AI Bias in Insurance: Can Algorithms Discriminate Against Consumers?

There’s a phrase that keeps appearing in insurance industry presentations and regulatory filings, and it deserves more scrutiny than it typically gets: “data-driven decision-making.” When an insurer says its pricing is data-driven, the implicit promise is objectivity — that a model built on numbers is, by definition, free from the kind of human prejudice that gave us redlining, discriminatory underwriting, and the long history of minority communities being denied coverage or overcharged for it. The problem with that promise, and the reason courtrooms across the country are currently full of cases that challenge it, is that data doesn’t exist in a vacuum. It was generated by a society with structural inequalities baked into it, and when you feed that data to a machine learning model, the model learns those inequalities just as efficiently as it learns everything else. Lets dive in to AI bias in insurance.

This isn’t a fringe concern raised by AI sceptics. It’s a documented, well-researched phenomenon that regulators in multiple jurisdictions are now treating as a live enforcement priority. The American Academy of Actuaries published a report titled “Unmasking Hidden Bias” in 2025, detailing how behavioural data and proxy variables institutionalise discrimination in underwriting. Research from the University of New South Wales found that with AI and big data, some proxy variables are no longer easy for even actuaries to identify. When the people building the models can’t reliably spot the discriminatory variables inside them, the idea that consumers can meaningfully contest those outcomes becomes almost theoretical.

What follows is an attempt to lay out how algorithmic discrimination in insurance actually works, who it’s hurting, what the courts are doing about it, and what consumers have the right to expect from an industry that is deploying AI at a scale that makes getting this wrong enormously expensive — not just legally, but morally.

How a Model Can Be Biased Without Trying To Be

The most important concept to grasp when thinking about AI bias in insurance is proxy discrimination. It’s the mechanism that allows a model to produce racially or economically discriminatory outcomes without ever referencing race or income directly. The model doesn’t need to know you’re Black, or Hispanic, or living paycheck to paycheck. It just needs to know your ZIP code, your credit score, your education level, and perhaps what kind of device you used to access the quote page — and from those inputs, it can infer enough about you to price you differently.

In what the Consumer Federation of America has called the “credit penalty,” property insurance companies charge the typical homeowner $1,996 more each year — almost double — just for having a lower credit score than their otherwise identical neighbours. This pricing practice disproportionately impacts Black and Hispanic homeowners, who tend to have weaker credit histories due to the racial wealth gap and persistent structural barriers. Notably, companies generally do not disclose their algorithms to the public, nor are the resulting underwriting decisions, claims, and pricing disclosed publicly.

The ZIP code problem is equally fundamental and similarly obscured by the appearance of actuarial legitimacy. The Casualty Actuarial Society has acknowledged the potential impact of systemic racism on insurance underwriting, rating, and claims practices, noting that while the use of geographic location is supported by its correlation to loss, location may also be correlated with race due to ongoing societal segregation. That correlation is the crux of the issue. A model that penalises you for living somewhere is, in many American cities, functionally penalising you for living in a neighbourhood that was racially segregated through explicit historical policy — and then continued to be so through the economic consequences of that policy. The model didn’t create the segregation. But it’s pricing from it.

A 2017 ProPublica study found that insurers such as Allstate and Geico were charging algorithmically determined premiums that were, depending on the jurisdiction, as much as 10 to 30 percent higher on average in ZIP codes where most residents are minorities than in whiter neighbourhoods with similar accident costs. In 2020, The Markup found that a proposed rate-setting algorithmic system from Allstate would have disproportionately affected people living in communities that were 75 percent or more nonwhite. These findings predated the current generation of AI-powered underwriting tools by several years. The models being deployed now are significantly more powerful, work with significantly more variables, and are significantly harder to audit for the same patterns.

Real Cases, Real Consequences

The argument that algorithmic bias is a theoretical future problem rather than a present reality is being decisively dismantled by the litigation record. All three major AI discrimination lawsuits in insurance survived motions to dismiss in 2025. Huskey v. State Farm, filed in 2022, alleges the insurer’s machine-learning fraud algorithms use racial proxies, and the case is now in discovery. In November 2025, Los Angeles County launched a civil investigation into State Farm’s use of AI tools to review wildfire claims, and California’s insurance commissioner opened a parallel market conduct examination. Surviving a motion to dismiss is legally significant: it means a judge has found the allegations plausible enough to proceed. In AI discrimination cases, it also means the insurer must open its model architecture, training data, and decision logs to the opposing side — which is precisely what the industry has resisted doing.

