Chapter 35: Ethics, Bias, and Fairness in Insurance Underwriting

"The line between charging for risk and discriminating against people is not a wall. It is a contested border that every generation of underwriters has had to redraw — and the algorithm did not erase it, it only hid it inside the math." — constructed line, in the voice of this book's argument.

Overview

Here is the hardest sentence in this book, and you should sit with how uncomfortable it is: insurance works by discriminating. The whole apparatus you have spent thirty-four chapters learning — classification, loss runs, rating factors, experience modification, the model that scores a risk a 7 — exists to sort people and businesses into groups that pay different prices because they present different risks. A society that forbade insurers from discriminating between a careful driver and a reckless one, between a sprinklered warehouse and a tinderbox, would not have fairer insurance. It would have no insurance, because the law of large numbers (Chapter 1) and adverse selection (Chapter 1) would tear the pool apart. Risk-based discrimination is not a regrettable side effect of insurance. It is the mechanism.

And yet. The same word — discrimination — names the thing the law forbids and the thing civilization has spent a century trying to end: treating people worse because of their race, their religion, their national origin, their sex. Insurance must do the first kind. It must never do the second. The entire ethical weight of underwriting sits on the line between them — and that line is not bright, not fixed, and not where most people assume it is. A rating factor can be perfectly legal, perfectly predictive, and still be doing the forbidden thing by proxy. A model can contain no protected variable at all and still learn to charge Black neighborhoods more, because it learned from a history that already did. This is the chapter where the craft stops being only technical and becomes, unavoidably, moral.

You are not here to be lectured. You are here because an underwriter who cannot reason carefully about fairness will eventually make a decision that is either unlawful, indefensible, or both — and "the model said so" is not a defense to a regulator, a court, or your own conscience. So we will reason it through: what makes risk-based pricing legitimate, exactly where it tips into unfair discrimination, how a legal factor becomes an illegal proxy, how an algorithm launders historical bias into a clean-looking number, why geography is insurance's most haunted rating dimension, what the patchwork of regulation actually requires, and how to hold actuarial fairness and social fairness in the same hand the way a real underwriter must.

In this chapter, you will learn to:

• Explain the paradox that insurance must discriminate by risk to exist, and why that is lawful.
• Locate where risk-based pricing becomes unfair discrimination, using the legal and ethical tests.
• Define proxy discrimination and disparate impact and spot a neutral factor standing in for a protected one.
• Explain algorithmic bias — how a model encodes history even with no protected variable present — and name the controls that detect it.
• Trace redlining and its legacy, and explain why geography is so contested.
• Distinguish actuarial vs. social fairness, map the regulatory patchwork, and state the underwriter's duty.

Learning Paths

🏠 Personal Lines: This is your battleground. Auto and home are where credit-based insurance scores, territory, and proxy concerns are litigated and legislated (§35.2, §35.3, §35.5). Re-read Chapter 14's rating factors with this chapter's lens — every one of them is a fairness question waiting to happen. 🏢 Commercial Lines: The proxy problem is thinner but not absent; §35.7 and the Harbor Steel beat show why commercial risk-based pricing is usually defensible and why the protection-gap question still reaches a coastal business. Watch the line between rating a business and rating the people in it. 📊 Analytics: §35.4 is the core — proxy discrimination, disparate impact, and the fairness metrics (and their genuine incompatibilities) are now part of model governance. This is where Chapter 32's lift and Gini meet a constraint that has nothing to do with predictive power. 📜 Certification: §35.6 maps the regulatory landscape (unfair-trade-practices acts, FCRA, the NAIC's AI/Big-Data work, GINA) tested across AINS/CPCU ethics modules; the actuarial-vs-social-fairness framing in §35.7 is the conceptual core examiners return to.

35.1 The paradox: insurance must discriminate by risk

Start with the word, because the word is where the confusion lives. In ordinary speech, discrimination is a slur — it means treating someone unjustly because of who they are. In insurance, the word is a neutral technical term that means, simply, to distinguish. An insurer that sets one price for a teenage driver with two speeding tickets and a different price for a fifty-year-old with a clean thirty-year record is discriminating in the technical sense, and it is doing exactly what it is supposed to do. We met this in Chapter 1 under its proper name: it is fair discrimination — distinguishing among risks on the basis of their expected loss — and it is the foundation of the whole system. Chapter 4 gave you the legal frame: state law affirmatively permits fair discrimination and forbids only unfair discrimination, the term Chapter 4 owns.

Why must insurance discriminate? Because the alternative is not fairness; it is collapse. Imagine an insurer forced to charge every driver the identical premium regardless of record, vehicle, or use. The careful driver is now overpaying to subsidize the reckless one. Being rational, the careful driver buys less coverage or drops out; the reckless driver, getting a bargain, buys all he can. The pool fills with the worst risks, losses come in above the flat price, the price rises, the next-best risks leave — and you are watching the adverse-selection death spiral from Chapter 1 play out in real time. Risk classification is the cure for adverse selection, and a cure that worked by ignoring risk would be no cure at all. This is the paradox stated plainly: the thing that makes insurance fair to the pool — pricing each risk for what it is — is a form of discrimination, and removing it would make insurance unfair to everyone in a different way, by making it unaffordable or unavailable.

📋 At the Desk When a broker, a journalist, or a friend at a dinner party says "insurance discriminates," your job is not to deny it — that makes you sound evasive — but to split the word. "Yes, by design, insurance distinguishes between risks; that is what keeps it from collapsing. The question that matters is on what basis it distinguishes. Distinguish by how a person drives, and you are pricing risk. Distinguish by the color of their skin, and you are breaking the law and betraying the function. The whole art is in keeping the first and refusing the second — and noticing when the first is quietly doing the second." Practice saying that out loud. You will need it more often than you think.

The reason this is genuinely hard, and not just a matter of good intentions, is that the two kinds of discrimination are not always distinguishable by inspection. A factor can look like pure risk and function as pure prejudice. A model can contain nothing but neutral inputs and produce a racially skewed price. The rest of this chapter is the disciplined examination of exactly where, and how, the legitimate thing turns into the forbidden thing — because an underwriter who cannot tell them apart will, sooner or later, do harm while believing they are doing math.

It is worth being clear about why the law lands where it does, because the structure is not arbitrary. Insurance occupies a peculiar moral position: it is a private business that performs a public function (theme 6, which this chapter advances harder than any other). When you decline a risk or price it out of reach, you are not merely losing a sale — you may be denying a family the ability to rebuild after a fire, or a business the coverage a lender requires to lend. That is why society lets insurers discriminate by risk (the system would collapse otherwise) but draws hard lines it will not let them cross. The lines are not the regulator's whim; they encode a judgment that certain bases for sorting people — the ones tied to identity rather than conduct, to who you are rather than what you do — are intolerable in a market that allocates a near-necessity. Keep that frame in view. It explains why an at-fault accident (conduct) is a permissible factor and race (identity) is a categorically forbidden one even if both happened to predict loss: the law is sorting your behavior into the priceable column and your identity into the untouchable one, and most of the genuine difficulty in this chapter comes from factors that sit ambiguously between the two.

🔍 Check Your Understanding 1. A colleague says, "We should be proud — our rates don't discriminate at all." What is wrong with that statement as a description of how insurance works, and how would you correct it without conceding that the rates are unfairly discriminatory? 2. Why would a law requiring every policyholder to pay the same premium, regardless of risk, make insurance less available rather than more fair? Name the Chapter 1 concept at work.

35.2 Where risk-based pricing becomes unfair discrimination

If fair discrimination is pricing by risk, then unfair discrimination is pricing by something the law has decided may not be priced — or pricing by risk so crudely, or so disconnected from actual loss, that it fails one of the tests regulators and courts apply. Chapter 4 owns the legal definition of unfair discrimination; our job here is to make the line operational — to give you tests you can actually run on a rating factor. There are four that matter.

First, the protected-class test. Some characteristics are simply off-limits as rating factors, full stop, no matter how predictive they might be. Race, religion, and national origin are universally prohibited; sex, marital status, sexual orientation, and others are prohibited variably by state and line (§35.6). The prohibition is categorical — it does not yield to a strong correlation with loss. Even if you could prove that some protected characteristic predicted claims, you may not use it. The law has made a judgment that some forms of sorting are intolerable in a civilized market regardless of their statistical truth, and that judgment overrides the actuarial one.

Second, the actuarial-justification test. A permitted factor must still be supported by a genuine relationship to expected loss. You cannot charge red cars more because you dislike red cars; you may charge a vehicle symbol more because the data shows that model's repair and theft costs run higher (Chapter 14). Most states' unfair-trade-practices statutes forbid rates that are "unfairly discriminatory," and the standard interpretation is that price differences between insureds must reflect expected cost differences. A factor that sorts people into price tiers without a defensible loss rationale is unfairly discriminatory even if it touches no protected class — it is just arbitrary, and arbitrary sorting is forbidden too.

Third, the disparate-impact test. Here is where it gets subtle, and where the modern fights are. A factor can pass both tests above — it is not a protected class, and it does have a real loss relationship — and still produce a sharply unequal result across protected groups. We will define disparate impact properly in §35.4; the short version is a practice that is neutral on its face but falls much more heavily on a protected group in effect. Whether disparate impact alone makes an insurance rating factor unlawful is genuinely contested and varies by jurisdiction and line — some regulators treat a strong disparate impact as a red flag requiring justification, others require proof of intent. But ethically, and increasingly under emerging regulation (§35.6), a factor that is predictive and produces a large protected-group disparity demands scrutiny, not a shrug.

Fourth, the causation-versus-correlation test. The strongest defense of a rating factor is that it has a plausible causal connection to loss, not merely a statistical one. A driver's at-fault accident history causes higher expected loss in an intelligible way. A factor that predicts loss only because it correlates with something else — and that something else is a protected characteristic — is the proxy problem (§35.3), and the weaker the causal story, the more suspect the factor. Contrast two factors that might both show the same lift in a model. A building's fire-protection class (Chapter 9) predicts fire loss because better protection actually causes fewer and smaller fires — pull the thread and you reach a physical mechanism. The applicant's first name might also "predict" loss in a large dataset — but only because, in a segregated society, names carry information about race and ethnicity; pull that thread and you reach not a mechanism of loss but a demographic correlate. Both factors could have identical statistical power. One is a cause and one is a coincidence-with-an-identity, and the causal test is what tells them apart. This is why, when a factor's predictive power is strong but its causal story is missing or strained, the disciplined response is not to celebrate the lift but to get suspicious — strong prediction with no causal mechanism is the classic signature of a proxy.

