Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Algorithmic Liability Underwriting and Proxy Fairness

Abstract. Algorithmic liability underwriting classifies the anticipated frequency, severity, defense cost, timing, and insurability of third-party claims using statistical models, automated rules, and external data. Its central fairness problem is the possibility that a formally neutral variable or decision process carries protected-characteristic information, historical exclusion, or a spurious correlation into acceptance, premium, limit, deductible, exclusion, or renewal outcomes. This monograph develops a comparative, source-grounded account of that problem across commercial general liability, professional liability, employment practices liability, product liability, and related casualty insurance. It distinguishes a proxy's statistical association from its causal function, its predictive contribution from its legal permissibility, and insurance underwriting conduct from liability arising out of the insured's own activities. It analyzes risk classification, selective labels, missing protected-characteristic observations, disparate outcomes, calibration, counterfactual comparisons, error-rate disparities, explainability, contestability, vendor governance, and contractual translation. Primary regulatory materials from New York, the NAIC, the EU, and the United Kingdom are read alongside actuarial standards and mathematical fairness research. Two fully specified synthetic illustrations show how portfolio composition can produce disparate outcomes despite identical within-stratum claims rates, and how a restriction on proxy use changes a risk-selection decision. The synthesis concludes that meaningful proxy review needs a documented chain from source data to insured exposure, model output, underwriting action, applicable legal standard, and available remedy. The work offers original comparative analysis and reproducible illustrations without asserting a new scientific theory, proprietary fairness metric, or empirical estimate of actual insurer performance.

Keywords: Algorithmic underwriting; casualty insurance; liability insurance; proxy discrimination; indirect discrimination; disparate impact; actuarial fairness; insurance pricing; risk classification; artificial intelligence governance; model validation; commercial insurance; causal inference; automated decisions; explainability; insurance regulation; contestability.

Introduction: The Liability Underwriting Problem

Liability insurance prices contingent obligations arising from alleged or legally established harm to third parties. Algorithms can help classify the insured's operations, premises, products, professional activities, workforce, territory, litigation environment, and past loss experience. These classifications may determine whether coverage is offered and on what contractual terms. The decision has several subjects: the person or organization seeking coverage, persons affected by its conduct, the insurer bearing a defined part of its risk, and the wider market in which selection may make cover unavailable to some applicants. A fairness analysis must specify which subject and outcome it addresses.

The point of departure is the difference between estimating a loss distribution and deciding how an insurer may act on that estimate. The first calls for predictive and actuarial evidence. The second also requires contractual interpretation, regulatory authorization, distributive judgment, and a procedure through which errors can be discovered and corrected. The NAIC model bulletin on AI expressly connects insurer AI governance to existing unfair-trade-practices and property-and-casualty rating standards. It provides a model for state adoption; the model document itself has no automatic force in every US state. Its practical significance is that algorithmic technique does not displace the law governing the underlying insurance decision.

A typical liability underwriting chain begins when a broker or applicant submits exposure information. The insurer adds public, licensed, or vendor data; transforms those fields; estimates claim propensity, severity, or suitability; places risks into classes; applies capacity, wording, and price rules; and produces an offer, referral, or declination. Proxy fairness can become relevant at every transformation. A geographic variable may record physical hazard, local litigation conditions, a proxy for a person's protected status, or several of these at once. A profession or business category may express a genuine exposure mechanism while also reproducing historical barriers to business formation. A combined model can reconstruct a sensitive attribute from variables that look neutral when reviewed one by one.

The research questions are therefore concrete. How can an insurer establish that a model input measures an exposure relevant to the covered obligation? What evidence distinguishes a protected-status proxy from an acceptable risk predictor in a particular jurisdiction? How should the insurer compare legitimate risk differentiation, disproportionate adverse outcomes, alternative model specifications, and procedural remedies? This dossier answers with comparative legal reading, actuarial reasoning, fair-prediction literature, and transparent synthetic calculations. Its findings describe evidentiary and operational consequences rather than prescribe a universally binding legal rule.

Liability Insurance, Risk Classification, and Contractual Boundaries

The insured loss is a legally conditioned variable

A liability claim can be modeled as a sequence: an event or conduct, a possible injured claimant, allegation, defense, finding or settlement, allocation to an insured cause, attachment to a policy, and payment subject to limits, deductibles, exclusions, conditions, and exhaustion. An underwriting score predicting 'claims' may target reports, demands, incurred ultimate cost, defense spend, paid indemnity, or combined loss-and-expense. These labels are different statistical objects. The underwriting decision becomes difficult to interpret if the model developer, insurer, and affected applicant silently use the same word for different endpoints.

Let N denote the number of covered occurrences or claims in a fixed exposure period; let X_j represent ultimate indemnity and attributable covered defense expenses for claim j, measured according to an explicitly chosen policy and accounting convention; and let L = sum(j=1 to N) X_j. A pure expected-cost estimate is E[L | X], with X the admissible exposure information at the underwriting decision date. A premium quotation may add expenses, margin, reinsurance costs, taxes, and commercial constraints. A referral or declination may reflect portfolio appetite or contractual insurability rather than a simple threshold on E[L | X]. Accordingly, pricing, eligibility, wording, and limits each require their own outcome definitions and fairness assessment.

Liability underwriting includes general liability, employers' and employment practices exposures, products, professional errors and omissions, and specialized technology-related activities. Policy triggers, occurrence versus claims-made treatment, retroactive dates, claims-made-and-reported conditions, aggregation, defense-inside-limits, exclusions, and jurisdictions can change the expected insured loss while the insured's underlying conduct is unchanged. A generic risk model transported across these lines can be inaccurate even where it predicts a broad historical claims category well. Scientific comparison must therefore preserve the line, policy form, jurisdiction, underwriting year, and the development age of losses.

