Congress Probes Credit-Based Insurance Scores as Rates Climb

On August 4, 2026, Senator Elizabeth Warren and Representative Ayanna Pressley led twenty members of Congress in sending letters to USAA, State Farm, Progressive, Liberty Mutual, Farmers, and Allstate, demanding answers by August 17 on how credit-based insurance scores drive homeowners underwriting and pricing (Pressley press release, August 2026). The letters reopen the longest-running rating-variable dispute in personal lines at the moment state AI-governance frameworks have given regulators new language for it.

What the Letters Ask, and Who Signed

The signatories span both chambers: seven senators, including Warren, Richard Blumenthal, Tammy Duckworth, Ruben Gallego, Adam Schiff, Chris Van Hollen, and Ron Wyden, and thirteen House members including Pressley, Alexandria Ocasio-Cortez, Ilhan Omar, Rashida Tlaib, and Jim McGovern (InsuranceNewsNet, August 2026). The demand is information, not legislation: how each carrier uses credit-based insurance scores in underwriting and pricing homeowners policies, with responses due August 17.

Two lines in the letters tell actuaries where the argument is headed. "Insurers should not charge consumers higher premiums simply because of their personal credit history," the lawmakers write, framing the variable as causally irrelevant to loss. The sharper claim is empirical: "Credit scores impact homeowners insurance premiums as much, if not more than, disaster risk in many parts of the country" (InsuranceNewsNet, August 2026). That comparison, credit weight versus catastrophe-peril weight in the rate, is a testable statement about relativities, and it is the one carriers' actuarial responses will have to engage. If a below-average credit tier carries a larger rate relativity than the spread between a low-hazard and high-hazard wind or wildfire territory, the lawmakers' framing will survive contact with the filings.

Two Decades of Evidence, Read Honestly

The evidence record on credit-based insurance scores is deeper than almost any other rating variable, and it points in two directions at once. The Federal Trade Commission's 2007 report to Congress, the most cited study in the field, found that credit-based insurance scores are effective predictors of claims risk under automobile policies: scores correlated with claim frequency even after controlling for other rating variables (FTC, July 2007). The Texas Department of Insurance reached similar predictiveness conclusions from carrier data in its mid-2000s studies. On the actuarial merits, the variable works: it separates loss experience.

The same FTC report documented the distributional fact that keeps the fight alive: scores are not evenly distributed across racial and ethnic groups, with Black and Hispanic consumers overrepresented in the lowest score bands. Predictive lift and disparate distribution are both true simultaneously, which is why the debate never resolves. Actuarial standards address risk classification on the first axis; the political system responds to the second. Every few years the second axis produces a probe, a bill, or a rulemaking, and the industry's defense returns to the first. The August letters are the newest iteration, aimed at homeowners rather than the auto book the FTC studied, and homeowners is the harder line to defend right now: premiums have climbed steeply with catastrophe costs, so any variable that adds rate on top of peril-driven increases is politically exposed.

The State Map: Where Credit Already Cannot Be Used

Congress is probing a practice that several states have already restricted, which gives both sides natural experiments. California, Hawaii, and Massachusetts prohibit credit-based insurance scores in auto insurance, and Michigan, Oregon, and Utah restrict them; for homeowners, California, Massachusetts, and Maryland bar the practice, while Hawaii's ban is auto-only (Experian, 2026). Washington's insurance commissioner tried to ban credit scoring by emergency rule in 2021 and was struck down in state court the following year, a reminder that the authority to remove a rating variable usually runs through legislatures, not commissioners. Bills to bar credit in homeowners or auto pricing are pending in Iowa, New York, Oklahoma, and Pennsylvania (Live Insurance News, 2026).

The ban states teach a lesson the letters do not mention: removing a predictive variable does not remove the premium, it redistributes it. When credit leaves the rating plan, the rate that credit was carrying migrates to the variables that remain, territory, prior claims, home characteristics, and the policyholders who benefited from strong credit see increases while those penalized by weak credit see decreases. The aggregate rate need stays anchored to losses. Whether that redistribution is good policy is a values question; that it happens is arithmetic. Carriers' August 17 responses would be more persuasive engaging that trade-off directly than restating predictiveness statistics the FTC already conceded two decades ago.

Testing the Central Claim from Public Filings

The lawmakers' strongest sentence, that credit moves homeowners premiums as much as disaster risk, can be checked by anyone with SERFF access, and carriers should assume committee staff will do exactly that. The test has four steps. Pull the current approved homeowners rating manual for a given carrier and state. Extract the credit-tier factor table and compute the ratio of the worst tier to the best tier; in many public filings that spread runs well past 2-to-1. Extract the territory factor table for the same program and compute the spread between the lowest-hazard and highest-hazard territories the carrier actually writes. Compare the two ratios. Where the credit spread exceeds the territory spread, the letter's claim holds for that program, full stop, and no predictiveness argument changes the arithmetic the committee will publish.