The health insurance sector has its own cluster of cases, and the scale of the alleged harm there is considerable. Lawsuits faced by major insurers including State Farm and Cigna have highlighted how algorithmic bias in AI decision-making can create legal, ethical, and reputational risks for insurers. Many AI systems operate as “black boxes,” making it difficult for insurers, regulators, and consumers to understand how decisions are made — which complicates regulatory oversight and accountability, especially in underwriting and pricing practices that impact consumer rights and access to insurance.

The academic evidence reinforces what the courtrooms are discovering. A 2026 study published in Springer’s Lecture Notes in Computer Science examined a real-world insurance machine learning model used to identify claims likely to become costly. The baseline analysis revealed significant discrimination against female claimants compared to male claimants. While mitigation methods successfully improved fairness metrics, these improvements came at a cost to predictive performance — a trade-off that sits at the heart of the entire debate. That trade-off — between a more accurate model and a fairer one — is not a technical footnote. It’s the central policy question of AI-driven insurance pricing, and different jurisdictions are landing in very different places on how to resolve it.

For anyone trying to understand the full picture of how AI systems make consequential decisions that can’t easily be contested, our piece on AI liability insurance and who pays when algorithms make expensive mistakes covers the legal accountability gap in detail — including why “the AI decided” is not a defence that courts are currently accepting.

The Black Box Problem: When Nobody Can Explain the Decision

One of the properties of modern machine learning models that makes algorithmic bias particularly difficult to address is their opacity. A traditional actuarial model uses a defined formula with visible variables and weights that a regulator can examine and a consumer can, in principle, understand. A gradient boosting model trained on 1,500 variables doesn’t work that way. Its outputs are mathematically determined by patterns learned from the training data, but tracing any individual output back to a specific cause is genuinely difficult — sometimes impossible.

Many AI systems operate as black boxes, making it difficult for insurers, regulators, and consumers to understand how decisions are made. This lack of explainability complicates regulatory oversight and accountability, especially in underwriting and pricing practices that directly impact consumer rights and access to insurance. For a consumer who’s been quoted a premium 40% higher than their neighbour’s, the inability to get a meaningful explanation isn’t just frustrating — it forecloses their ability to identify discrimination, much less challenge it.

Historical and ongoing racial discrimination has created an enormous racial wealth gap, and because American society remains significantly segregated, almost all the data held by data brokers reflects and encodes racial disparities. When predictive models are built using this data, people of colour are consistently disadvantaged — in ways that are legally cognisable under the Fair Housing Act when they occur in the housing context, and under Title VII in employment. But federal insurance law has no equivalent bright-line protection. That legal gap is a significant part of why the litigation is proceeding at the state level and why regulatory responses have been fragmented.

This opacity issue is directly connected to the broader surveillance infrastructure now feeding insurance models — the behavioural data, location history, and purchase patterns flowing into pricing algorithms through channels consumers have no visibility into. The dynamics are explored in depth in our analysis of telematics insurance privacy risks and what your car is actually tracking, and at the checkout-page level in our piece on embedded insurance and the hidden data collection inside apps.

The Regulatory Response: Patchwork, Imperfect, but Moving

The regulatory picture in 2026 is genuinely more active than it was three years ago, even if it falls well short of a coherent national framework. The most consequential state-level interventions have come from Colorado and New York, which have established different but complementary approaches to holding AI-driven underwriting accountable.

New York’s DFS Circular Letter 2024-7 requires insurers to demonstrate that AI and external data systems do not proxy for protected classes or generate disproportionate adverse effects. Insurers must keep explanatory documentation, allow the Department of Financial Services to review vendor tools, require vendor audits, and ensure internal oversight. Colorado Revised Statutes Section 10-3-1104.9 prohibits the use of external consumer data sources and predictive models that result in unfair discrimination, requiring performance of quantitative testing to detect disparate impact even if the data or model is facially neutral. Colorado’s law is particularly notable because it extends to situations where bias emerges from models that appear neutral on the surface — which is precisely the proxy discrimination mechanism described above.

At the federal level, progress has been slower and more contested. No standalone federal law specifically addressing algorithmic discrimination in insurance has been enacted yet. The European Union’s AI Act, enacted mid-2024 and being phased in through 2026, classifies AI systems used in credit underwriting as high-risk, imposing rigorous requirements for bias testing, technical documentation, and post-deployment monitoring on any insurer with European exposure. The contrast between the EU’s binding framework and the US patchwork of state regulations is increasingly consequential for multinational insurers, who must now maintain compliance with two substantially different governance regimes simultaneously.

California’s SB 1120, effective January 2025, prohibits health insurers from denying coverage based solely on an AI algorithm. The NAIC’s 2025 health insurance survey found that nearly one-third of health insurers don’t regularly test their AI models for bias. That last finding remains, two years on, the most damning single statistic in the entire AI bias debate: a third of health insurers are deploying systems capable of producing discriminatory pricing outcomes at scale without routinely checking whether they’re doing so. The gap between what is legally required and what is actually being done is where most of the harm lives.