The ASOP 12 risk-classification standard addresses actuaries working on classification systems. Its professional relevance is the discipline of defining classes for the financial-security system in question and examining the credibility and interpretation of experience. The ASOP 56 modeling standard addresses model design, selection, evaluation, use, and appropriate validation in actuarial services. These are US professional practice standards with stated scope; they are not transnational legislative permissions to use otherwise prohibited attributes.

The underwriter's decision creates several distinct harms

Discriminatory or erroneous action may affect the insured directly through a premium increase, restricted limit, intrusive endorsement, exclusion, nonrenewal, or refusal. It may affect a third-party claimant indirectly if exclusion or inadequate insurance changes the resources available after injury. It may affect competition if certain small businesses systematically face reduced access to coverage. It may create loss for the insurer through adverse selection, legal challenges, supervisory sanctions, or unexpected claims. The policy covering the insured's third-party liability and any separate coverage of the insurer's own professional or regulatory exposures are separate contracts: an underwriting error does not automatically make the resulting damages an insured loss.

The relevant unit may be an incorporated enterprise, partnership, sole proprietor, employee, client, or household. Protected-characteristic testing designed for individuals does not transfer automatically to a corporation, although the circumstances of a sole proprietor can make a personal characteristic and business eligibility inseparable in practice. The New York DFS 2024 circular letter explicitly acknowledges that commercial property/casualty products may or may not produce a direct consumer impact and uses sole proprietors as an example. This observation prevents both a blanket exemption of commercial insurance and the unjustified assumption that every commercial risk is equivalent to personal insurance.

Coverage, timing, and selection

The underwriting model should retain a reproducible definition of the exposure base: annual revenue, payroll, employees, units produced, insured vehicles, transactions, occupied floor area, contract value, or another causal measure. Each denominator changes the meaning of frequency. In claims-made lines, reporting incentives and a policy's reporting conditions influence observed claims; in occurrence lines, discovery and settlement can extend many years beyond the underwriting period. Immature claims data can yield artificial geographic or sector differences if development rates vary by class. The fairness of decisions built on such labels depends first on their validity.

A particularly consequential feedback loop arises when the insurer rejects a class. Rejected applicants cease to generate that insurer's observed claim experience, so later model retraining treats the selected insured portfolio as if it were the full applicant population. Acceptance itself depends on earlier scores and proxy rules. This missing-outcome mechanism can turn an underwriting policy into a progressively narrower empirical basis for its own justification. The relationship is a standard selection-bias problem, and it requires separate data, plausible transport assumptions, or carefully bounded conclusions rather than an unobserved 'true loss rate' asserted from insurer records alone.

Evidence, Data Provenance, and Model Inputs

Inventory the complete decision chain

A scientifically reviewable underwriting record identifies the applicant and policy form, time of capture, original field, source owner, data collection purpose, legal basis where applicable, transformation, feature version, model version, calibration cohort, decision threshold, human override, and final action. Vendor labels such as 'community stability', 'business resilience', or 'management quality' need a data dictionary and a testable connection to the liability exposure. The analysis must distinguish an observed variable, an inferred score, and a human judgment embedded in a training label.

Data quality includes population coverage, measurement reliability, exposure consistency, duplicates, missingness, and whether an attribute was recorded before the decision or leaked from later events. Loss ratios collected after underwriting may be influenced by changes in deductibles, claims handling, reserve adequacy, inflation, defense strategy, and portfolio mix. A feature correlated with loss cost is not thereby evidence that it caused an insured hazard. Model users should first rule out administrative proxies for claims management, policy structure, or geography-specific reserving practice.

External data intensify accountability. The NAIC model bulletin expects insurers to address third-party AI systems within their governance program. New York's vendor oversight provisions similarly keep responsibility with the insurer and describe reasonable oversight and contractual audit cooperation. The operational implication is that a vendor's assurance is evidence to examine, not a substitute for an insurer's own decision record. Source licensing, confidentiality, and trade secrets must be managed together with regulator access and meaningful reasons for individual decisions.

Separate measurement from legitimacy

A variable can be statistically useful while legally restricted. Conversely, a lawful and intuitively appropriate variable can be measured poorly and introduce arbitrary classification. The audit should consider four questions in order. First, what exact information does the feature contain? Second, why might it predict the specified liability loss? Third, what protected-status association or historical exclusion does it carry, alone and jointly with other variables? Fourth, what does the jurisdiction's applicable law allow the insurer to infer or do with it? These questions are analytic stages; they do not imply that a particular correlation is automatically unlawful.

An interpretable traditional rating factor may outperform a complex external score in traceability even where their predictive accuracy is close. When risk engineering identifies a direct mechanism, such as maintained safety controls or documented equipment inspections, a model may gain both robustness and explanatory strength. This is a testable empirical proposition, dependent on data coverage, gaming risk, and measurement quality, not a claim that any particular safety questionnaire is inherently fair. The alternative feature must itself undergo proxy analysis, especially where compliance costs vary by business size or access to capital.

Missing protected-characteristic data

Insurers sometimes lack reliable group indicators. Ignoring groups is then a missing-data problem, not proof of equal impact. Imputation from names or geography introduces uncertainty and privacy obligations; it can misclassify mixed identities, smaller groups, and people with unusual histories. The DFS quantitative-assessment guidance contemplates available or reasonably imputed protected-status information while clarifying that it does not expect insurers to collect additional data merely to perform its exemplary analysis. Analysts should record the method, validation reference, uncertainty interval, access restrictions, and exactly which fairness metrics are identifiable with that proxy.

The ICO guidance on bias and discrimination explains why omitting protected characteristics from the deployed predictor does not by itself remove discriminatory consequences and why fairness analysis can sometimes require controlled processing of sensitive attributes. A principled test distinguishes protected data used strictly inside an access-controlled audit from protected data used to alter an individual operational decision. Lawful basis, minimization, retention, security, and permitted use remain jurisdiction-specific.

Proxy Variables, Discrimination, and Causal Identification

Four distinct meanings of a proxy

A proxy is a measured or constructed variable that stands in for another characteristic. In liability underwriting it may represent (i) a directly relevant exposure that is costly to measure, (ii) an unmeasured but potentially relevant exposure, (iii) membership in a protected class, or (iv) a historical decision process, such as prior access to insurance or claims litigation. One field may serve several functions simultaneously. Calling an attribute a 'proxy' describes an empirical or causal relation; it does not complete the legal assessment of its permitted use.

Consider location. A parcel's exposure to a known hazard may be relevant to premises liability; legal jurisdiction can affect defense cost; local claims-reporting patterns can affect observed experience; and neighborhood can correlate with race or economic status. Removing geography without considering the first three relationships may discard substantial signal, while retaining an opaque geographic score can reproduce the fourth. An adequate test examines the actual underlying mechanisms, the granularity of the variable, and whether direct measurements of the legitimate hazard can replace a coarse location label. Relative risks and fairness disparities should be estimated with uncertainty rather than inferred from raw maps.

Direct discrimination involves an explicit rule based on a protected trait where the relevant law prohibits that rule. Indirect or proxy discrimination can arise when neutral inputs carry that trait's influence into the outcome. Historical discrimination adds a third pathway: the recorded target is shaped by unequal prior treatment, so a model trained to emulate the target can reproduce prior exclusion without requiring an overtly sensitive input. Finally, selection bias can arise when the data reflect only those whom earlier underwriters accepted. These pathways have different remedies and different evidentiary burdens.

The discrimination-free pricing analysis of Lindholm, Richman, Tsanakas, and Wüthrich (2022) formally distinguishes direct and indirect discrimination and shows why simply excluding a protected characteristic from the final formula may leave information about it in other covariates. Their published mathematical proposal supplies a substantive research comparator; it is not adopted here as a mandatory insurance tariff or claimed as an original construction of this dossier. Its concrete applicability depends on policy line, risk distribution, privacy constraints, and which differences the law treats as impermissible.

Causal structure and competing explanations

A causal diagram can clarify competing hypotheses. Let A denote a protected attribute, H a measurable physical or operational hazard, Z a geographic or vendor proxy, Y eventual covered loss, and D the underwriting decision. Plausible paths include H → Z, H → Y, A → Z, Z → D, and A → historical economic opportunity → H. The same observed association between Z and Y may be explained by common causes, causal effects, differential measurement, or prior selection. A simple statistical regression of Y on Z and A cannot identify these mechanisms without assumptions, data, and a defensible estimand.

The counterfactual fairness paper by Kusner, Loftus, Russell, and Silva (2017) develops fairness reasoning under a structural causal model. Its appeal is that the analyst asks how a prediction would behave if protected status changed in a specified hypothetical world. But the conclusion depends on the assumed causal graph and on variables deemed descendants of that status. In an insurance case, legal coverage, operational processes, and the insured's own adaptation can complicate that graph. Counterfactual fairness is a research diagnostic requiring explicit causal assumptions, not a shortcut to an operative legal judgment.

A comparative analysis should therefore record at least three candidate explanations for a disparity: a legitimate difference in covered exposure, an impermissible proxy mechanism, and data-collection or selection distortion. Each explanation generates falsifiable implications where adequate observations exist. For example, if a postal code's premium effect disappears when independently verified safety controls and jurisdictional defense costs enter the model, the original code may have been a coarse composite proxy. If the effect persists, further investigation is required; persistence alone does not identify whether it is legitimate or prohibited.

The danger of path dependence

If an insurer repeatedly declines customers in a neighborhood, the next training set will contain fewer observations from that neighborhood. If claim frequency is then inferred from the remaining accepted risks, apparent stability may be a product of who remained eligible. This is selective-label bias. A risk-based business choice can alter future research evidence, so longitudinal validation needs applicant data, acceptance indicators, changes in underwriting criteria, and a clear account of which losses are observed. Where outcomes for rejected applicants cannot be collected ethically or reliably, the uncertainty must be stated rather than filled with invented experience.

A further complication is the difference between individual and structural fairness. Conditional comparison of otherwise identical firms can reveal arbitrary treatment but may normalize unequal access to capital or compliance technology when these are embedded in the conditioning variables. A demographic parity test may reveal unequal total outcomes yet conflate legitimate variations in insured hazard with the discrimination pathway. Both comparisons are informative; neither is self-interpreting. The recent statistical review of insurance fairness by Côté, Côté, and Charpentier (2026, advance online publication) places actuarial risk sensitivity, solidarity, and causal reasoning in the same methodological discussion. This dossier applies that distinction to liability underwriting and emphasizes its contract-specific and jurisdiction-specific interpretation.

Fairness Measurement and the Limits of Single-Metric Evaluation

Prediction and decisions are different endpoints

Let Y be an indicator of a defined covered loss event during a specified period, S a continuous predicted probability, A a group indicator available for lawful audit, and D a decision such as declination. Model calibration within a group means P(Y=1 | S=s, A=a) approximately equals s over reliably estimable score ranges. Acceptance parity concerns P(D=0 | A=a), while conditional acceptance parity compares decisions within documented legitimate risk strata. Error-rate parity compares false-positive and false-negative rates against an observed outcome definition. Each measures a different property, and every estimate depends on whether Y is observed consistently across applicants and groups.

The incompatibility shown in the risk-score trade-off theorem by Kleinberg, Mullainathan, and Raghavan (2017) establishes that several attractive conditions on risk scores cannot generally coexist when group base rates differ, outside special cases. The independent disparate-impact study by Chouldechova (2017) demonstrates tensions between calibration and error-rate balance. Both investigate predictive-risk problems rather than the full contractual and statutory structure of liability insurance. Their transferable result is methodological: a fairness scorecard requires an explicit choice of what it is meant to reveal, rather than claiming that one scalar proves an underwriting decision fair.