⚖️ Compliance Corner The operative statutory language across most states comes from the NAIC's model Unfair Trade Practices Act, which prohibits "unfair discrimination between individuals of the same class and essentially the same hazard." Read that phrase carefully, because two clauses do all the work. "Same class and essentially the same hazard" means the prohibition is against treating like risks differently — it does not forbid treating genuinely different risks differently (that is the fair discrimination the system runs on). The fight is always over what counts as the "same class": the insurer argues two risks are in different classes because a predictive factor differs; the challenger argues the factor is a pretext and the risks are really alike. Note also what the statute does not on its own say: it does not, by its terms, prohibit disparate impact. That gap is exactly what the new generation of AI-fairness regulation (Colorado's SB21-169, the NAIC's model bulletin on AI; §35.6) is trying to fill.

The practical upshot for you at the desk: when you defend a price, you are implicitly asserting that the factors driving it pass these tests. Most of the time they do, and you never think about it. But the moment a factor is challenged — by a broker, a regulator, a plaintiff's lawyer — "the manual says so" is not an answer. You need to be able to say why the factor is in the manual: that it is not a protected class, that it has a loss justification, that its disparate impact (if any) has been examined, and that it is closer to a cause than to a coincidence. A factor that cannot survive those four questions does not belong in your price.

35.3 Proxy discrimination: when a legal factor stands in for an illegal one

Now we name the central danger of the whole subject, and the term this chapter owns. Proxy discrimination is the use of a facially neutral, legally permitted factor that functions as a stand-in for a prohibited characteristic — so that the prohibited sorting happens anyway, through the side door. The factor is legal. The variable is permitted. And yet the effect is that a protected group is systematically charged more, because the legal factor is so tightly correlated with the illegal one that using the first is nearly the same as using the second.

The mechanism is almost arithmetic. Suppose, hypothetically, that ZIP code is strongly correlated with race because of a century of residential segregation — which, in the United States, is not a hypothesis but a documented fact. Now suppose an insurer rates by ZIP code. The insurer has used no racial variable; race appears nowhere in the model. But because ZIP code carries racial information, pricing by ZIP code prices, in part, by race. The insurer can truthfully say "we do not use race" and still produce racially disparate prices. That is proxy discrimination: the laundering of a prohibited basis through a permitted one. It is the single most important concept in modern insurance fairness, because it explains how a scrupulously "colorblind" pricing system can still produce a discriminatory result.

The strength of a proxy is a matter of degree, and that degree is what you have to reason about. A factor that correlates only weakly with a protected class leaks only a little of it; a factor that correlates tightly with a protected class is, for pricing purposes, nearly the protected class itself wearing a different label. Think of it as a dial, not a switch. At one extreme, a factor with essentially no correlation to any protected characteristic (the fire-protection class of a steel plant, say) is no proxy at all — you can use it freely. At the other extreme, a factor so tightly bound to a protected class that knowing one tells you the other (a hypothetical variable that happened to be 95% determined by race) is a proxy so pure that using it is functionally indistinguishable from using the forbidden trait, and a court or regulator will treat it that way. Most contested factors live in the wide middle of that dial — predictive, neutral-looking, partially entangled with a protected class — and the whole disputed art of §35.6's emerging regulation is deciding how much entanglement is too much, and what evidence of independent causal value can justify a factor despite it. The underwriter's job is not to draw that legal line (that is above your desk) but to recognize where a factor falls on the dial and to escalate the ones that sit dangerously far along it rather than waving them through because "we don't use the protected class itself."

📄 Read the Submission

text FIGURE 35.1 — "The factor that knows too much" [constructed teaching example] THE SUBMISSION A pricing team proposes adding "average length of prior auto-policy tenure" as a rating factor; it is predictive of loss and uses no protected variable. THE CONTEXT In backtesting, the factor lowers the loss ratio. But length-of-tenure correlates with income and housing stability, which in this market correlate with race and age. WHAT IT SHOWS The factor is genuinely predictive and facially neutral — it would pass a naive "we use no protected class" review. WHAT IT DOESN'T It does not show the factor is causally tied to *driving* risk rather than to socioeconomic status; nor whether its benefit survives once a disparate-impact test is run. THE DECISION Don't adopt it on predictive power alone. Test it for protected-group disparity; demand a causal loss story; if the lift is really a proxy for income/race, the factor is suspect. THE LESSON "We don't use race" is not a defense. A neutral factor that carries protected information can discriminate by proxy — predictive power and fairness are different tests.

The example that has actually been fought over for decades is the credit-based insurance score — the factor Chapter 8 owns and Chapter 14 explores. Credit-based scores are strongly predictive of insurance loss; that is not seriously disputed by the industry, and multiple studies have found the correlation real. But credit history also correlates with race and income in the United States, for reasons rooted in historical and ongoing economic inequality. So the credit-based insurance score sits exactly on the proxy fault line: a factor that is predictive (passing the actuarial-justification test) and facially neutral (passing the protected-class test) but that produces a disparate impact on protected groups (failing, or at least straining, the third test). This is why credit scoring in insurance is permitted in most states, banned or restricted in a few (California, Massachusetts, and others limit or prohibit it for some lines), and perpetually contested everywhere. Reasonable, informed people disagree about whether a predictive, neutral factor with a disparate impact should be allowed — and that disagreement is not going to be resolved by more data, because it is not, at bottom, an empirical question. It is a values question wearing empirical clothes.

⚠️ Underwriting Trap The trap is believing that removing the protected variable solves the problem. It does not. An insurer that deletes race from its data has not become unable to discriminate by race; it has only become unable to see that it is doing so. If the remaining variables carry racial information — and in a segregated, unequal society, many of them do — the model will reconstruct the protected characteristic from its proxies and price accordingly, while everyone involved can honestly swear that race was never used. The disciplined practice is the opposite of colorblindness: you must measure the protected-group impact of your factors precisely so that you can detect proxy effects you would otherwise be blind to. You cannot fix what you refuse to look at.

There is a genuine tension here that the honest underwriter must hold rather than resolve glibly. On one side: if a factor truly predicts loss, banning it forces the good risks within a disadvantaged group to subsidize the bad ones, and pushes the insurer toward adverse selection — the careful low-income driver overpays so that the system can avoid a disparate impact. On the other side: a "predictive" factor that merely encodes the accumulated disadvantage of a protected group is not measuring risk so much as punishing history, and a society may legitimately decide that some predictive accuracy is a price worth paying to refuse that. The book's position, which is the regulator's growing position, is not that proxies must always be banned or always allowed, but that they must always be examined — that the burden falls on the insurer to show a factor's benefit is real, causal where possible, and not simply a relabeling of a prohibited basis.

35.4 Algorithmic bias: how models perpetuate history

Everything in §35.3 gets worse — much worse — when the pricing is done by a machine-learning model rather than a transparent rating table. This is the term the chapter owns: algorithmic bias is the tendency of a predictive model to produce systematically unfair outcomes for a protected group, arising not from a programmer's intent but from the data the model learned on and the way it learned. The model does not want to discriminate. It cannot want anything. It simply does what it was built to do — find patterns that predict the target — and if the patterns in the historical data reflect a biased world, the model will reproduce the bias faithfully, and call it accuracy.

There are several distinct ways this happens, and conflating them is itself a mistake.

Biased training data. A model learns from history. If the history embeds discrimination — if, say, claims were adjusted more harshly in some neighborhoods, or certain groups were historically steered into worse policies — the model treats that discriminatory pattern as ground truth and perpetuates it. The model is not biased relative to the data; it is faithful to data that is itself biased. Garbage in, gospel out.

Proxy variables (the §35.3 problem, automated and amplified). A machine-learning model is far better than a human or a linear model at reconstructing a hidden variable from its correlates. Remove race, and a gradient boosting machine (Chapter 32) will happily learn race from the interaction of ZIP code, first name, shopping behavior, and a hundred other features — not because anyone told it to, but because race carries predictive signal and the model is built to extract every drop of predictive signal it can find. The very power that makes machine learning valuable (Chapter 32's lift and Gini) is the power that makes its proxy discrimination harder to see and harder to stop.

Feedback loops. This is the most insidious mode. Suppose a model charges a neighborhood more, so fewer people there buy coverage, so the data on that neighborhood thins out and skews toward the few who did buy — who, having paid a high price, may be a selected subset. The model retrains on this distorted data, draws an even more extreme conclusion, prices even higher, and the distortion compounds. The model's own decisions poison the data it learns from next, so a small initial bias can ratchet into a large one over successive retrainings. (Chapter 32 raised this; here we name its fairness consequence.)

🤖 Model vs. Judgment The model that scored Harbor Steel a 7 (Chapter 32) is the same kind of model we are now interrogating for bias — and the relationship between override and fairness is more pointed than it first appears. Chapter 32's override was about accuracy: the model could not see that the fires predated a new plant manager. A fairness override is different and harder, because the model may be perfectly accurate as a predictor and still be producing an outcome you should refuse — accurate because it faithfully learned a biased world. Here is the asymmetry that should keep you honest: when the model flags a risk high for reasons you suspect are proxy effects, you can override down and you are both fairer and, if you are right about the proxy, more accurate. But when the model's bias runs the other way — when it under-prices a favored group and over-prices a disfavored one — the individual underwriter rarely sees it, because it is a pattern across thousands of files, invisible in any single one. That is why fairness cannot be left to the heroic individual override. It has to be engineered into the model's governance, because the bias that matters most is exactly the bias no single desk can see.

So what do you actually do about it? You cannot eliminate algorithmic bias by good intentions, and you cannot detect it by looking at one file. You detect and control it with disparate-impact testing built into model governance. Let us define the term the chapter owns. Disparate impact is a discriminatory effect on a protected group produced by a facially neutral practice, regardless of intent — measured by comparing outcomes across groups, not by reading the practice's stated rules. To test for it, you do something that feels paradoxical: you bring the protected attribute back in — not as a pricing input, which remains forbidden, but as an audit variable — and you measure whether the model's prices, declines, or scores differ across protected groups after controlling for legitimate risk. There are several formal fairness metrics, and here is the part that trips up everyone who expects a clean answer:

THREE FAIRNESS METRICS THAT CANNOT ALL HOLD AT ONCE        [constructed teaching example]

  DEMOGRAPHIC PARITY   the model's average price (or approval rate) is equal across groups
                       → "treat the groups the same on average"
  EQUALIZED ODDS       the model is equally accurate (same error rates) for each group
                       → "be equally right and equally wrong for everyone"
  CALIBRATION          a given score means the same expected loss regardless of group
                       → "a 7 is a 7 whoever you are"

  ── THE IMPOSSIBILITY ──────────────────────────────────────────────────────────────────
  When the underlying loss rates genuinely differ across groups, these three definitions
  are MUTUALLY INCOMPATIBLE — you can satisfy at most a subset. A model calibrated so that
  "a 7 means the same risk for everyone" will NOT, in general, produce equal average prices
  across groups. Choosing which fairness to enforce is therefore a VALUES decision, not a
  technical one — and pretending it is technical is itself a way of hiding the choice.

This is the single most important technical fact in insurance fairness, and it is why the field has no tidy answer. The mathematics is settled — it is a theorem that these definitions conflict when base rates differ — so the disagreement is not about whether they conflict but about which fairness an insurer should honor. An actuary who insists on calibration (a 7 is a 7 for everyone) is choosing a defensible notion of fairness; so is a regulator who insists on demographic parity (the groups should not be charged differently on average); and they cannot both have their way when the loss rates differ. There is no escaping the choice into pure objectivity, and the underwriter who understands this is far better equipped than the one who thinks "fair" has a single meaning the model can be told to compute.

📋 At the Desk Suppose you sit on the model-governance committee — increasingly part of a senior underwriter's job (Chapter 38). Here is the disparate-impact check you would actually run, in spirit: ```python

Audit a model for disparate impact across a protected group.

The protected attribute is used ONLY to AUDIT outcomes, never to PRICE — that distinction is the

whole legal and ethical point. Illustrative data; do not execute at build time.

import pandas as pd

def disparate_impact_report(df, score_col, group_col, loss_col): out = [] for g, sub in df.groupby(group_col): out.append({ "group": g, "avg_score": sub[score_col].mean(), # demographic parity view "decline_rate": (sub[score_col] >= 7).mean(), # approval/decline disparity "actual_loss": sub[loss_col].mean(), # is the score CALIBRATED for this group? "loss_per_score": sub[loss_col].mean() / sub[score_col].mean(), }) rep = pd.DataFrame(out) # If decline_rate differs sharply across groups but loss_per_score is ~equal, the model is # calibrated yet produces a disparate impact — the §35.4 dilemma, made visible. return rep `` Read the output like an underwriter, not a statistician: if the decline rate is far higher for one group butloss_per_score` is roughly equal across groups, the model is calibrated and yet disparate — and now you, not the model, must decide what to do about a result that is simultaneously "fair" by one definition and "unfair" by another. That decision is a governance act, and it must be documented.

35.5 Redlining and its legacy

No discussion of insurance fairness is honest without confronting redlining — the term the chapter owns, and the historical wound that explains why geography is insurance's most haunted rating dimension. Redlining is the practice of denying or pricing up financial services — historically mortgages and insurance — for entire neighborhoods based on their racial or ethnic composition, marked literally with red lines on maps. The name comes from the color-coded residential security maps drawn in the 1930s by the federal Home Owners' Loan Corporation, which graded neighborhoods for lending risk and colored the predominantly Black and immigrant ones red, deeming them "hazardous." Those maps shaped where capital and coverage flowed for generations.

The insurance dimension of redlining is sometimes forgotten behind the mortgage story, but it was just as real. Property insurance was historically denied, or offered only at punitive prices and on inferior terms, in redlined neighborhoods — and without property insurance you cannot get a mortgage, cannot rebuild after a fire, cannot accumulate or protect wealth through a home. Insurance redlining was thus a mechanism by which disinvestment compounded: the neighborhood deemed risky was denied the coverage that might have let it recover, which made it riskier, which justified the next denial. It is the §35.4 feedback loop, but run over decades and across a whole society, with human beings making the decisions that the algorithm now makes faster.

⚖️ Compliance Corner Explicit racial redlining is illegal — prohibited by fair-housing and civil-rights law and by every state's unfair-discrimination statutes. The hard modern question is not whether you may draw a red line around a Black neighborhood (you may not, unambiguously) but whether a risk-based geographic factor that happens to track those same historical lines is lawful and ethical. Geography genuinely matters to insurance risk: a coastal ZIP code really does have more hurricane exposure (Chapter 30); an urban ZIP code really may have higher theft or fire-loss frequency. So territory (the rating territory Chapter 14 owns) is a legitimate factor. But because territory in a segregated country is correlated with race, it sits squarely in the proxy zone of §35.3 — legitimate on its face, suspect in its effect. Some states regulate the granularity of territory, prohibit certain geographic factors, or require that territorial rates be justified by loss data at a defined level. The line you must hold: price the peril that genuinely attaches to a place; never let "territory" become a sanitized label for the population that lives there.

The legacy is not history; it is present tense. Studies and investigations over the years have repeatedly found that, controlling for risk, coverage in predominantly minority neighborhoods can be more expensive, harder to obtain, or offered on worse terms — though, in keeping with this book's discipline, we will not attach a fabricated figure to that pattern; the qualitative finding is well attested and the precise magnitudes vary by study, place, and line, and should be read from current sources. The point for the underwriter is structural: even with no one in your company harboring a discriminatory intent, a pricing system that leans heavily on geography in a segregated society will reproduce the geography of historical discrimination unless it is deliberately examined and corrected. Redlining teaches the chapter's hardest lesson — that you can perpetuate a historical injustice with no intent to do so, simply by faithfully pricing a world that injustice already shaped.

🔍 Check Your Understanding 1. Explain how insurance redlining created a feedback loop that made a disadvantaged neighborhood appear riskier over time. Which §35.4 concept does this illustrate at the scale of a whole society? 2. A coastal ZIP code and an inner-city ZIP code both carry higher-than-average property rates. One is far easier to defend than the other against a proxy-discrimination challenge. Which, and why? (Think about the causal loss story.)

35.6 Protected classes and the patchwork of regulation

You cannot reason about fairness in the abstract; you have to know what the law actually forbids, and the honest answer is that it depends — on the state, the line of business, and the year, because the rules are changing fast. Insurance is regulated primarily by the states (the McCarran-Ferguson Act, Chapter 4), which means there is no single national rulebook on rating fairness but a patchwork of fifty-plus regimes, overlaid with a few federal statutes and a growing layer of model regulation. Here is the landscape an underwriter actually has to navigate.

The universally prohibited classes. Race, color, religion, and national origin are off-limits as rating or underwriting factors everywhere, for every line — there is no actuarial exception, no matter the correlation. This is the categorical core of §35.2's first test.

The variably prohibited classes. Sex/gender, marital status, age, sexual orientation, disability, genetic information, and credit history are restricted differently across states and lines. Sex-based auto rating is prohibited in some states and permitted in others; some states have moved to ban or limit it; gender-based life and annuity pricing remains common (mortality genuinely differs by sex) while gender-based health rating was largely eliminated by the Affordable Care Act (ACA) (Chapter 18). Credit-based insurance scoring (§35.3) is permitted in most states, restricted or banned in a few. The lesson: you cannot assume a factor's legality from one jurisdiction — the same rating variable can be required in one state, optional in a second, and illegal in a third.

It helps to see the landscape at a glance, even though every cell carries an asterisk that says "check the current rule in this state and line." The point of the table is not to memorize it — it will be out of date somewhere by the time you read it — but to internalize its shape: a small categorical core that never moves, and a large contested middle that moves constantly.

Genetic information. A federal statute draws a bright line here. The Genetic Information Nondiscrimination Act (GINA) prohibits the use of genetic information in health insurance (and in employment). But — and this is the gap that matters for §35.7 and for the David Okafor thread — GINA's protection does not extend to life, disability, or long-term-care insurance. A life insurer may, in most states, ask about and use genetic test results. This "GINA gap" is one of the most ethically fraught open questions in underwriting, and we return to it in the Underwriting File.

The new layer: AI and Big-Data regulation. This is where the patchwork is being actively rebuilt, and where a current underwriter must keep reading. Colorado's SB21-169 (enacted 2021) directs insurers to test their algorithms, external data, and predictive models for unfairly discriminatory outcomes against protected classes — importantly, focusing on effect, not just intent, and thereby reaching the disparate-impact gap that the old unfair-trade-practices acts left open (§35.2). The NAIC has issued a model bulletin on the use of artificial intelligence by insurers, setting governance expectations, and has a standing body of work on accelerated underwriting, big data, and algorithmic accountability. These are real, named, current developments — and because they are evolving, the specific obligations they impose should be read from the current statutes and bulletins, not from any textbook's snapshot.

⚖️ Compliance Corner The structural shift you are living through is the move from intent to effect as the test of unfair discrimination. The old model (the Unfair Trade Practices Act, §35.2) implicitly asked: did you treat like risks differently, and did you mean to discriminate? The new model (Colorado SB21-169, the NAIC AI bulletin) asks: whatever you intended, does your system produce a discriminatory outcome — and have you tested for it? For the underwriter and the model-governance committee, this is a profound change in burden: it is no longer enough to refrain from using a protected class; you may be required to affirmatively demonstrate that your models and data do not produce a prohibited disparate impact. "We didn't intend to discriminate" is becoming an insufficient defense. Document your testing as if you will have to prove it, because increasingly you will.

The practical reality is that a national or multi-state insurer maintains a matrix — by state, by line, by factor — of what may and may not be used, and updates it constantly. As an underwriter you are not expected to memorize fifty regimes, but you are expected to (a) know that the rules differ and never assume, (b) recognize when a factor or model output raises a fairness question that needs escalation, and (c) treat compliance as a floor, not a ceiling — the law tells you the minimum, and §35.7 is about the rest.

35.7 Actuarial fairness vs. social fairness (and the underwriter's duty)

We arrive at the deepest concept in the chapter and the one it owns: the tension between actuarial vs. social fairness. These are two genuinely different, genuinely defensible ideas of what "fair" means, and the reason insurance fairness is hard is that they often point in opposite directions and cannot both be fully satisfied. Hold both definitions in your head at once:

Actuarial fairness says: a price is fair when it accurately reflects the expected cost of the risk. Two risks that are genuinely different should pay different prices; charging them the same would force the good risk to subsidize the bad one, which is itself a kind of unfairness — and would invite adverse selection. On this view, a rating factor is fair precisely to the extent that it tracks real loss. Actuarial fairness is the fairness of the individual risk: you pay for what you bring.

Social fairness says: a price is fair when it does not deepen existing inequality or deny essential protection to those who need it — even if that means some risks pay less than their actuarial cost and others pay more. On this view, a factor that accurately prices risk can still be unfair if it makes coverage unaffordable for the already-disadvantaged, or if it prices people for circumstances beyond their control, or if it perpetuates historical injustice (§35.5). Social fairness is the fairness of the whole society: some risks should be protected even at a subsidized price, because protection is a precondition of participation in economic life.

These collide constantly. Charging a flood-zone homeowner the full actuarial cost of their hurricane exposure (Chapter 30) is actuarially fair — they bring that risk — and may be socially catastrophic if it renders an entire coastal working-class community uninsurable, which is exactly the protection gap (Chapter 30 owns the term) now opening across coastal and wildfire regions. Charging a low-income driver more because a credit-based score predicts loss is actuarially defensible and socially questionable. There is no formula that dissolves the tension, because it is not a math problem; it is a conflict between two values, both real.

ACTUARIAL FAIRNESS  ◄──────────── the contested middle ────────────►  SOCIAL FAIRNESS
 "price the risk accurately"                                      "protect access; don't deepen inequality"

  pure individual risk pricing                                    community rating / subsidy
  credit score allowed                                            credit score banned
  full coastal cat rate                                           rate caps, residual markets, public programs
  every factor that predicts loss                                 only factors within a person's control
        │                                                                    │
        │  adverse selection if you ignore risk  ◄───── trade-off ─────►  exclusion if you ignore access
        ▼                                                                    ▼
  the pool stays sound but some are priced out         everyone is covered but the pool may need a backstop

  The market, the regulator, and the underwriter all sit somewhere on this line — never at a pure pole.

The mechanisms society uses to push toward social fairness are worth naming because you will work alongside all of them: community rating (Chapter 18 — charging a class one price regardless of individual risk, as the ACA did for health); residual markets and FAIR plans (the insurer-of-last-resort pools that cover risks the voluntary market won't, as in coastal and wildfire states); rate caps and rate suppression (regulators holding prices below the actuarial indication, which trades access for the long-run risk of insurer withdrawal); and public programs (the NFIP for flood, Chapter 15; crop insurance, Chapter 26). Each is a deliberate departure from actuarial fairness in the name of social fairness — and each carries the actuarial cost of that departure: subsidy has to come from somewhere, suppressed rates drive carriers out, and a backstop that under-prices risk accumulates a liability the public eventually pays. There is no free lunch on either side of this line, and an underwriter who understands the trade-off is more honest than an advocate for either pole.

⚠️ Underwriting Trap The trap on this topic comes in two opposite flavors, and you must avoid both. The first is the technocrat's trap: "I just price the risk; fairness is the regulator's problem, not mine." This is false, because your pricing decisions are where actuarial and social fairness collide, and refusing to see the social dimension does not make it disappear — it just means you stop noticing the harm you may be doing. The second is the advocate's trap: "all risk-based pricing is unjust; insurance should just be fair." This is equally false, because pricing that ignores risk destroys the pool through adverse selection and leaves everyone worse off, including the people it meant to help. The disciplined position is uncomfortable and correct: price by risk because the system requires it, examine your factors for proxy and disparate-impact effects because justice requires it, support the social mechanisms that handle what pricing cannot, and never pretend either the actuarial or the social claim can be ignored.

So what, concretely, is the underwriter's duty? It is not to single-handedly solve a tension that societies have not solved. It is fourfold, and it is enough to keep you busy and honest. First, competence: understand your factors and models well enough to know what they are doing — including the proxy and disparate-impact effects (§35.3, §35.4) that "I don't use protected classes" conceals. Second, candor: be able to defend every price you set with a real loss rationale, and escalate, don't bury, a factor or output that you suspect is doing the forbidden thing by proxy. Third, compliance-plus: meet the legal floor of §35.6 and treat the disparate-impact testing now being required not as a box to check but as a genuine audit. Fourth, humility about the limits of your seat: recognize that the biggest fairness questions — community rating, residual markets, the protection gap — are decided above your desk, and that your job is to contribute honest analysis to those decisions, not to smuggle your own answer into a rate filing. An underwriter who does these four things is doing the ethical work the job actually demands — which is not to be a saint, but to be clear-eyed about a genuinely hard thing, and unwilling to let the math hide the choice.

This advances the book's sixth theme — insurance serves a social function — more directly than any other chapter: behind every price is a person who needs protection, and the underwriter decides who gets it and at what cost. And it advances the first theme — underwriting is judgment — at its most demanding, because the fairness judgment is the one the model most wants to make for you and is least equipped to make: it can optimize a loss ratio, but it cannot decide which fairness a society should honor when the definitions conflict. That decision is, irreducibly, yours and your institution's.

⚖️ Compliance Corner One more named practice belongs here because it crystallizes the whole tension: price optimization — the term this chapter owns. Price optimization is the practice of setting an individual's premium based not only on their expected loss but on their predicted price sensitivity — how much they will tolerate before shopping, how likely they are to renew without comparing. Charge the customer who won't shop a little more, the price-sensitive one a little less; the price now reflects willingness to pay, not just risk. Many regulators have concluded this is unfairly discriminatory — because it breaks the §35.2 principle that price differences between insureds must reflect cost differences, not the insurer's ability to extract more from a captive customer — and a number of states have issued bulletins prohibiting or restricting it. Price optimization is the clearest case in the book of a practice that is profitable, data-driven, and legal-sounding yet severs price from risk — exactly the line §35.2 says you may not cross. It is worth remembering as the limiting case: when the algorithm starts pricing what a person will tolerate rather than what they cost, it has stopped underwriting and started extracting, and the law has noticed.

🗂️ The Underwriting File

Is the Harbor Steel pricing fair? You have, by now, built most of a quote: a debit-rated property rate reflecting the aging roof and the two fires (Chapter 11), a debit workers'-comp X-mod (Chapter 22), a named-windstorm deductible (Chapter 12), telematics on the fleet (Chapter 23). A model scored the account a 7 and you overrode to a 6 (Chapter 32). Now apply this chapter's lens and ask the fairness question directly — and notice, with some relief, that commercial underwriting like Harbor Steel's is where the fairness problem is at its mildest. Every factor driving Harbor Steel's price has a clean causal loss story: the roof's age causes wind/water exposure; the hot-work operations cause fire frequency; the fleet's radius and driver records cause auto risk; the coastal location causes hurricane exposure (Chapter 30). You are pricing the peril attached to a business's operations and assets — there is no protected class, and no plausible proxy for one, in "thirty-year-old built-up roof" or "welding governing class." The protected-class test (§35.2) is not even close; the actuarial-justification test is easily met; the causal story is strong. Conclusion: the Harbor Steel pricing is defensible on fairness grounds — it is risk-based, not a protected-class proxy. Document that you considered it; do not belabor it.

But the chapter's harder lesson reaches even this file, from a different direction. Port Hadley is a working-class Gulf Coast town, and the same catastrophe exposure that makes you load Harbor Steel's rate is opening a protection gap (Chapter 30) across its whole community — homes and small businesses that the voluntary market is pricing out or non-renewing, exactly as Harbor Steel's prior carrier did. Your individual decision on Harbor Steel is actuarially fair. But the aggregate of every carrier making the same actuarially-fair decision is a coastal town that may become uninsurable — a social-fairness problem no single underwriter created and none can solve at the desk (§35.7). Note this honestly in the file: the Harbor Steel price is fair; the protection-gap question it sits inside is real, it belongs above your desk (the residual-market and public-program mechanisms of §35.7), and pretending the actuarial answer settles the social one would be the §35.7 technocrat's trap.

And a thread from outside the commercial file, because the chapter promised it. Recall David Okafor (Chapter 6, 17) — the borderline life applicant, standard-versus-preferred, with a family history of heart disease. Suppose Okafor's file now includes a genetic test result. Under §35.6, the GINA gap means a life insurer may, in most states, use that result — GINA protects only health insurance. So a person who did the responsible thing and got tested could be rated up, or declined, on a genetic predisposition they cannot change and did not choose. Is that actuarially fair? Arguably yes — it predicts mortality. Is it socially fair? Many would say no, and the policy debate is live. Okafor's file is where the chapter's abstractions become a single human being, and where you feel, in your gut, why "it predicts loss" is not the end of the fairness conversation. Running disposition: Harbor Steel pricing defensible (risk-based, no protected-class proxy); the coastal protection-gap noted as a social-fairness matter above the desk; the GINA-gap thread flagged on the David Okafor life contrast.

Conclusion

Insurance is built on discrimination in the technical sense — it must sort risks by their expected loss, or the pool collapses through adverse selection — and that fact is not a scandal but the mechanism. The scandal is when the legitimate sorting becomes the forbidden kind: when a price turns on a protected class, or on a factor with no real loss rationale, or on a neutral variable that functions as a proxy for a prohibited one. We gave you four operational tests for that line — protected-class, actuarial-justification, disparate-impact, and causation-versus-correlation — and we showed how the proxy problem makes "we don't use race" an empty defense, because a factor can carry protected information without naming it.

Algorithms make all of this harder, not easier: a machine-learning model learns proxies better than any human, perpetuates whatever bias lives in its training data, and can ratchet a small bias into a large one through feedback loops — all while producing a number that looks objective. The only defense is to measure what you would rather not look at: to bring the protected attribute back as an audit variable and test for disparate impact, knowing that the fairness metrics themselves conflict and that choosing among them is a values decision the math cannot make. Redlining showed the stakes at the scale of a whole society — that you can perpetuate an injustice with no intent to, simply by faithfully pricing a world that injustice shaped. And the deepest tension, actuarial versus social fairness, has no formula because it is a conflict between two real values; the underwriter's duty is competence, candor, compliance-plus, and humility — to be clear-eyed about a hard thing and unwilling to let the math hide the choice.

The Harbor Steel file came through this examination clean on its own terms — its price is risk-based, with no protected-class proxy — while reminding us that the actuarially-fair decision sits inside a social-fairness question (the coastal protection gap) that no desk can settle alone. In the next chapter we turn to the future the whole of Part VI has been building toward — AI as co-pilot, climate change repricing entire lines, and the underwriter of 2035 — carrying with us the conviction that the fairness judgment is exactly the kind of judgment that will not be automated away, because it is the one the machine is least able to make.

Key Terms

• Proxy discrimination — the use of a facially neutral, legally permitted factor that functions as a stand-in for a prohibited characteristic, producing the forbidden sorting through a permitted variable.
• Algorithmic bias — the tendency of a predictive model to produce systematically unfair outcomes for a protected group, arising from biased training data, proxy variables, or feedback loops rather than intent.
• Disparate impact — a discriminatory effect on a protected group produced by a facially neutral practice, regardless of intent, measured by comparing outcomes across groups.
• Redlining — the historical practice of denying or pricing up insurance and lending for whole neighborhoods based on racial or ethnic composition; the source of geography's contested status as a rating dimension.
• Actuarial vs. social fairness — two competing notions of a fair price: actuarial (the price accurately reflects expected risk) versus social (the price does not deepen inequality or deny essential access), which often conflict.
• Price optimization — setting a premium based partly on a customer's predicted price sensitivity (willingness to pay) rather than solely on expected loss; widely deemed unfairly discriminatory because it severs price from cost.

Spaced Review

1. Explain why insurance must discriminate by risk, and why a law forcing one flat price for all would make coverage less available rather than more fair. Name the Chapter 1 concept that drives the collapse. (§35.1)
2. A predictive model uses no protected variable yet charges minority neighborhoods more. Name the two mechanisms from this chapter that explain how that happens, and the one audit step that lets you detect it. (§35.3, §35.4)
3. Recall from Chapter 8 what a credit-based insurance score is. Explain precisely why it sits on the proxy fault line — passing two of §35.2's tests while straining a third — and why no amount of additional data will resolve the controversy. (§35.2, §35.3; Ch.8)
4. Distinguish actuarial fairness from social fairness with one original example where they point in opposite directions, and name one social-fairness mechanism (from Chapter 15, 18, or 30) that society uses to bridge the gap. (§35.7)
5. (The recurring pricing-discipline question.) A regulator caps coastal property rates below the actuarial indication to keep coverage affordable. Would this help or hurt insurers' combined ratios in that market, and what does the §35.7 trade-off predict carriers will eventually do in response? (§35.7; Ch.3, Ch.11)