What the four-step test misses is also worth stating, because it is the carriers' best factual defense. Territory factors are not the whole catastrophe load: hurricane and wind-hail deductibles, roof schedules, by-peril rating, and inspection-driven surcharges carry hazard signal outside the territory table, so a naive two-column comparison understates the disaster-risk share of the rate. A response that walks through the full peril-load architecture, with the program's own numbers, engages the claim on its merits. A response that recites the FTC's predictiveness finding and stops will read as a concession that the relativity comparison was never run. The difference between those two responses is the difference between an actuary drafting the answer and a government-affairs office drafting it.

Why This Lands Inside the AI Rating Debate

The probe arrives with different regulatory machinery than the last credit fight. More than 20 jurisdictions have adopted the NAIC's model bulletin on insurer AI use, and Colorado's quantitative testing regime under SB 21-169 now requires life insurers to test models and external data for unfairly discriminatory outcomes, a template its regulators have signaled extends to other lines (NAIC, 2026). Credit-based insurance scores are exactly the kind of external, non-causal consumer data those frameworks were written for. A carrier that runs credit through a machine-learning rating plan has converted a familiar scalar variable into a feature inside a model subject to governance, documentation, and outcome-testing expectations that did not exist in 2007.

That convergence cuts both ways. It gives critics a governance vocabulary, proxy discrimination, outcome testing, explainability, that is harder to dismiss than fairness rhetoric. It also gives carriers a defensible playbook: the disparate-impact testing methods developed for AI compliance, including inferred-demographics analysis of rating outcomes, work just as well on credit as on any novel variable. A carrier that has quantified the demographic footprint of its credit relativities, tested whether alternative variables recover the same predictive lift with less disparity, and documented the result has an answer for Congress, for state regulators, and for its own filings. A carrier that has never run the test is betting that nobody with subpoena power will ever ask. Twenty members of Congress just asked.

The Dislocation Problem a Ban Would Create

Suppose the probe eventually produces what its authors want, a homeowners book priced without credit. The transition mechanics deserve more attention than the policy debate gives them. Removing a variable with a 2-to-1 or wider factor spread repricing an entire book produces dislocation in both directions: the policyholders who were receiving the best-tier discount face increases at renewal, and they are, by construction, the segment with the most financial flexibility to shop. The likely result is asymmetric retention: the newly surcharged good-credit segment defects to carriers or channels that price them better, exempt surplus-lines paper among them, while the newly discounted segment stays. The remaining pool's loss experience deteriorates relative to the rate level assumed at transition, which forces the next rate increase, which drives the next round of defection. That spiral is not hypothetical to anyone who has managed a book through a mandated rating change; it is why transition rules, renewal capping, and multi-year glide paths dominate the implementation discussions in states that have actually removed variables.

The capping math contains its own trap. A carrier that caps renewal increases at, say, 10% a year while the uncapped indication for a formerly best-tier insured runs 25% carries the difference as a systematic rate inadequacy that earns through for years, concentrated in its most desirable segment. Pricing actuaries modeling a credit-free plan today should be modeling the capped transition path, not just the end-state relativities, because the transition is where the underwriting margin actually lives or dies. The states with pending bills, Iowa, New York, Oklahoma, and Pennsylvania, will write the transition rules that decide this, and carriers with modeled answers will have more influence over that drafting than carriers arriving with talking points.

What the Six Carriers Face Before August 17

The immediate task for the recipients is a response that is accurate, filing-consistent, and does not create discovery material for the class-action bar. Every quantitative claim in the responses will be checked against the carriers' own public rate filings, where credit tier factors are visible to anyone who reads them. The deeper task belongs to pricing actuaries over the next several quarters. If the probe matures into legislation or emboldened state action, the homeowners book faces the same transition auto carriers navigated in ban states: re-fitting rating plans without credit, managing the dislocation for the large block of policyholders whose rates rise, and defending the replacement variables, which will attract the same proxy-discrimination scrutiny the moment they absorb credit's predictive role. The carriers that handle it best will be the ones that started modeling the credit-free version of their rating plan before anyone required it. The August 17 responses are the opening filings in that longer case.

The calendar compresses the question. The NAIC Summer National Meeting convenes in Columbus August 11 through 14, with the Big Data and Artificial Intelligence Working Group on the agenda for August 13, three days after the congressional letters land in carrier mailrooms and four days before responses are due. State regulators will be sitting in a room discussing AI-governance instruments in the same week national reporters are writing about credit scores in homeowners rates. Whether or not the working group takes up credit directly, the coincidence hands commissioners a live example of exactly the external-data governance question their frameworks claim to answer, and it puts the six recipients' regulatory-affairs teams in Columbus and on deadline simultaneously. Watch what the responding carriers say publicly that week; the ones that say nothing are the ones still deciding what their filings can support.