A workable evaluation separates three levels. Model performance: calibration, frequency and severity error, ranking, stability, and predictive uncertainty. Decision outcomes: distributions of acceptance, premium, deductible, limits, exclusions, and referral, including high-impact tails. Legal or ethical assessment: which distinctions are permitted, what justifications and alternatives are required, whether disclosure and review are adequate, and how the duty changes by jurisdiction. Statistical pass thresholds do not displace legal interpretation. A model can be accurate in aggregate and unlawful in a particular use; a legally permitted classification can still be unreliable for the exposure actually being underwritten.

Generalization and small populations

Casualty portfolios often contain few very large claims. Rare liability events make point estimates of group-level severe-loss risk unstable, particularly when protected categories, class of business, jurisdiction, and policy wording are cross-classified. High apparent disparity can arise from a small number of events, and genuine disparity can remain hidden behind broad confidence intervals. The audit must report sample size, exposure volume, claim development, confidence intervals or justified uncertainty bounds, and the pooling strategy. A non-significant test does not certify equal outcomes, especially when power is weak.

For binary outcomes, a group-specific rate estimate r_a = k_a / n_a has estimated binomial standard error sqrt[r_a(1-r_a)/n_a] under independent identical-trial assumptions. Insurance observations often violate those assumptions through clustering by industry and economic cycle, repeat exposures, common catastrophes, and delayed claims reporting; the formula is best understood as a first-order illustration. Bootstrap, cluster-robust uncertainty, hierarchical models, and development-aware actuarial methods may be appropriate where their assumptions are documented. The chosen unit must remain stable: policies, insured entities, exposures, dollars of insured value, and claimants are not interchangeable denominators.

Disparate outcomes also change with economic and legal conditions. Litigation trends, judicial decisions, inflation in defense expenditure, changes in safety standards, and the distribution of commercial clients can shift model calibration. Monitoring should disaggregate time drift from group composition and insurer decision drift. A periodic fairness measurement made only on bound policies can miss a change in the flow of rejected applicants; the applicant funnel should be included wherever lawful data permits.

Proxy detection methods and their interpretations

Mutual information or predictive recoverability of A from a candidate feature set can indicate that protected-status information remains encoded after a direct attribute is removed. Conditional association tests can evaluate whether a variable adds group information beyond verified exposure variables. Feature attribution can show how a model uses a field for prediction, while counterfactual perturbations can show how a decision changes when inputs are varied. None of these by itself establishes whether the changed input corresponds to a physically possible intervention or whether the law permits the resulting differential treatment.

The proper comparator must reflect the decision and the fairness question. To evaluate price, one may compare pure expected loss and final loaded premium for the same covered risk. To evaluate access, one needs declinations, coverage availability, and post-referral outcomes, not merely quoted price. To evaluate explanation, one needs decision-relevant factors and source provenance. To evaluate long-run exclusion, one needs applicant and renewal cohorts, including those never entering the insured portfolio. These are complementary observations tied to different causal and contractual questions.

Mathematical and Contractual Illustrations

All numbers in this chapter are constructed for transparent calculation. They are not records from insurers, estimated industry effects, measured demographic disparities, claims trends, regulatory thresholds, or recommendations for actual rates. Group labels A and B are abstract indicators used solely for mathematical illustration.

Illustration: different aggregate loss rates from identical conditional rates

Suppose two groups each contain 1,000 insured exposures, with high-hazard and low-hazard strata. In both groups, the event probability is 0.20 for the high-hazard stratum and 0.05 for the low-hazard stratum. Group A comprises 800 high-hazard and 200 low-hazard exposures; group B comprises 200 high-hazard and 800 low-hazard exposures. Expected events for group A are (800 × 0.20) + (200 × 0.05) = 170, or 17%; for B they are (200 × 0.20) + (800 × 0.05) = 80, or 8%.

Hypothetical group High-hazard exposure Low-hazard exposure Expected events Aggregate event rate
A 800 200 170 17%
B 200 800 80 8%

Within each hazard stratum, the assumed probabilities are identical across A and B. Aggregate parity fails because the exposure composition differs. This calculation demonstrates why a raw 17%-versus-8% comparison cannot itself attribute a lawful or unlawful cause. It also shows why simply conditioning on the hazard stratum does not answer whether that stratum was genuinely measured, whether its distribution was shaped by exclusion, or whether a legally prohibited factor produced it. A legal conclusion needs both the observed mechanism and the governing rule.

If each event had an assumed expected covered cost of $50,000 and expenses were omitted, expected aggregate insured cost would be $8.5 million for A and $4.0 million for B; the difference follows from the constructed frequencies, not from any asserted valuation of actual persons or firms. The sensitivity to cost severity is obvious: a twofold severity difference could reverse or intensify some premium comparisons, showing why comparing event rates without indemnity and defense-cost definitions is incomplete for liability lines.

Illustration: an unjustified proxy produces unequal declinations

Consider a second hypothetical portfolio with 200 applicants in A and 200 in B. Each group has the same 100 high-hazard and 100 low-hazard applicants; therefore each has the same assumed expected event rate: (100 × 0.20 + 100 × 0.05)/200 = 12.5%. Now suppose a vendor flag Z equals 1 for 160 applicants in A but 40 applicants in B. Within both groups, assume Z is independent of true high/low hazard and of Y conditional on that hazard. This independence assumption is supplied by the example, not inferred from an external dataset.

Let a flawed classifier assign S=0.125+0.030Z and let an underwriting rule decline every applicant with S greater than 0.15. Applicants with Z=1 receive S=0.155 and are declined; those with Z=0 receive S=0.125 and are accepted. The decline rates are 160/200 = 80% for A and 40/200 = 20% for B, although the assumed mean loss rate is 12.5% in each group and Z contains no additional legitimate signal. Removing the unsupported coefficient sets S=0.125 for everyone and removes the proxy-induced difference under this particular threshold. Revalidation is required because the simplified score ignores true individual hazard distinctions present in the hypothetical portfolio.

This is an intentionally stark failure case. It proves algebraically that a neutral-looking feature plus a threshold can produce disproportionate actions unrelated to the constructed target risk. It does not establish that every unequal acceptance ratio represents unfair discrimination; it depends on the stipulated data-generating process and the applicable legal rules. In real underwriting, a vendor factor can combine genuine exposure and sensitive information, leading to a harder problem than the zero-signal case.

A limit, deductible, and wording example

Assume two applicants have the same independently estimated ultimate liability loss distribution but obtain different automated outputs: a $2 million limit with a $25,000 retention for one, and a $500,000 limit with a $100,000 retention for the other. Equal predicted frequency alone would not establish equivalent treatment, because the contract transfers different layers of severity and defense expense. Comparison should model a common underlying liability distribution, attach both policy forms, and separately estimate insurer payment E[min(max(L-d,0),l)] only where this stylized retention-and-limit formula fits the actual contract. Defense outside limits, multiple claims, aggregate erosion, and allocation provisions require additional terms. A fairness review based solely on premium can miss a more consequential restriction in contractual protection.

Claims-made structures introduce a further issue. A class with elevated late reporting may look safer in a recent training period even if its ultimate exposure is higher. A model can then offer a low price but restrictive retroactive terms to that class; the observed premium gap appears favorable while the available indemnity is smaller. The audit should evaluate the joint bundle of price, coverage, duration, and underwriting conditions. This is an analytical implication of liability contract mechanics, not a claim derived from confidential insurer claims files.

Comparative Regulatory and Professional Standards

The United States: state insurance law and differentiated sources of authority

US regulation of insurance underwriting is principally organized through state statutes, regulations, examinations, and applicable federal protections. The NAIC's 2023 model bulletin is a shared supervisory template and should be distinguished from the adopted requirements of a particular state. It calls for governance, risk management, a written AI systems program, appropriate monitoring, documentation, and third-party oversight while grounding obligations in applicable insurance law. The NAIC adoption overview describes ongoing state implementation and regulatory work on third-party data and models. A statement that 'the NAIC prohibits a variable' would be incomplete without the controlling state legal provision and its scope.

New York supplies especially detailed interpretive guidance. The July 11, 2024 Circular Letter No. 7 addresses external consumer data, AI systems, and other predictive models used in underwriting or pricing. Its definitions expressly extend across property and casualty as well as health and life, and the department states that using a vendor does not eliminate insurer responsibility. Section II directs insurers to examine relationships between external data and protected classes and describes an assessment of disparate adverse effects, legitimate explanations, and less discriminatory alternatives where indicated. The circular distinguishes assessments with available or reasonably imputed group information from demands to collect new personal data. It also calls for pre-deployment and periodic testing and emphasizes disclosure of adverse-action reasons. Its legal effect is guidance about the department's interpretation and expectations under identified New York laws; each underlying mandatory provision must be identified for an enforcement claim.

The less-discriminatory-alternative inquiry is particularly significant for liability underwriting. If a broad location proxy predicts defense cost but a narrower legally relevant court-venue characteristic performs similarly with less protected-status association, the insurer can compare the alternative empirically. If a digital reputation score penalizes small enterprises but documented controls predict loss better, the insurer has a testable substitution question. The comparison must account for actual policy outcomes, predictive error, calibration, and data-collection burden; no source establishes an automatic safe harbor from merely running such a comparison.

Colorado illustrates why the regulated line matters. The 2025 Regulation 10-1-1 adoption notice identifies individual life, private passenger automobile, and health benefit plan insurers using specified external data, algorithms, and predictive models, effective October 15, 2025. This stated scope should not be silently expanded to all commercial general liability underwriting. Equally, a commercial carrier subject to another state rule cannot invoke Colorado's narrower named classes to avoid its own law. The correct review uses the actual authorized line, insurer, data practice, and jurisdictional provision.

Federal rules may apply by the nature of the data and the person affected. Under 15 U.S.C. § 1681m, a user taking adverse action based in whole or in part on a consumer report has specified notice obligations. Those obligations must be analyzed together with the Fair Credit Reporting Act's definitions and exceptions, including whether the applicant and report are within their scope. A commercial firm's financial score, a personal consumer report about a sole proprietor, and a non-consumer business-data product are distinct situations. Automated use does not erase a notice duty that otherwise attaches, and a generic 'credit-based factor' label does not prove that every business-data decision triggers the same federal provisions.

European Union: insurance governance, personal data, and changing AI rules

EU insurance and AI governance is layered. The EIOPA 2025 supervisory Opinion clarifies risk-based supervisory expectations under existing insurance-sector legislation concerning governance, records, fairness, explainability, cybersecurity, and human oversight. EIOPA expressly distinguishes that Opinion from obligations of systems already classified as prohibited or high-risk under the AI Act; the Opinion itself does not create a freestanding new statute. Its relevance to liability underwriting is the discipline of governance for systems that fall outside the AI Act's specific high-risk enumeration but remain subject to insurance-sector obligations.

The amended EU Artificial Intelligence Act identifies high-risk Annex III uses that include risk assessment and pricing for natural persons in life and health insurance. The classification does not automatically make every general or commercial liability underwriting model a high-risk Annex III system. Other triggers and prohibitions must still be assessed where applicable. As of the research cutoff, the EU's revised implementation timetable sets December 2, 2027, for stand-alone high-risk Annex III obligations and August 2, 2028, for relevant product-embedded high-risk obligations under Regulation (EU) 2026/1744. The future date is a legal implementation detail; it does not suspend other already applicable EU and national insurance, equality, consumer, or data-protection law.

Where underwriting involves identifiable natural persons, the EU General Data Protection Regulation applies on its own terms. Article 22 addresses decisions based solely on automated processing that produce legal or similarly significant effects, subject to conditions and safeguards; the regulation also addresses lawful processing, purpose limitation, transparency, data accuracy, and rights in other provisions. A human sign-off is meaningful only when the person can genuinely assess and change the result. Where the underwriting subject is a corporation without identifiable personal data, the GDPR analysis changes; a sole proprietor's application or a professional-liability score attached to an individual may present direct personal-data questions. The relevant automated decision and affected data subject must be identified before asserting an Article 22 right.

European product liability reform concerns a related but different object: the insured's exposure to product-related third-party claims, including software and AI as products in specified circumstances. The revised Product Liability Directive requires Member State transposition by December 9, 2026. The official May 2026 corrigendum corrects Article 2(1) to apply to products placed on the market or put into service after December 8, 2026. On November 16, 2026, that application condition is prospective. The Directive may affect the liability distribution an insurer underwrites in future product-liability lines; it is not itself a general antidiscrimination rule for casualty insurers' pricing models. Preserving this boundary avoids confusing two separate forms of algorithmic liability: liability generated by an AI-enabled product and fairness of the insurer's algorithmic risk selection.

United Kingdom: automated decisions and insurance product value

UK law has its own post-Brexit statutory text and should not be presented as a copy of the EU GDPR. Section 80 of the Data (Use and Access) Act 2025 substitutes the UK GDPR automated-decision-making scheme with Articles 22A–22D. The text distinguishes significant decisions with no meaningful human involvement and sets safeguards including information, representations, human intervention, and challenge in the circumstances specified. A UK insurer must assess current commencement, the applicable processing category and legal basis, and whether an actual underwriting decision meets the statutory significance condition. The data-protection account of automated decisions should not be imported unchanged from EU Article 22.

The Equality Act 2010 insurance provisions contain carefully bounded insurance-related exceptions. Their wording, affected protected characteristics, and reasonableness or actuarial-data conditions must be reviewed in their operative version and factual context. The existence of a statutory exception is not a general permission for any proxy or any insurance class. In addition, the FCA general-insurance review documents supervisory concern about product governance and fair customer value under PROD 4. The FCA's thematic findings are relevant to evaluating the entire contract, especially where an applicant obtains a cheap policy with materially narrower coverage. They are empirical supervisory findings about sampled firms and products, not a measurement of proxy discrimination in liability models specifically.

What comparative analysis establishes

The compared sources share a concern for accountable decisions, reliable data, fair treatment, oversight, and remedies, but the actionable duties differ. US state guidance can focus on unfairly discriminatory rates and insurer governance; EU rules additionally involve sectoral supervision and the GDPR's conditions for personal-data processing; UK law combines its own data-protection amendments, equality rules, and FCA conduct and product governance. The same underwriting model may be acceptable for one business-data-only use yet raise personal-data, rate-filing, or protected-class issues when transferred to a different jurisdiction or an individual applicant. A technically portable algorithm is not a legally portable underwriting decision.

Explanation, Recourse, and Governance of Underwriting Decisions

An explanation must describe the actual outcome

Decision explanations should identify the action, the version of the relevant risk model, the decisive data, the source of disputed information, and how correction or further review can occur. A global feature-importance plot averaged across policies cannot explain why a particular applicant was declined under a specific threshold, referral rule, or exception. Conversely, a reason code detached from the decision engine may say little about the score's operative causal role. Explanation can be tested by reproducing the exact output from the logged model version and input record and checking whether the displayed reasons correspond to features that actually changed that output.

New York's underwriting disclosure guidance says a carrier should provide specific information about adverse underwriting decisions and should not use vendor secrecy to justify insufficient explanation. That is a concrete supervisory expectation under named state provisions, not a claim that every jurisdiction compels disclosure of full source code. Good governance can preserve proprietary information while disclosing the data, material factors, error-correction channel, and decision rationale needed for an applicant to understand and contest its treatment.

Contestability requires more than an email address. It has an evidentiary path: the applicant identifies an inaccurate fact, the insurer retrieves the source record, a trained reviewer examines the decision and relevant overrides, the data provider corrects an error if established, and the insurer recalculates the quotation or explains an unchanged decision. The process should be able to change a result where warranted; otherwise nominal human review creates a misleading appearance of accountability. Where a legal notice right attaches, statutory deadlines and content prevail over an insurer's general service commitment.

Vendor dependencies and internal accountability

A carrier may use separately purchased business descriptions, predictive loss models, fraud flags, geographic classifications, and pricing software. Responsibility becomes diffuse if each vendor sees only its own component. An auditable final decision needs data-lineage identifiers connecting all relevant components and clear accountability for feature authorization, model validation, compliance review, and underwriting override. Contractual audit rights, retention duties, incident escalation, and correction obligations are useful controls where feasible, but the insurer's legal obligations depend on the jurisdiction. The idea that outsourced software eliminates carrier accountability conflicts with the NAIC governance expectations and New York's explicit guidance.

A governance inventory should record the purpose and permitted users of each model, each version's substantive changes, boundaries of the training population, monitored fairness groups where lawful, sources of validation, exceptions, and retirement conditions. Independent challenge should address both underperformance and the normative assumption embedded in the target. A model predicting past acceptance may be accurate as a predictor of historical insurer decisions while being irrelevant to the risk of future covered liability. The absence of documentation here is a scientific defect because the target itself is not established.

The voluntary NIST AI Risk Management Framework, published in 2023, organizes risk work around Govern, Map, Measure, and Manage. It gives an adaptable vocabulary for documentation and monitoring, not an insurance-specific legal compliance certificate. Its value lies in forcing model owners to articulate the use case, stakeholders, measurement basis, and actual decision. For an insurer, the framework must be populated with insurance contract definitions and jurisdiction-specific legal requirements rather than treated as a replacement for them.

Fairness as a continuing obligation of evidence

When inputs or model parameters change, the old fairness and performance evidence may no longer describe the deployed decision. A migration between vendor scores, a revision to a class code, a new catastrophe exclusion, or an updated claims-development assumption can change group outcomes without changing the model's marketing name. Effective governance requires version-specific validation, monitored drift, pre-deployment review of material changes, and a record of complaints and correction results. It must also explain the limits of group measurements where protected data are missing or unreliable.

An insurer can examine the effects of safer exposure variables, less discriminatory alternatives, and contract adjustments by simulating decisions on a documented applicant sample. Such comparisons should preserve technical fairness and commercial economics in separate columns. A more uniform acceptance distribution can result from tightening all offers, which may reduce unequal treatment but decrease coverage availability for everyone. A lower premium can coincide with a higher deductible. The substantive object is the total set of terms, with attention to who bears the residual liability risk.

Validation, Uncertainty, and Scientific Research Design

Source hierarchy and evidence units

This research distinguishes legislation and official legal corrections; binding regulations and operative national law; supervisory model documents and interpretive letters; professional actuarial standards; peer-reviewed mathematical and empirical scholarship; and this dossier's synthetic calculations. Each is used for a different inferential purpose. A statute establishes legal text, although application may depend on interpretive authorities and facts. A supervisory opinion describes an authority's expectations and scope. A peer-reviewed mathematical theorem establishes a conditional analytical result, not an insurance discrimination finding across every jurisdiction. A hypothetical computation establishes only the consequence of its premises.

The source review took place with a November 16, 2026 cutoff. Official source pages and scholarly publication records were checked for identity, subject, and scope. Where a page is primarily an abstract or record, the text does not attribute unverified detailed empirical estimates to the underlying full study. No proprietary insurer dataset, claims ledger, vendor training corpus, or regulator examination file was supplied or analyzed. No measured market prevalence, effect size, causal estimate, or litigation rate is inferred from the illustrations. Regulatory interpretations are bounded to the specified legal instruments and dates; later changes require a new dated review.

Reproducible empirical protocol for future testing

An insurer-specific study would identify the underwriting cohort by jurisdiction, product, date, and applicant eligibility; retain both submitted and bound risks; define each ultimate loss target; document the availability of group indicators; and freeze all transformations before outcome evaluation. It would split data by decision period, control leakage, measure development maturity, and compare out-of-time performance. Independent tests would estimate discrimination and calibration uncertainty by class and time, examine model feature relationships to protected traits, and test defensible alternatives. The study would separately record legal conclusions supported by competent jurisdictional analysis and observed statistical results.

For high-dimensional models, training data should be distinct from final validation and should record any model-search or feature-selection process performed before locking the candidate. For rare events, stratification and pooling may be needed, but pooling assumptions must be disclosed. Where applicants were denied and no valid counterfactual claims history exists, validation should state the population to which claims-based estimates apply and evaluate sensitivity to plausible selection mechanisms. Fairness analyses based on a representative selected sample cannot automatically generalize to all applicants.

Robustness analysis should ask whether conclusions change when the risk horizon, severity cap, loss-development factor, feature grouping, exposure unit, proxy-imputation method, legal classification, or contract layer changes. If a purported fairness gain depends on one precise target definition and vanishes when defense expense is included, it is fragile for liability underwriting. If the effect persists across independent risk definitions, it becomes more informative without automatically settling the legal characterization. These tests are proposals for empirical implementation, not claims that unpublished proprietary data were reviewed.

Internal verification and limitations

The arithmetic in the two constructed cases can be checked from their stated cells and does not depend on unverifiable external data. The first generates expected event counts 170 and 80 through different hazard mixtures. The second yields equal 12.5% base event rates and 80% versus 20% declinations through a deliberately unsupported three-percentage-point proxy coefficient. The coverage formula is expressly limited to a simplified one-claim layer and must not be transferred to complex defense or aggregate arrangements without modification. Bibliographic sources are traced to the chapter claims for which they are invoked; the accompanying EVIDENCE.json records primary-document scope, literature-record scope, and issues requiring renewed review if the work is materially revised.

The scientific limitations are substantive. Cross-country comparisons cannot turn guidance into binding law. An apparent group effect cannot identify discrimination without a defined rule and comparator. Accurate predictions do not automatically confer permission to use a variable. A group-level parity target cannot assure individualized fairness. Removing all protected features does not guarantee their information has vanished. A validated portfolio model cannot demonstrate fairness to rejected applicants when their outcomes are systematically absent. These are methodological conclusions grounded in the combined legal and statistical record.

Integrative Scientific Synthesis

Algorithmic liability underwriting is a meeting point of contract risk, actuarial inference, market access, and procedural accountability. The analysis supports six interrelated findings. First, a risk predictor requires a clearly specified insured-loss target and policy form. A liability score whose training target is inconsistent across claims-made periods, coverages, or defense-expense treatment is not an adequate foundation for a fairness claim. Second, proxy assessment is a multivariable and partly causal inquiry. The removal of an overtly protected column cannot certify that the remaining variables are neutral in effect.

Third, group disparity is an observation requiring explanation. The first synthetic illustration demonstrates a large aggregate difference compatible with identical within-stratum event probabilities. The second demonstrates a large declination disparity produced by a score term stipulated to contain no target information. The examples prove that both exposure composition and unjustified proxy dependence can produce visible disparity, and that separating them requires more than one statistic. Fourth, predictive calibration, equal treatment, and equalized classification errors are different goals with documented mathematical trade-offs. The policy objective and legal comparator determine which observations are material.

Fifth, jurisdiction matters at the point of decision. The NAIC model document, New York DFS letter, Colorado product-specific regulation, EU sectoral rules, amended AI Act, EU GDPR, and UK legislation carry different subjects and forms of authority. A vendor's common technology cannot make those distinctions disappear. Sixth, fairness attaches to an action with consequences: premium, availability, deductible, limit, exclusion, explanation, and recourse. Statistical assurance without a reconstructable decision trail offers little protection to an applicant challenging incorrect inputs or an unjustified referral threshold.

A practical consequence is to assemble evidence in the order in which the decision operates: data acquisition; exposure and coverage definition; target and model construction; proxy and performance testing; underwriting rule; final policy terms; jurisdictional review; and meaningful correction. This is an editorial synthesis of documented processes, not a new scientific theory, doctrine, or named model. It is designed to keep a mathematical statement, a legal statement, and an operational recommendation traceable to their distinct forms of evidence.

Conclusion

The fairness of algorithmic liability underwriting is determined through the relation between the covered liability exposure, information used to predict it, the actual insurance decision, the people and enterprises affected, and the governing legal and supervisory system. The scientific task is to establish these relations with valid labels, credible comparisons, explicit causal assumptions, and complete documentation. The institutional task is to make decisions reviewable and correctable while maintaining a lawful and economically coherent insurance product.

Current scholarship explains why correlated inputs can reconstruct protected information and why conventional fairness criteria may conflict. Actuarial practice requires sound classification and model validation. Official insurance and data-protection authorities require the analyst to identify the legal force, jurisdiction, scope, and date of each obligation. Together these findings support decision-level assessment covering access, price, terms, and recourse, with specific attention to selection effects and missing protected-status data.

Future empirical work should prioritize dated, multi-line insurer cohorts with both applicant and insured populations, audited external feature provenance, policy-form-specific loss development, and observed adverse decisions. It should report effects with uncertainty, distinguish lawful exposure variation from impermissible proxies under the operative rule, and test alternatives against the complete coverage bundle. No universal empirical claim or new normative doctrine is required to reach the central conclusion: a fair underwriting decision must be accountable for the information on which it acts, the liability risk it transfers, and the consequences it produces.

References

  • Actuarial Standards Board. (2005; last revised 2011). Actuarial Standard of Practice No. 12: Risk Classification (for All Practice Areas). Effective May 1, 2006. Official ASOP 12.

  • Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. Official ASOP 56.

  • Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163. DOI: 10.1089/big.2016.0047.

  • Colorado Division of Insurance. (2025, August 20). Notice of Adoption: Amended Regulation 10-1-1 Governance and Risk Management Framework Requirements. Official adoption notice.

  • Côté, O., Côté, M.-P., & Charpentier, A. (2026, November 12, advance online publication). A statistical perspective on fairness in insurance pricing. Annual Review of Statistics and Its Application. DOI: 10.1146/annurev-statistics-042925-075327.

  • European Commission. (2026). AI Act: Regulatory framework and revised implementation timeline. Commission implementation page.

  • European Insurance and Occupational Pensions Authority. (2025, August 6). Opinion on Artificial Intelligence Governance and Risk Management. EIOPA Opinion.

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

  • European Parliament and Council of the European Union. (2024, as amended in 2026). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence. EU AI Act.

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

  • European Union. (2026, May 7). Corrigendum to Directive (EU) 2024/2853 (Official Journal 2026/90364). Official corrigendum.

  • Financial Conduct Authority. (2024; updated 2025). TR24/2: General Insurance and Pure Protection Product Governance Thematic Review. FCA thematic review.

  • Information Commissioner's Office. (n.d.). Guidance on AI and data protection: What about fairness, bias and discrimination? ICO fairness guidance.

  • Kleinberg, J., Mullainathan, S., & Raghavan, M. (2017). Inherent trade-offs in the fair determination of risk scores. In 8th Innovations in Theoretical Computer Science Conference, LIPIcs 67, Article 43. DOI: 10.4230/LIPIcs.ITCS.2017.43.

  • Kusner, M. J., Loftus, J., Russell, C., & Silva, R. (2017). Counterfactual fairness. In Advances in Neural Information Processing Systems 30, 4066–4076. NeurIPS proceedings.

  • Lindholm, M., Richman, R., Tsanakas, A., & Wüthrich, M. V. (2022). Discrimination-free insurance pricing. ASTIN Bulletin, 52(1), 55–89. DOI: 10.1017/asb.2021.23.

  • National Association of Insurance Commissioners. (2023, December 4). Model Bulletin: Use of Artificial Intelligence Systems by Insurers. Adopted model bulletin.

  • National Association of Insurance Commissioners. (2026). Insurance Topics: Artificial Intelligence. NAIC AI overview.

  • New York State Department of Financial Services. (2024, July 11). Insurance Circular Letter No. 7: Use of Artificial Intelligence Systems and External Consumer Data and Information Sources in Insurance Underwriting and Pricing. NYDFS circular letter.

  • National Institute of Standards and Technology. Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. NIST AI RMF.

  • Parliament of the United Kingdom. (2010, as amended). Equality Act 2010, insurance-related provisions. Legislation.gov.uk text.

  • Parliament of the United Kingdom. (2025). Data (Use and Access) Act 2025, Section 80 and resulting UK GDPR provisions. Legislation.gov.uk text.

  • United States Congress. (2026 consolidated text). 15 U.S.C. § 1681m: Requirements on Users of Consumer Reports. U.S. Code.

  • European Parliament and Council of the European Union. (2026, July 8). Regulation (EU) 2026/1744 amending Regulation (EU) 2024/1689, and related instruments, as regards simplification of the implementation of harmonised rules on artificial intelligence. Official Journal of the European Union. Regulation (EU) 2026/1744.