Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Algorithmic Price Discrimination in Insurance Markets

Abstract. Algorithmic insurance pricing combines statistical prediction, contract design, consumer demand estimation, and competitive strategy within an institution whose defining product is contingent risk transfer. This monograph examines how increasingly granular information changes the distribution of premiums, insurance participation, selection, insurer solvency incentives, consumer surplus, and economic welfare. It distinguishes the prediction of expected claims from the estimation of willingness to pay and retention, and distinguishes both from contract characteristics that alter the coverage supplied. A formal decomposition of the premium separates expected losses, administration, risk and capital loadings, and demand-sensitive margins. Its central analytical finding is that predictive refinement, individualized prices, and improved welfare are logically distinct outcomes: the sign of welfare change depends on demand elasticity, the composition of the insured risk pool, market structure, access to information, cost of classification, and the objectives and legal limits imposed on pricing. This conclusion follows from established insurance economics and industrial organization rather than from a proposed new theory. Comparative analysis covers United States insurance regulation, the United Kingdom's renewal-pricing intervention and subsequent evaluation, European Union data and discrimination law, and international supervisory guidance. The empirical section separates observed regulatory findings from stylized calculations and explains identification requirements for causal welfare claims. A research-grade audit agenda connects model validation, market outcomes, consumer experience, and social distribution. The scope is international, with jurisdiction-specific legal conclusions expressly confined to their sources and dates.

Keywords: Algorithmic pricing; insurance price discrimination; risk classification; personalized premiums; price optimization; willingness to pay; adverse selection; consumer inertia; insurance demand; underwriting; actuarial fairness; competition; welfare economics; renewal pricing; insurance regulation; algorithmic governance; consumer protection.

Introduction

Insurance replaces a potentially large uncertain financial loss with an ex ante contractual premium. For a risk-averse person, the exchange can create value even when the premium exceeds the actuarial expectation of insured loss. The institution depends on risk pooling, the separation of covered and uncovered states, the enforceability of contracts, and a financial arrangement capable of paying claims. The price of insurance therefore affects several margins simultaneously: whether the contract is bought, which risks enter a pool, the level and form of protection, and the finances of the risk carrier.

Data-intensive pricing changes the resolution at which insurers observe these margins. A predictive system may identify more homogeneous expected-loss classes, a marketing system may estimate who is likely to accept a quote, and a renewal system may infer who is unlikely to search elsewhere. These systems may operate on the same customer record while answering different economic questions. The NAIC Price Optimization White Paper (2015) explicitly describes the distinction between actuarially indicated rates and granular adjustments informed by consumer price sensitivity. The FCA's subsequent evaluation (2025) provides unusually direct evidence of the consequences of price walking and the policy response in United Kingdom home and motor insurance.

The research object is the economic effect of algorithmic segmentation, personalized premiums, and dynamic insurance prices on consumer surplus, selection, and competition. The methodological unit is an insurance offer with an identifiable coverage definition, risk exposure, customer, distribution channel, time, and set of market alternatives. This prevents a nominal price change from being attributed to discrimination when coverage, deductibles, expected claims, distribution costs, or taxes have changed. It also prevents an apparently risk-neutral margin from being mistaken for a risk-based actuarial adjustment.

The analysis draws on three bodies of knowledge. Insurance economics explains asymmetric information and selection, particularly the conditions studied by Rothschild and Stiglitz (1976) and the empirically identifiable demand-cost relationships in Einav, Finkelstein, and Cullen (2010). Industrial organization establishes the economic meaning of price discrimination and its dependence on market power, reviewed by Varian (1989). Actuarial science and insurance law establish which risk characteristics and pricing practices can be used in a particular regulated market, a distinction developed by Avraham, Logue, and Schwarcz (2014).

The analysis uses a comparative institutional method, transparent algebra, internally reproducible hypothetical illustrations, and critical reading of published empirical estimates. Source-verified assertions are linked at their point of use. Mathematical examples are analytical devices, not estimates of actual insurer behavior. No proprietary policyholder dataset, machine-learning model, survey, or original causal experiment was collected. The contribution is an original synthesis of the interactions between cost-based classification, demand-sensitive pricing, market selection, and observed regulatory outcomes; it makes no claim to establish a new theory or new priority in scientific discovery.

Economic Foundations of Price Discrimination

Insurance as a contingent good

A premium buys a contingent claim, so prices across consumers can differ for reasons that would not constitute economic price discrimination for an identical good. Let individual i face a loss L_i with probability q_i and choose indemnity I_i subject to exclusions, deductibles, and limits. Under the simplest one-period assumptions, the conditional expected claim is c_i = q_i E[I_i | loss]. In practice, expected costs include claims severity, frequency, adjustment expense, allocated and unallocated expenses, taxes and levies, reinsurance, capital charges, and systematic uncertainty. Two people paying different premiums for materially different expected claims have purchased insurance capacity with different expected costs.

Economic price discrimination requires additional analytical care. A conventional definition is a difference in price not fully explained by the economic costs of serving different buyers of a comparable product. Varian's price discrimination survey (1989) distinguishes first-degree, second-degree, and third-degree discrimination, but the insurance application demands explicit standardization of risk and coverage before the classification is made. Individualized risk-based underwriting can coexist with a common percentage expense loading; individualized demand-based markups can coexist with identical expected loss costs. Both may be calculated by algorithms, yet their welfare implications differ.

Consider the following accounting decomposition for an offered annual premium:

P_i = C_i + A_i + K_i + M_i,

where C_i is expected insurance claims cost, A_i is expected service, distribution, tax, and administration cost, K_i is the priced risk or capital contribution, and M_i is the remaining competitive or demand-sensitive margin. This is an accounting representation for analytical purposes; the components need not be additive, separately observable, or constant in a firm's production system. Insurers can subsidize acquisition and earn margins through other periods, products, or investment returns. Reinsurance, policyholder dividends, fixed expense allocation, and cost uncertainty may redistribute what is economically classified as each element.

For an identical coverage and risk profile, a change in M_i arising from an estimated propensity to renew is the clearest case of demand-sensitive premium differentiation. A change in C_i because credible claims information has improved is a different operation. Changes in K_i may respond to portfolio concentration and capital needs, while changes in A_i may reflect distribution channels. Every empirical claim about discrimination should begin with an attempt to distinguish these components and a disclosure of what cannot be isolated.

Efficient differentiation and limits of pooling

Insurance uses heterogeneous risks because individuals or firms pool exposures across states of the world. Risk classification can mitigate cross-subsidies that would otherwise encourage lower-risk buyers to depart a voluntary pool; it can also reduce protection for higher-risk participants when their individualized premiums rise beyond their willingness or ability to pay. The competitive insurance equilibrium analysis of Rothschild and Stiglitz (1976) shows why the informational structure of the market matters for contract equilibria and potential nonexistence of equilibrium. Their canonical model does not predict that granular big-data pricing is always welfare improving; its economic conditions must be specified before extrapolating.

Crocker and Snow's categorical discrimination study (1986) establishes efficiency properties of costless imperfect categorization in their theoretical framework, while its evaluation of costly classification produces an ambiguous efficiency ranking. This qualification is especially important for algorithmic systems. Data acquisition, governance, technology, privacy loss, model maintenance, litigation, regulatory compliance, and consumer comprehension are real resource costs. An additional classifier that predicts a tiny amount of risk may have negative net value once these costs are included.

The strongest interpretation of this literature is conditional. If classification is accurate, legally admissible, economically inexpensive, and capable of reducing distortions in voluntary participation, it can improve the allocation of risk. If classification imposes large costs, extracts rents from inelastic customers, or prices essential coverage beyond household resources, the same improvement in predictive accuracy can coexist with reduced consumer welfare. The sign of the change cannot be read from the model's predictive score.

Surplus, transfers, and resource costs

Let V_i denote an individual's willingness to pay for a standardized insurance contract, including expected indemnity value and the individual's valuation of risk reduction. Let P_i be the price. The consumer's gross contractual surplus is V_i - P_i if they buy, abstracting from liquidity constraints and complicated utility effects; the insurer's expected contribution is P_i - C_i - A_i - K_i. Adding the two eliminates the premium transfer and yields V_i - C_i - A_i - K_i. This algebra distinguishes a redistribution of surplus from a change in the real value of trade. It does not imply that distribution is irrelevant: weighting the welfare of differently situated households, the solvency of insurers, and the public cost of uncompensated loss is a separate normative question.

A personalized markup might raise an insurer's revenue from a customer who would have bought anyway. The immediate effect is a transfer from consumer surplus to producer surplus, before considering dynamic competition or investment. If the higher price causes a buyer whose willingness to pay exceeded the social cost of insurance to exit, a lost mutually beneficial trade is an efficiency loss. If a reduced price induces a buyer with willingness to pay above marginal cost to enter, the extensive-margin gain may be positive. A price cut can also attract high-cost risks at a premium below sustainable cost, generating a need for cross-subsidization or a later capital adjustment. The welfare calculation must therefore include selection-sensitive marginal cost and the time dimension.

The welfare estimation approach of Einav, Finkelstein, and Cullen (2010) makes the price variation, demand, and cost relationship central to identifying losses from selection. Its case study detects adverse selection while finding relatively small welfare effects in its particular employer health-insurance application. That empirical finding belongs to the studied setting; it provides neither a universal welfare coefficient nor a claim about algorithmic pricing across all insurance lines.

Dynamic contracts and the unit of discrimination

Insurance renewal is a repeated transaction. The first-year premium can be low because the insurer anticipates future retention, intermediary commissions, or customer acquisition costs. A renewal premium can rise even if expected claims remain unchanged, and a rising customer tenure can be predictive of reluctance to switch. The unit of analysis is then a sequence of prices and coverages rather than one year's offer. Price walking involves a systematic tenure-related increase in margins where customer inertia is exploited. The United Kingdom's renewal pricing rules, in force from January 2022 for covered home and motor business, address the relation between an existing customer's renewal quote and the equivalent new business price; they do not impose a universal prohibition on all risk-differentiated premiums or on changes in insured cost.

Suppose an annual risk cost remains $600, standardized non-claim cost $100, and capital charge $50. An illustrative premium of $800 creates a $50 residual margin; $950 creates a $200 margin. If the same customer is quoted $950 at renewal solely because a retention algorithm predicts a high probability of acceptance, the $150 difference is a price-sensitive margin under these assumptions. If expected claims rise by $150, the comparison changes fundamentally. Without documented movement in these components, the label applied to the premium cannot establish its cause.

Algorithmic Risk Classification

From broad classes to high-dimensional estimates

Risk classification is an attempt to estimate conditional expected loss and group exposures of comparable economic significance. Historical rating systems used a manageable number of variables, territories, classifications, experience modifiers, and expert judgments. Modern models can integrate claims histories, geospatial signals, transactional traces, connected-device data, and complex interactions. The economic advance arises when predictors carry validated incremental information about expected claims or the insured's response to coverage, and when that information can be used within applicable legal boundaries.

The Actuarial Standard of Practice No. 12 (ASB, 2005/2006, with later revision metadata) concerns the design, review, and modification of risk-classification systems; its existence reflects a professional requirement for reasoned risk grouping independently of the modeling technique. A machine-learning label is still a risk class for practical governance if it determines premiums or access. The classification should be interrogated for relevance to cost, credibility, stability, and communication rather than accepted because its algorithm is complicated.

Let X_i be underwriting data available at quote time and Y_i the realized claim outcome. A cost model estimates μ(X_i) = E[Y_i | X_i], ideally with exposure adjustments and a defined contract. A retention model estimates r(P_i, Z_i) = Pr(renewal | price P_i, information Z_i, available alternatives), while an acquisition model estimates n(P_i,Z_i). The inputs can overlap, but the targets answer different questions. The first predicts cost; the others predict behavior. A complete pricing engine may combine them inside a constrained objective, so governance must separately document target variables and their causal interpretations.

Predictive accuracy is conditional

A model that predicts historical paid claims well may fail when coverage limits change, a new catastrophe regime emerges, repair inflation rises, medical utilization changes, or consumer behavior responds to monitoring incentives. Claims development, censoring, incomplete observation, reporting practices, and changes in legal liability can bias fitted losses. A claims outcome after insurance was purchased is not necessarily the loss distribution of an otherwise identical uninsured customer. Coverage itself changes utilization, protection, and observed costs; observational prediction is not automatically causal.