Understanding how algorithmic underwriting works mechanically — what variables feed into these models and how they produce pricing outputs — is essential context for evaluating the bias risk. Our deep dive on how AI underwriting algorithms set your insurance premiums in 2026 covers the technical architecture in plain language.

Frequently Asked Questions

Q: Is it actually legal for insurers to use my ZIP code or credit score to set premiums?

In most US states, currently yes — though this is being actively contested. Insurance companies generally do not disclose their algorithms to the public, and the resulting underwriting decisions, claims, and pricing are not publicly disclosed either. The credit penalty practice — charging homeowners nearly double for having lower credit scores — is currently lawful in most jurisdictions, despite its documented disproportionate impact on Black and Hispanic communities. California and a handful of other states have moved to restrict credit score use in auto insurance, but no uniform federal prohibition exists.

Q: How do I know if an algorithm has treated me unfairly?

Practically speaking, it’s very difficult to detect without regulatory tools. New York now requires insurers to keep explanatory documentation and allow regulatory review of vendor tools — but that information isn’t directly accessible to consumers. If you receive an adverse underwriting decision, you have the right to request the specific reason in writing in most states. If your claim is denied, federal lawsuits against Cigna and UnitedHealthcare have established that alleging algorithmic claim denial is a viable legal theory. Your state insurance commissioner is the appropriate first port of call for a discrimination complaint.

Q: Can insurers fix algorithmic bias without making their models less accurate?

Research published in 2026 found that while de-biasing techniques successfully improved fairness metrics in real insurance models, these improvements came at a cost to predictive performance. De-biasing techniques applied to insurance data have been shown to close 65 to 82% of bias gaps — but accuracy losses are real and create commercial resistance to implementation. This trade-off is genuine, and regulators are increasingly requiring insurers to navigate it explicitly rather than prioritising accuracy over fairness by default.

Q: What is “disparate impact” and why does it matter for AI insurance decisions?

Disparate impact is the legal standard under which a facially neutral policy or practice can be found discriminatory if it produces disproportionately adverse outcomes for a protected class — even without any intent to discriminate. Colorado’s insurance AI regulation requires quantitative testing to detect disparate impact even when data or models are facially neutral — meaning an insurer cannot defend a discriminatory pricing outcome by pointing to the neutrality of its inputs if the outputs produce disproportionate harm to a protected group. This is the legal framework most likely to be applied to proxy discrimination cases in insurance going forward.

Q: What should I do if I think my insurance premium reflects algorithmic discrimination?

Start by requesting a written explanation of how your premium or decision was calculated. In most states you have the right to request the specific reason for an adverse decision in writing and to appeal, including the right to a human review. If you believe discrimination has occurred, file a complaint with your state insurance commissioner, and consider contacting a consumer advocacy organisation such as the Consumer Federation of America, which has published detailed research on bluelining and algorithmic insurance exclusion. In states with active AI bias regulations — Colorado, California, New York — regulators have the tools to compel audit access that individual consumers cannot obtain independently.

The Bottom Line

The insurance industry’s adoption of AI has genuinely produced efficiency gains, faster decisions, and — for some consumers — fairer prices that reflect their individual behaviour rather than crude demographic averages. None of that is in dispute. What is also not in dispute, at this point, is that the same systems are producing discriminatory outcomes for a significant portion of the population — not because insurers are deliberately targeting protected classes, but because the data those models are trained on is the archaeological record of structural inequality, and the models are learning from it faithfully.

Insurers seem captivated by two trends enabled by AI: data-intensive underwriting that analyses more and newer types of data to assess risk more precisely, and behaviour-based insurance that monitors individual consumers in real time. While these trends bring many advantages, they may also have discriminatory effects on society at a scale that traditional insurance never could. Scale is precisely the issue. A biased human underwriter affects one application at a time. A biased model affects every application it processes, at whatever speed the compute allows.

The courts are beginning to provide some accountability. Regulators in several jurisdictions are demanding explainability and bias testing. And the legal standard of disparate impact means that good intentions are not a defence against discriminatory outcomes. The burden of proof is shifting — slowly, and not uniformly — from consumers who must prove they were discriminated against, to insurers who must prove their models don’t discriminate. That’s progress. Whether it arrives fast enough, and comprehensively enough, to match the speed at which AI-driven pricing is being deployed is the question the next decade of insurance regulation will answer.

This article is for informational purposes only and does not constitute legal or insurance advice. Regulatory requirements vary significantly by jurisdiction. Always consult a qualified professional for guidance specific to your situation.

External Authority Sources Used: