consumer Protection Act 1987, Part I, section 3: Meaning of defect. Original source

  • Parliament of the United Kingdom. (2018). Automated and Electric Vehicles Act 2018, Part 1. Original source

  • Parliament of the United Kingdom. (2024). Automated Vehicles Act 2024. Original source

  • Cornell Law School Legal Information Institute. (2026). Products liability (Wex legal encyclopedia). Original source

  • National Highway Traffic Safety Administration. (2026). Standing General Order on Crash Reporting: ADS and Level 2 ADAS. Original source

  • National Highway Traffic Safety Administration. (2022). NHTSA updates cybersecurity best practices for new vehicles. Original source

  • National Institute of Standards and Technology. (2020). NISTIR 8259A: IoT Device Cybersecurity Capability Core Baseline. Original source

  • National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework 1.0. Original source

  • International Organization for Standardization and SAE International. (2021). ISO/SAE 21434:2021, Road vehicles — Cybersecurity engineering. Original source

  • International Association of Insurance Supervisors. (2024). Insurance Core Principles and ComFrame. Original source

  • International Association of Insurance Supervisors. (2026). Cyber risk and operational resilience: IAIS resources. Original source

  • International Accounting Standards Board. (2023). IFRS 17 Insurance Contracts. Original source

  • Buiten, Miriam C.. (2024). Product liability for defective AI. European Journal of Law and Economics 57:239–273. DOI 10.1007/s10657-024-09794-z. Original source

  • Schellekens, Maurice. (2022). Human–machine interaction in self-driving vehicles: a perspective on product liability. International Journal of Law and Information Technology 30(2):233–248. DOI 10.1093/ijlit/eaac010. Original source

  • Henderson, James A., Jr., and Aaron Twerski. (1998). Achieving Consensus on Defective Product Design. Cornell Law Review 83(4). Original source

  • Organisation for Economic Co-operation and Development. (2026). AI risks and incidents: OECD topic and incidents monitor. Original source

  • National Institute of Standards and Technology. (2026). NISTIR 8259 Series (including 8259 R1, revised April 2026). Original source