Chapter 35 — Key Takeaways

A one-page card for the chapter's hardest subject. If you remember nothing else, remember that insurance must discriminate by risk — and that the whole ethical weight of the job sits on the contested line between that and discriminating against people.

The core claims

  • The paradox. Insurance works by discriminating in the technical sense — sorting risks by expected loss. Remove that and the pool collapses through adverse selection (Chapter 1). Risk-based discrimination is the mechanism, not a flaw. The forbidden kind is sorting by protected class or by factors unmoored from cost.
  • Four tests for the line. A rating factor is suspect unless it passes: (1) protected-class (is it a prohibited characteristic? — categorical, no exception); (2) actuarial justification (real loss relationship?); (3) disparate impact (does it fall much harder on a protected group in effect?); (4) causation vs. correlation (closer to a cause of loss, or a coincidence with a protected trait?).
  • Proxy discrimination kills "colorblindness." A neutral, legal factor can stand in for a prohibited one (ZIP code ↔ race in a segregated society). Deleting the protected variable does not stop the discrimination — it only blinds you to it. You must measure protected-group impact to detect proxies.
  • Algorithms make it worse. Models perpetuate bias via (a) biased training data, (b) proxy variables reconstructed from correlates, and (c) feedback loops that compound. The fix is disparate-impact testing in model governance, not heroic individual override.
  • The fairness metrics conflict. Demographic parity, equalized odds, and calibration cannot all hold at once when base loss rates differ — a theorem. Choosing among them is a values decision, not a technical one. "Fair" has no single computable meaning.
  • Redlining's legacy is present tense. Geography is insurance's most haunted factor: a real peril (cat exposure) and a proxy for historical segregation. You can perpetuate an injustice with no intent, just by faithfully pricing a world injustice shaped.
  • The deep tension has no formula. Actuarial fairness (price reflects risk) vs. social fairness (don't deepen inequality or deny access) often point opposite ways — e.g., the coastal protection gap (Chapter 30). Society bridges it with community rating, residual/FAIR plans, rate caps, and public programs — each a deliberate, costly departure from pure risk pricing.
  • The law is moving from intent to effect. Old unfair-trade-practices acts asked about intent; new AI regulation (Colorado SB21-169, the NAIC AI bulletin) asks about outcomes and can require you to prove no prohibited disparate impact.

The rule of thumb

A practice's profitability, predictive power, and apparent neutrality tell you nothing about its fairness. Ask every time: Is the price tracking cost? Is a neutral factor carrying protected information? The data will never tell you when it has crossed a line it cannot see.

The limiting case

Price optimization — pricing what a customer will tolerate rather than what they cost — is the clearest line-crossing in the book: profitable, data-driven, legal-sounding, and still unfairly discriminatory, because it severs price from risk. When you price captivity instead of peril, you have stopped underwriting and started extracting. (Widely banned/restricted in personal lines.)

Key terms

proxy discrimination · algorithmic bias · disparate impact · redlining · actuarial vs. social fairness · price optimization — plus the GINA gap (genetics protected in health, not in life/disability/LTC).

What you could defend to your manager

"I considered the fairness of this price under all four tests. The factors driving it have clean, causal loss stories — there's no protected class and no plausible proxy for one. I documented that. Where a factor strained the disparate-impact test, I flagged it for governance testing rather than burying it. And I noted, honestly, where an actuarially-fair decision sits inside a social-fairness problem that belongs above my desk. I priced the risk; I didn't price the person."