Calibration asks whether realized losses approximate predicted losses across pre-specified classes; discrimination asks whether a score ranks different risk levels. The economic importance of a model also depends on the stability and credibility of those comparisons, absolute monetary calibration, tail losses, robustness to shocks, and effects of deploying the model on who enters the pool. Group calibration can conceal localized harms, especially in small segments. A model can be statistically accurate and still employ prohibited characteristics or proxies. The formal fairness trade-off results of Kleinberg, Mullainathan, and Raghavan (2016) demonstrate incompatibilities among certain statistical fairness conditions in specified mathematical settings, while Chouldechova's analysis (2017) illustrates distinct error-rate and group-outcome considerations in criminal-risk prediction. These are methodological insights, not direct empirical findings on insurance premiums; translating them requires insurer-specific validation and legal analysis.

Proxy attributes and endogenous risk

An innocuous-looking geographic indicator may encode access to services, household wealth, histories of segregation, or other legally or ethically sensitive information. A device-based health indicator may reflect the possession of equipment and time for exercise as much as future morbidity. In data-intensive markets, a proxy may be selected because it predicts both claims and demand, thereby embedding the extraction of consumer surplus into a seemingly actuarial rating feature. The OECD's analysis of digital health tools (2024) identifies concerns about privacy, data quality, safety, effectiveness, and fairness in policyholder-facing tools.

Proxy evaluation cannot stop at the coefficient on a protected characteristic. Removing the characteristic may leave correlated variables that reproduce group differences; adding it only for fairness auditing can implicate data protection obligations. The relevant technical question is whether a feature contributes to the lawful target at an acceptable cost and what distributional effect arises once the full price and coverage system acts on the result. Validation therefore needs counterfactual testing of feature removal, causal scrutiny of collection practices, and documentation of disparate outcomes under realistic enrollment responses.

The two estimators inside an insurance price

A practical mechanism often begins by estimating a technical premium T_i = C_i + A_i + K_i. A second stage chooses a quoted price P_i that responds to predicted demand while respecting constraints. If the firm operates a differentiated pricing portfolio, it can solve a constrained optimization problem such as maximizing expected contribution Σ_i q_i(P_i)[P_i - T_i] subject to lawful rating plans, premium adequacy, portfolio risk constraints, fairness rules, and price-disclosure requirements. Here q_i(P_i) is a predicted probability of purchase, and it need not be a true causal response function unless identified with suitable evidence.

The distinction matters because modeling consumer behavior is not equivalent to measuring a consumer's risk. The NAIC's price optimization study expressly discusses elasticity of demand and the possibility of differing quotes for similarly situated insurance risks. Its account is a description of methods and regulatory choices, not a determination that any particular algorithm violates every state's law. The relevant regulator, insurance line, rating rule, and evidence of actual discriminatory effect remain decisive.

Model governance as economic infrastructure

Data provenance, data rights, error correction, model versioning, drift monitoring, vendor accountability, human review, and appeal channels affect the real cost and distribution of algorithmic pricing. If a consumer cannot learn why a price rose, correcting a bad input is expensive and difficult. If the insurer cannot reproduce its own quote after a vendor update, risk and compliance costs rise. Transparent technical documentation reduces these frictions and supplies evidence for supervisors. The NAIC AI model bulletin framework (adopted nationally as a model in 2023, with jurisdiction-specific adoption) and the IAIS AI supervision paper (2025) both associate insurance AI governance with accountability, robustness, transparency, fairness, and redress.

Risk classification is consequently a production process with economic costs and institutional constraints. The measurable output is a sequence of offers and expected claims, not a laboratory prediction score alone. A reliable assessment relates statistical performance to coverage take-up, market access, claims outcomes, distributional incidence, and insurer capital. This wider evaluation is particularly necessary when algorithms are optimized simultaneously for profitability and retention.

Heterogeneous Willingness to Pay

Preferences, constraints, and the meaning of demand

The willingness to pay for insurance is shaped by exposure, the distribution of possible losses, wealth, risk aversion, liquidity, trust, financial literacy, switching costs, the perceived reliability of claims payment, and the availability of alternatives. It is therefore a valuation of a complicated contingent contract under uncertainty. A household's unwillingness to purchase a premium at one moment does not establish the absence of an insurable interest or the absence of social value from coverage. Similarly, observing a renewal does not show that the quoted price was freely chosen in an environment of frictionless comparison.

A behavioral model should distinguish an economically rational valuation, a liquidity-constrained maximum affordable payment, and an observed acceptance decision. A person may value protection at $900 but be unable to pay more than $700 this month. Another may pay $1,100 for a contract that would be available for $800 after a costly search. A third may place substantial value on an insurer's claims reputation or a superior service network, so a higher price reflects a genuinely differentiated product. A machine-learning system trained to predict purchase or retention cannot identify these explanations from an acceptance label alone.

The broader economics of privacy literature synthesized by Acquisti, Taylor, and Wagman (2016) shows that information collection and disclosure alter bargaining positions and the informational asymmetry between firms and consumers. In insurance, the asymmetry is reciprocal: applicants may know more about their exposure or health, while firms may know more about the customer's projected renewal behavior and the distribution of competitors' quotes. The availability of personal information can change both the technical risk price and the residual bargaining margin.

Elasticity and first-order pricing logic

For a simplified product with constant marginal economic cost c and purchase probability q(p), an insurer choosing a single price p maximizes expected contribution π(p) = q(p)(p - c). At an interior differentiable optimum, q(p) + (p - c)q'(p) = 0. Let ε(p) = [p/q(p)]q'(p), the price elasticity of purchase, ordinarily negative in this representation. Then (p - c)/p = -1/ε(p), when the assumptions and interior solution hold. This classical markup relation is an algebraic implication of the specified objective, not an empirical statement about realized insurance rates. It requires a causal demand response, not simply a correlational customer score.

With customer-specific demand q_i(p), the corresponding optimizing margin may vary across individuals with different elasticities even when c_i is identical. Less elastic demand tends to support a higher monopoly-style markup under the same simplified conditions. A pricing algorithm can find such differences in purchase histories, renewal tenure, marketing responses, or clickstream behavior. Market power, price comparison technology, regulation, and rival reactions constrain the result. In a nearly frictionless competitive market with homogeneous coverage and transparent prices, a rival may attract the same customer and erode the markup. Consequently, the existence of advanced demand modeling is insufficient evidence of a sustained discriminatory premium.

Three practical complications prevent a straightforward inference from an optimized price to welfare. First, price may affect the composition and claims cost of enrolled buyers, so c is endogenous. Second, insurance purchase can change utilization, prevention, and reporting, so future claim cost depends on product design and behavior. Third, insurers are regulated entities subject to reserves, capital, rating restrictions, distribution rules, and market-conduct expectations. The pricing objective can be constrained by solvency and permissible pricing practices, and an internally calculated unconstrained optimum may never be offered to the public.

An illustrative two-customer transfer

Take two prospective policyholders with equivalent coverage, identical expected total economic cost of $750 each, and maximum willing payments of $900 and $820, respectively. Assume those values are stable, that consumers act on them, and that there are no competition, capital interactions, or privacy costs. At a uniform price of $800 both purchase. Their combined consumer surplus is ($900 - $800) + ($820 - $800) = $120; the insurer earns 2 × ($800 - $750) = $100. Total gains from these two trades are $220.

If the firm can set personalized prices of $890 and $810, both still purchase. Consumer surplus falls to $10 + $10 = $20; insurer contribution becomes ($890 - $750) + ($810 - $750) = $200. The sum remains $220, because trading quantities and real costs are unchanged. This is a redistribution of gains, not an increase in productive efficiency. The result assumes prices can vary without cost, customers cannot arbitrage, willingness to pay has been correctly inferred, and no competitive response occurs. It shows why a report of higher insurer revenue after algorithmic personalization cannot be equated with an improvement in aggregate welfare.

Now suppose both are charged $850, without personalized variation. Only the first purchases, and the gains from the second potentially beneficial contract disappear. Total gains become $900 - $750 = $150, or $70 less than the two-trade benchmark. Conversely, a personalized schedule charging $850 to the first and $800 to the second can restore both trades, creating total gains of $220 while changing the division of surplus. The broad economic effect of price discrimination can therefore be expansion or contraction of trade, not only extraction of willingness to pay. No conclusion about legal permission or distributional legitimacy follows from these stylized totals.

The example makes a further point: equal risk cost and unequal price can be observed while total trade expands. A legal prohibition on demand discrimination can still be justified by distributional, transparency, procedural, or social insurance considerations, but those are separate criteria requiring independent justification. Conversely, a legal allowance for cost-based pricing does not authorize every markup based on estimated inertia.

Demand identification and causal inference

Suppose an insurer trains a model on prior quotes and observes that people receiving expensive premiums renew more often. The correlation might arise because loyal customers accept higher prices, because higher-risk individuals face fewer alternatives, because coverage differs, or because the firm deliberately raised prices for people it already expected to retain. Estimating a price coefficient from such data without accounting for the assignment mechanism risks reverse causality and selection bias. An observational elasticity regression identifies a causal demand schedule only under strong exogeneity and sufficient overlap conditions.

A credible empirical design requires variation in price that is plausibly independent of unmeasured willingness to pay and claims risk, conditional on relevant controls. Randomized price tests can be informative if permitted by applicable insurance law and ethical safeguards, although intentionally assigning unequal premiums raises independent legal issues. Other possibilities include valid instruments, regulatory or tax discontinuities, pre-specified rating changes, and credible difference-in-differences designs with appropriate parallel-trend and spillover checks. When the exposure affects both customer demand and risk classification, the exclusion restriction becomes especially demanding. Identification should establish the price variation actually used, and distinguish offered quotes, bound contracts, and claims-bearing exposures.

A dataset must retain rejected quotes and uninsured observations where lawful and feasible, since the realized policyholder pool is selected by acceptance. It must record coverage, deductibles, channels, promotional discounts, taxes, duration, and renewal characteristics, with temporal alignment between price and information availability. External competitor quotes can help identify substitution but introduce matching and unobserved-quality problems. A price change attributed to an algorithm may have followed a simultaneous underwriting revision or a catastrophe-related repricing. Suitable methods can mitigate these concerns; no econometric estimator can eliminate them by labeling its output causal.

Consumer inertia and measured welfare

The health insurance inertia study by Handel (2013) provides empirical evidence that consumers may remain in plans despite available alternatives and that reducing inertia can, in the studied setting, worsen adverse selection enough to reduce aggregate welfare. Its result is particularly useful because it demonstrates interactions between individual decision quality and market equilibrium. It does not vindicate an insurer's exploitation of inertia or imply that obscuring prices is desirable. It indicates that an intervention has to be evaluated jointly with risk sorting, prices, and product take-up.

Online comparison can lower search costs and intensify competition. Brown and Goolsbee (2002), in their term life insurance study, estimated that expanding use of online comparison in the 1990s reduced term life prices by approximately 8–15 percent in their empirical setting. This evidence provides a counterweight to the assumption that algorithmic technology necessarily raises margins: better consumer information can also reduce them. Yet comparison platforms may collect consumer data, rank suppliers according to commercial arrangements, and create new intermediary bottlenecks. The realized effect depends on how ranking, coverage normalization, customer switching, and underwriting disclosures operate.

A research program should therefore estimate at least three linked quantities: the relationship between price and acceptance, the relationship between price and the marginal cost of the buyers attracted, and the consumer response to information and transaction costs. The combined structure is more informative than retention-prediction accuracy because the same observed renewal rate can arise under different welfare states.

Personalized Premiums and Selection

Selection as a response to the price schedule

An insurance portfolio is formed by the people who accept its prices and contracts. If unobserved risk is correlated with demand, changing prices changes the mix of losses insured. The central contribution of Einav, Finkelstein, and Levin (2010) is to connect empirical demand, costs, and selection rather than relying only on a binary test for asymmetric information. The key comparison is between average cost of existing insureds and marginal cost of additional insureds at the price that brings them into the market. A uniformly increasing or decreasing demand curve is not enough to establish the sign of selection.

Let a population have types i with expected costs c_i and price-dependent purchase probabilities q_i(p). Average insurer claims cost among those purchasing at price p is AC(p) = [Σ_i c_i q_i(p)]/[Σ_i q_i(p)], assuming the denominator is positive and standardized exposure. A common premium p is sustainable in this stylized market only if it covers the selected average claims, other costs, capital provisions, and any required return. When a price rise disproportionately causes lower-risk people to leave, AC(p) can increase, creating adverse selection. When higher-risk people are more price sensitive, a price rise could instead reduce average cost. The sign is empirical rather than guaranteed by the word insurance.

In an illustrative pool of 100 people, 50 have expected covered claims of $200 and 50 have expected covered claims of $600. The combined mean expected claim is $400. If a common gross premium of $450 induces 30 of the 50 lower-risk people to leave while all 50 higher-risk people retain coverage, the selected mean claim becomes (20 × $200 + 50 × $600)/70 = approximately $485.71. Even before expenses and capital, the $450 quote is inadequate for the remaining pool. If only 20 lower-risk people leave, the mean is (30 × $200 + 50 × $600)/80 = $450, again leaving no room for expenses. These calculations illustrate how a seemingly generous initial margin can disappear through changed composition. The numerical values are hypothetical and depend entirely on the posited acceptance pattern.

Greater prediction can reduce one distortion and create another

Imagine that new information allows the insurer to predict the two risk types reliably and offer different technical premiums. Under appropriate conditions, separate cost-based pricing can restore the participation of lower-cost customers by preventing a high cross-subsidy. For higher-risk customers, premiums may increase. The effect on total coverage depends on the value of risk transfer, affordability, existing public subsidies, mandatory purchase requirements, alternative insurers, and the composition of the social safety net. Accurate risk pricing can sustain a market without providing universal access to that market.

A highly predictive system can also reduce the scope of pooling across observed risks. In a purely individual actuarial price for every risk, ex ante cross-subsidization across observable expected losses is reduced; insurance against the realization of a loss remains valuable because a given person's outcome is still uncertain. Future classification updates, however, can make coverage against changes in risk status less secure. A healthy person whose newly observed risk becomes high may face a sharp renewal premium; long-duration guarantees, community rating, regulated renewability, and public risk adjustment can address this intertemporal exposure under their respective institutional conditions.

The relevant distinction is between insuring a realized event and insuring the future loss of favorable risk status. Short-term individually repriced coverage can perform the first function while leaving the second largely with the household. Contract duration and the capacity to commit to future coverage therefore influence the welfare implications of better classification. Comparing a one-year price algorithm with a guaranteed-renewal long-duration contract without accounting for these differences confuses two distinct insurance products.

Personalized margins interact with risk sorting

If a demand-sensitive algorithm raises premiums most for customers least likely to shop, the resulting selection need not mirror a purely risk-based price change. A higher markup on inert but low-risk customers could drive some unexpectedly price-sensitive low-risk households out of the pool, increasing average loss cost for the remaining business. A discount for price-sensitive high-risk consumers can attract disproportionately costly risks. Conversely, retention data may correlate with favorable claims behavior and give a firm reasons to retain lower-cost customers. Every scenario requires empirical estimation of how demand and risk interact.

Price optimization can be represented as a portfolio problem: choose the vector of offers P subject to an expected profit or risk-adjusted objective and regulatory restrictions, accounting for Pr(buy_i | P_i) and conditional insured cost E[C_i | buy_i, P_i]. Treating C_i as fixed after the quote may be wrong when buying behavior selects unmeasured risks. The actuarial cost and demand stages need to be reconciled rather than bolted together. A model that perfectly predicts acceptance but ignores which claim liabilities arrive with accepted policies can recommend prices that are superficially profitable and ultimately unstable.

Reinsurance and financial capital also affect the equilibrium. Insurers with correlated catastrophe exposures may charge portfolio loadings that differ from the individual expected claim, because a marginal contract adds tail risk rather than only a mean loss. Segment-specific margins may finance costly catastrophe reinsurance or regulatory capital. A scientific assessment cannot identify such loadings as pure extraction without examining credible capital and portfolio cost evidence. At the same time, invoking capital cost as an undifferentiated justification for all margins prevents meaningful analysis. The financial accounting should make capital attribution explicit and disclose its model dependence.

Evidence from insurance research and its limits

The demand and cost identification framework of Einav, Finkelstein, and Cullen (2010) illustrates how price variation can reveal selection and welfare in one employer's health plan choices. Handel's estimated inertia and selection effects (2013) show why policyholder responses affect equilibrium. Both settings differ institutionally from privately sold home, motor, and life insurance; the degree of choice, employer subsidies, plan menus, risk adjustment, and cost observability all matter. Moving numerical estimates from health plan choice to algorithmic automobile pricing would require an explicit transportability argument and comparable data.

A complementary empirical contrast is the online comparison evidence from term life insurance. Falling search costs were associated with lower prices, illustrating that lower information frictions on the demand side may offset the pricing power created by richer data on the supply side. It would be incorrect to claim that the study isolated an AI personalization mechanism in the 1990s. Its relevance is a demonstrated change in competition through consumer information within an insurance line.

The United States price optimization white paper identifies the use of price elasticity and alternative information in personal insurance pricing, while the FCA's insurance market evaluation supplies evidence on renewal behavior and distributional outcomes under a specific regulatory intervention. These sources cover mechanisms and policy effects, respectively; neither yields a global census of firms deploying individualized algorithms. A scientifically disciplined synthesis keeps theoretical possibility, documented institutional practice, and empirically estimated outcome distinct.

Counterfactual experiments and validation strategy

An assessment of personalization must compare policies that can actually be implemented. The relevant counterfactual could be a uniform tariff, an accepted broad class rating plan, individualized technical premiums with a common margin, demand-sensitive price optimization, or a renewal-price parity constraint. Each counterfactual must use an explicitly defined consumer choice environment and legal feasibility condition. The comparison should account for strategic rival responses and insurers' capacity to adjust product characteristics, acquisition spending, or exclusions in response to a pricing restriction.

A minimum empirical program includes: a validated pre-price expected-cost model; take-up probabilities by price and risk; uncertainty intervals for demand and loss parameters; a defined mapping from quotes to issued policies; administrative and acquisition costs; solvency or capital impacts; and sensitivity to alternative consumer welfare weights. Where feasible, rejected quotes and eventual uninsured outcomes should be measured to avoid a survivorship account of only successful policies. Price experiments should respect applicable insurance law and the ethical implications of intentionally changing access to protection.

A robust sensitivity analysis varies the correlation between risk and elasticity, the degree of market power, switching frictions, the availability of government protection, and model errors concentrated in smaller subgroups. A reported welfare improvement should survive plausible perturbations of those assumptions or be accompanied by a clear boundary on the claim. An algorithmic price schedule evaluated against an unrealistically immobile competitor or a fully informed representative consumer can create misleading certainty. The researcher's task is to show how conclusions move across credible market states rather than present a single optimized objective as universal economic proof.

Competition and Redistribution

Information asymmetry as a source of pricing power

A firm's ability to estimate willingness to pay does not by itself establish power to enforce a personalized price. Market power depends on entry, switching, the substitutability of covered risks and contractual terms, information available to rival suppliers, and the institutions through which insurance is distributed. A small insurer can deploy a sophisticated retention model and still face a competitor able to undercut its offer. A dominant distribution platform may possess modest underwriting technology yet control customer access and make price comparison difficult. The mechanism through which information becomes market power must be identified.

In a transparent market, a higher markup can lead to immediate switching, particularly for standardized annual products. In an opaque market, comparison requires effort: contracts have different exclusions, excesses, renewal terms, claims procedures, and service quality. The costs of evaluating these dimensions reduce the effective elasticity of some consumers. Price discrimination can consequently survive even when insurers appear numerous. Consumer knowledge and search technology shape market outcomes; Brown and Goolsbee's life insurance comparison evidence demonstrates a historical case of substantial price reduction associated with lower search costs.

Competition in algorithmic markets can occur in the design of features rather than only the quoted premium. A lower published price might coexist with higher deductibles or narrower protection, while a higher price might include more extensive claims services. Commission arrangements, affiliate platforms, advertising, and price-comparison rankings affect the channel a consumer sees. A market study should therefore compare the full effective price per standardized level of coverage, distribution costs, transparency, and claims performance. Premium dispersion conditioned only on demographic group is a weak stand-in for welfare when benefit variation is material.

The strategic effect of changing the renewal rule

Consider a firm that offers steep first-year discounts financed by larger expected renewal margins. If consumers switch infrequently, acquisition spending may be recouped through their inertia. A rule requiring renewal parity with the equivalent new-business price can change the initial discount strategy: firms may compete less aggressively for short-run new business while competition for the longer-term value of a customer becomes more important. These adjustments mean that the distributional benefit to long-tenure customers need not equal the full historical price gap; new business premiums can rise as renewal margins fall.

The United Kingdom gives an empirical case for this transition. The FCA's ICOBS renewal pricing provision states that, for in-scope business, the renewal price may not exceed the equivalent new-business price. The FCA's 2025 evaluation of the remedies found evidence of reduced price walking while distinguishing changes in customer risk and inflation from tenure-based margin extraction. The evaluation reports that the average home-insurance price gap between existing and new customers fell from £95.38 to £49.17; it cautions that residual average differences do not necessarily demonstrate widespread rule violations. For motor insurance, the evaluation reports a £111.14 increase in average new-business prices compared with a £22.71 increase for existing customers, in a setting of major inflationary pressure and differing customer risk profiles.

The same evaluation estimated an average price reduction of £6.63 per motor policy associated with the intervention in its causal analysis, with a central ten-year monetary estimate around £1.6 billion and a broad sensitivity range from approximately £163 million to £3.0 billion. It found no statistically significant causal effect on home-insurance prices within its empirical specification. These are regulator estimates with stated identification limitations and a specific reform package, not direct observations of every individual's welfare or proof that the entire difference was created by algorithms. The authors' estimated benefit is distinct from the FCA's earlier ex ante projection of £4.2 billion across home and motor markets over ten years in the policy's cost-benefit analysis. Ex ante prediction and ex post estimation should remain separately labeled.

The observed results establish a disciplined inference. Restricting one form of tenure-sensitive discrimination can shift prices across customer groups and change competition; the effects require measurement after firms adapt. The evidence supports concern about concealed price walking and demonstrates one intervention's outcomes. It cannot establish that every differentiated quote is exploitative, that all analogous regulatory restrictions would achieve the same effect, or that the regulation alone caused all nominal premium changes during an inflationary period.

Distribution between households and between risks

Personalized pricing redistributes along several dimensions. At a fixed portfolio quantity, a demand-sensitive markup transfers surplus from customers to suppliers. A risk-based classification changes the incidence of cross-subsidization between groups with different expected losses. A competition-driven reduction in search costs can transfer surplus from producers to consumers while lowering the resource cost of obtaining comparable offers. A rule affecting premiums can also change insurer acquisition costs, coverage participation, and expenses carried by public systems if households become uninsured.

These transfers have different implications for policy. A high-risk person who faces an expensive premium may have a strong expected need for the product and limited alternatives. A low-risk person may pay a higher pooled rate to sustain access for others. A consumer with limited financial literacy may face a large retention markup even though their expected loss is ordinary. Treating these situations as equivalent because all involve different prices would erase their economic mechanisms. Social evaluation must decide which differences warrant compensation, redistribution, or legal restrictions while preserving an insurer's capacity to meet promised claims.

Avraham, Logue, and Schwarcz (2014) document material variation among United States insurance anti-discrimination rules across jurisdictions and insurance lines. This variation demonstrates why the relevant legal standard cannot be inferred from economic language alone. A predictor's apparent actuarial power may be insufficient when the classification is prohibited. A legally permissible risk class may still create a distributional problem that calls for public rather than purely private insurance instruments. The distinction between economic efficiency, legality, and social acceptability is indispensable in an international analysis.

Stability, capital, and the insurance function

Competition can make premium margins thin, while risk-based segmentation may increase uncertainty for certain groups and reduce broad pooling. Neither result directly settles market stability. Insurers face liabilities whose magnitude and correlation can change suddenly after catastrophes, inflation, or legal decisions. Capital adequacy, reinsurance, loss reserves, underwriting discipline, and product exit choices determine the reliability of the promised indemnity. A rate that looks attractive to consumers may become unsustainable when correlated losses materialize. Likewise, a capital charge used to justify a premium difference should be traceable to portfolio risk and applicable regulation.

Algorithmic feedback creates additional concerns. If several insurers rely on similar external risk scores or price-comparison inputs, they may segment the same classes similarly and simultaneously reduce offers to certain risks. Shared data errors can propagate across firms. A historical model can learn from prices and acceptance decisions that previous models generated, reinforcing classification patterns. These are plausible systemic mechanisms requiring firm- and market-level evidence; the present dossier does not assert that such feedback has been measured for the global insurance sector. A credible investigation would combine rate filings, market entry and exit, quote distributions, vendor concentration, and claim outcomes over time.

Welfare measurement in competitive equilibrium

The most informative welfare comparison estimates realized consumer utility and the real cost of supplying coverage under alternative legally feasible price schedules, allowing insurer responses, selection, and product attributes to adjust. In a simplified partial-equilibrium setup, aggregate surplus can be expressed as the integral of willingness to pay net of marginal economic cost over contracts actually traded. When marginal cost depends on selection, a price change shifts both the number of policies and the cost of those policies. The Einav–Finkelstein–Cullen framework makes this interdependence central to empirical welfare estimation.

Additional social weighting can account for income, vulnerability, essential coverage, and distributional objectives. Such weighting must be declared; it is not mechanically dictated by a prediction model. A policy that increases unweighted aggregate surplus can disadvantage a particularly exposed subgroup. A redistribution that reduces measured surplus could be justified under an explicit social objective and legal mandate. Reporting both efficiency and incidence enables the reader to see the trade-offs rather than concealing normative priorities inside an opaque optimization score.

Regulation, Fairness and Welfare

The structure of the legal comparison

Insurance pricing regulation protects several interests that economic models consider separately: premium adequacy and claims-paying capacity, fair access and legally protected characteristics, market conduct, contractual clarity, privacy, and the integrity of competition. These interests vary by line and jurisdiction. A responsible international synthesis compares operative texts and their scope rather than collapsing all norms into one universal prohibition on algorithmic discrimination.

Three regulatory designs recur in the examined materials. The first supervises risk classification and the permissible factors used in premium setting. The second restricts a specific pricing practice, such as renewal price walking, with a directly testable relationship between two prices. The third governs the data, models, accountability, explanations, and remedies associated with algorithmic decisions. These instruments can coexist, and none automatically substitutes for the others. A technically well-governed model can still produce a prohibited premium; a lawfully priced contract can still expose the insurer to model governance risk; a ban on one markup practice can leave many forms of actuarial segmentation in place.

United States: state insurance law and model governance

In the United States, state insurance departments regulate rating and unfair discrimination under applicable state insurance laws. The NAIC's 2015 white paper documents competing arguments about price optimization in personal lines and notes that acceptable departures from indicated rates depend on state law and supervisory judgment. It is a task-force research paper, not a nationwide binding ban. The broader empirical-legal work of Avraham, Logue, and Schwarcz emphasizes variation across states, covered insurance lines, and restricted characteristics. Consequently, a proposition about an American insurer requires the actual state, policy type, and rating framework.

The NAIC adopted its AI Systems Model Bulletin in December 2023. As a model supervisory instrument, it describes governance expectations relating to insurer use of AI and relevant information that regulators may request; its particular legal or supervisory operation depends on adoption and existing law in a jurisdiction. Treating national NAIC materials as automatically enforceable federal statutes would materially misstate their authority. Model governance also extends beyond a single underwriting score to data suppliers, third-party systems, ongoing validation, and consumer-impacting decisions.

Colorado provides a concrete state-specific development. Its Division of Insurance announced the 2025 amended Regulation 10-1-1, effective October 15, 2025, extending the external-consumer-data governance framework to certain life, private passenger automobile, and health-benefit-plan insurers using external consumer data and related algorithms or predictive models in insurance practices. This is a Colorado regulatory action with an identified scope, not a representation of all U.S. jurisdictions. Its substantive effect is to connect the use of alternative data with governance, risk management, and unfair-discrimination scrutiny. The exact application to a particular insurer depends on the rule's defined entities, uses, and compliance obligations.

A regulator can require insurers to show what information affected an offer, how it relates to risk, whether external data can reproduce protected classifications, and whether model governance is appropriate. An adequate analysis should distinguish measured disproportionate outcomes from the legal test that defines unfair discrimination. Statistical disparity is often relevant evidence but is not a universal self-executing legal conclusion; permissible classifications and standards vary.

United Kingdom: renewal parity and observed outcomes

The United Kingdom has a particularly instructive, practice-specific rule. Under the current FCA Handbook renewal standard, a firm must not quote an in-scope renewing home or motor customer above the equivalent new-business price; the rules contain detailed assumptions for channels, cover, and related products. Implementation from January 2022 targeted tenure-based price walking, while complementary measures addressed automatic renewal, product governance, and reporting. The rule compares like with like at the defined renewal point. It does not make every premium identical across people, insurers, times, or risks.

The FCA evaluation paper EP25/2, published in July 2025, provides evidence that the intervention reduced price walking and changed prices for new and renewing consumers, with caveats concerning inflation, risk composition, consumer behavior, and causal identification. Its findings warrant separate attention because they measure changes after adoption rather than resting only on regulator predictions. A full policy evaluation remains multidimensional: lower observed switching among loyal customers could reflect reduced exploitation, but switching alone cannot establish whether product quality, choice, or competition improved. The FCA itself described product-quality results as mixed and did not find a statistically significant home-price effect.

Comparative interpretation matters here. Renewal parity can be evaluated by comparing counterfactual new-business and renewal quotes for the same risk and coverage. Other jurisdictions may address related harms through different legal instruments. Transplanting the rule into a market with substantially different mandatory insurance coverage, intermediary remuneration, insurer concentration, or subsidy arrangements would change the economic mechanisms. The United Kingdom case is evidence of a specific regulatory experiment, not a universal estimate of algorithmic discrimination worldwide.

European Union: protected classes, personal data, and AI systems

European Union insurance pricing operates within overlapping instruments, each with a particular scope. The Test-Achats judgment of the Court of Justice of the European Union (2011) invalidated the indefinite derogation permitting sex-based premium and benefit differences under Directive 2004/113/EC, with the invalidity taking effect December 21, 2012. Its significance is categorical: predictive actuarial relevance does not automatically override an equality rule. The decision should be interpreted within the legal subject of sex-based differences in the specified insurance context, not as a general ban on all classifications.

The General Data Protection Regulation establishes obligations for processing personal data, including the Article 22 regime for certain solely automated decisions with legal or similarly significant effects, subject to enumerated exceptions and safeguards. The assessment requires attention to whether a decision is solely automated, the nature of its effect, lawful bases and exceptions, and the rights and measures engaged. The Court's SCHUFA scoring judgment (Case C-634/21, 2023) concerns automated credit scoring and a third party's strong reliance on that score for contractual decisions; it informs the interpretation of Article 22 under its conditions but is not itself an empirical insurance price study.

The EU Artificial Intelligence Act (Regulation (EU) 2024/1689, as amended) places qualifying AI systems intended for risk assessment and pricing of natural persons in life and health insurance in an Annex III high-risk use category, subject to statutory criteria and applicable implementation rules. The Act's classification does not by itself classify every insurance algorithm or every motor-insurance price model as high-risk under that entry. The exact applicability, exemptions, and implementation timetable require consultation of the current consolidated legislation when deployment decisions are made; legislative amendment proposals are distinct from enacted law. The EIOPA 2025 AI governance opinion clarifies expectations under existing sectoral insurance obligations and expressly separates its scope from prohibited or high-risk AI Act systems, avoiding a false claim that the opinion creates additional statutory duties for those systems.

Data protection and insurance fairness overlap but answer different questions. A lawful ground for processing a feature does not prove that its use in rating complies with insurance equality law. A statistically fair score does not determine whether its collection was transparent, proportionate, or authorized. A prohibition on direct sex pricing can remain relevant when indirect proxies are used. Compliance analysis has to follow each applicable legal instrument independently before a unified assessment of the deployed pricing process is possible.

International supervisory guidance and proportionality

The IAIS's 2025 AI application paper addresses risk-based supervision, accountability, security, explainability, fairness, and redress under the existing Insurance Core Principles. Its fair-treatment application paper acknowledges risk-based pricing while examining access, consumer needs, communication, and treatment across the insurance life cycle. These are international supervisory materials intended to support interpretation and convergence, not self-executing rules with identical domestic force everywhere.

The OECD study of digital wellness insurance (2024) reinforces a related point: data-enabled prevention can potentially change claims costs and consumer behavior, while privacy, data quality, digital inclusion, and uncertain long-term efficacy complicate the evaluation. A premium discount for a monitored healthy behavior can act as an incentive, an underwriting signal, and a mechanism that shifts costs onto those unable to participate. Evaluating the program requires evidence of actual risk reduction, selection effects, and distributional incidence, not only engagement counts or device ownership.

Proportionate governance should track the magnitude and character of the risk. An algorithm that merely supports clerical administration differs from a system determining access to essential protection or a substantial premium increase. Supervisors can request traceability of inputs, model objectives, counterfactual price outcomes, third-party dependencies, and complaint redress. The NIST AI Risk Management Framework (2023) is a voluntary cross-sector reference for risk management; its useful governance vocabulary should not be mistaken for insurance-specific binding law.

A disciplined assessment of fairness and welfare

A rigorous assessment makes its evaluation dimensions explicit. Predictive fairness concerns evidence that the model estimates its lawful outcome across meaningful groups. Distributive fairness concerns the incidence of premium burdens, access to essential coverage, and cross-subsidization. Procedural fairness concerns notice, explanations, correction, contestation, and accountable human decision-making. Market fairness concerns the exploitation of search frictions, hidden price walking, and misleading comparisons. These dimensions can reinforce one another, but they need not rank a given policy identically. Theoretical conflicts among statistical fairness metrics, highlighted by Kleinberg, Mullainathan, and Raghavan and Chouldechova, further caution against treating a single metric as exhaustive.

An empirical regulatory audit should ask whether the proposed input predicts expected claims reliably; whether the result survives drift, small-group stress, and out-of-sample assessment; whether price changes reflect costs or consumer inertia; whether groups with comparable standardized risk and cover face substantially different margins; whether consumers can contest material errors; and whether the pricing policy changes aggregate coverage, market entry, claims-paying capacity, or social costs. The choice of fairness objectives and their legal authority must be recorded separately from technical performance. This produces a reviewable sequence of claims and evidence rather than an opaque final score.

The legal and economic conclusion is that policy should target the identified mechanism and measurable harm. If the harm is a stale risk model, recalibration and data governance may be appropriate. If it is unlawful use of a protected characteristic, the applicable prohibition and remedies govern. If it is a discriminatory retention margin, a conduct and price-comparison rule can be directly relevant. If it is unaffordable coverage for a high-risk population despite accurate underwriting, redistribution, risk adjustment, public reinsurance, or coverage guarantees may be more closely connected to the objective. Each instrument has implementation costs and possible equilibrium responses that must be examined before calling it optimal.

Conclusion

Algorithmic price discrimination in insurance markets is an economic phenomenon arising from the intersection of heterogeneous insured costs, heterogeneous willingness to pay, and unequal access to information and alternative contracts. Its significance cannot be recovered from predictive accuracy or premium dispersion alone. The primary analytical distinction is between a price variation attributable to credible claims and capital costs, a price variation attributable to customers' responsiveness or inertia, and a price variation attributable to differences in coverage. Those variations may coexist within one algorithmic tariff and affect risk pooling, market participation, consumer surplus, and insurer viability in different ways.

Established insurance theory explains how risk classification can moderate adverse selection under specified information conditions. Established industrial organization explains how personalized markups can redistribute surplus and change trade volumes when firms possess pricing power. Empirical insurance studies show that selection, inertia, and search costs can be measured, while the effects are tied to the studied population and institution. The FCA's 2025 evaluation demonstrates that a particular restriction on price walking can alter renewal and new-business pricing, with a measurable motor-market effect and an imprecise or unidentified home-price effect under the study's methods. This evidence strengthens the general conclusion that policy evaluation must occur after market adjustment, with explicit counterfactuals and uncertainty.

The monograph's original analytical synthesis identifies the research consequence of those findings: a credible account of algorithmic insurance pricing must jointly measure risk cost, demand response, contract quality, competitive alternatives, and the distribution of coverage. Estimating only premium levels, predicting only purchase probability, or auditing only model calibration leaves crucial economic quantities unidentified. Transparent hypothetical calculations demonstrate that consumer-to-insurer transfers can occur without changes in total trade gains, and that exclusion of otherwise mutually beneficial buyers can impose a separate efficiency loss. Selection-sensitive expected costs and capital needs further condition both results.

International regulation supplies multiple instruments rather than one universal formula. United States state rating law and model governance, the United Kingdom's renewal parity regime, the European Union's equality and data legislation, the AI Act's defined insurance applications, and IAIS supervisory guidance address different mechanisms. Their implementation depends on exact scope and authority. An evidence-based scientific conclusion recognizes the legitimate economic role of actuarial risk differentiation, the measurable harms of exploiting consumer inertia, and the importance of the distribution of protection. The lasting public question is which individuals can obtain credible coverage on sustainable terms, how information changes that access, and whether the resulting market offers society more valuable risk transfer at a defensible distribution of costs.

The analysis is a scholarly synthesis of verified sources and explicitly hypothetical calculations. It introduces no new independent theory, newly named model, or claimed discovery priority. It does not substitute for jurisdiction-specific legal advice, actuarial certification, or insurer-specific empirical validation. Future studies with linked insurer quotes, exposure-level claims, randomized or quasi-experimental variation, and audited platform data can test the conditional mechanisms identified here.

References