Artificial intelligence is changing insurance pricing, but the hardest question is not whether algorithms should charge everyone the same price. Insurance has always depended on telling different levels of risk apart. The real challenge is deciding when technology identifies legitimate, lawful differences in risk, and when the data, proxy variables, or model design produce differences nobody can justify.
That distinction sits where prediction, actuarial practice, consumer protection, and regulation meet. AI can pick up more complex relationships than traditional models, and the same sophistication can make a decision harder to understand, audit, or challenge. There is also no single set of U.S. pricing rules: ACA rules sharply limit what can vary a premium in the individual and small-group markets, while auto, homeowners, and life insurance operate under entirely different frameworks.[^2] By 2026 the conversation has moved from broad principles toward governance, testing, explainability, monitoring, and third-party accountability, even as a federal push to preempt state AI regulation raises questions about how durable that framework will be.[^6]
One note for Tigerless AI readers: Lara is an insurance-specific platform for guidance, plan understanding, claims support, and workflow—not an insurer-side premium-setting engine.[^7]
When an Algorithm Helps Decide What You Pay
Picture requesting a quote online. You enter some information, wait a few seconds, and see a premium. Behind that number sits a set of calculations estimating how likely a future loss is and what it might cost—increasingly supported by AI.
But there is an important starting point: AI did not invent risk-based pricing. The premise was never that every customer pays the same amount; it is that price differences should bear an appropriate relationship to the risk insured, and should comply with applicable law—which differs sharply by product and jurisdiction. Variables that are routine in one line cannot simply be carried into another.[^2]
What AI changes is less the purpose of risk classification than its scale and complexity. New York's Department of Financial Services acknowledges that AI and external consumer data can speed up underwriting and pricing and may improve accuracy—while warning in the same document that unreliable data and complex systems can produce unfair or discriminatory outcomes.[^4]
It would be wrong, though, to picture every premium as already run by a sophisticated model. In NAIC surveys conducted between 2022 and 2023, 88% of auto insurers, 70% of homeowners insurers, and 58% of life insurers said they use, plan to use, or are exploring AI or machine learning somewhere in operations; in the 2025 health survey of 93 companies, 92% gave that same answer and 84% reported actually using it today.[^1] [^8] Those figures cover marketing, claims, and service—far more than premium setting.[^1]
So the shift is not humans → AI sets the price, but traditional risk variables → predictive analytics → increasingly complex AI-enabled decision support. That produces better predictions, and a harder question: how do you tell smarter risk pricing from algorithmic discrimination?
Where Fairness Questions Enter the Model
An AI system does not need a field labeled "race" to produce unequal outcomes. That is one reason AI bias in insurance is harder to solve than deleting sensitive columns from a database.
Problems often start with historical data. A model learns from what it is shown; where that reflects past economic, geographic, or institutional patterns, some of those relationships carry forward. Not every one is discriminatory—but nobody should assume training data is neutral.[^9]
Then comes the proxy problem. A proxy is an apparently neutral variable that indirectly carries information about something else. A model may never be supplied with race and still retain much of its predictive influence, because a rich enough set of correlated variables—location, purchasing behavior, device activity, financial history—can reconstruct it. Peer-reviewed actuarial research draws exactly this distinction between direct and proxy discrimination,[^10] and New York expects insurers using external consumer data to test whether their variables correlate with protected-class status in ways that may produce unlawful discrimination.[^4]
Data quality and model design matter separately. External information can be incomplete, outdated, or thin for certain populations, and New York treats its reliability as a supervisory concern in its own right.[^4] What outcome is predicted and who is represented in development data also shape results: a model with strong overall accuracy can still perform poorly for one subgroup. At the same time, a measurable difference between demographic groups does not automatically establish unfairness or illegal discrimination.[^10]
That distinction is fundamental:
Different outcomes → potential disparity → potential unfairness → potentially prohibited discrimination.
Related, but not synonyms—and research cautions that the group-fairness measures common in machine learning do not map neatly onto actuarial fairness or legal standards.[^10] Which leads to the hardest question in the debate: what does "fair" insurance pricing actually mean? Consumers, regulators, and insurers each answer differently, and those answers overlap without being identical.
That is why predictive power alone is not enough. New York expects external consumer data used in underwriting and pricing to bear a clear, empirical, rational, and non-unfairly-discriminatory relationship to the insured risk—and where a disproportionate effect appears, asks whether there is a legitimate explanation and whether a less discriminatory alternative could meet the same business need.[^4] There is a counterintuitive consequence: responsible fairness testing often requires more visibility into demographic effects, not less, since a model blind to a characteristic can be harder to audit for disparities involving it.
How Regulation and Governance Are Catching Up
The United States does not have one uniform "AI insurance law." Supervision remains predominantly state-based, layered with federal requirements that vary by product.
The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023—guidance rather than a model law, emphasizing governance, documentation, and the principle that consumer-impacting decisions supported by AI remain subject to existing insurance law.[^11] More than two dozen states have since adopted or adapted it. The work has since become operational: the NAIC's AI Systems Evaluation Tool asks insurers to inventory their AI systems, describe their governance framework, and account for high-risk models and the data feeding them, and a twelve-state pilot launched in March 2026, with adoption expected at the Fall 2026 National Meeting.[^3] A parallel third-party framework would require data and model vendors to register with state regulators; its P&C pricing and underwriting draft was exposed for comment through August 5, 2026 and remains ongoing work, not a finalized rule.[^12]
Colorado took the most explicit statutory route. SB21-169 prohibits insurers from using external consumer data and information sources (ECDIS)—or models built on them—in ways that unfairly discriminate on protected characteristics. The operative requirements sit in Regulation 10-1-1: a board-overseen governance framework, model inventory, quantitative testing, documentation, and an annual officer attestation. The life insurance framework took effect November 2023, and a 2025 amendment extended scope toward private passenger auto and health benefit plans, with line-specific rules still in rulemaking.[^5] Note that SB21-169 is not the Colorado AI Act—that law (SB24-205) was repealed and replaced in May 2026 by SB26-189, effective January 2027 and administered by the Attorney General.[^13]
New York offers the clearest practical governance example. Its 2024 Insurance Circular Letter No. 7 calls for actuarial validity, proxy assessments, discrimination testing before production and regularly thereafter, documentation, model monitoring, and oversight of third-party systems. It also takes a clear position on "the vendor built it": an insurer remains responsible when it relies on third-party systems, and cannot rest on a vendor's claim of non-discrimination or proprietary technology—either to establish compliance or to justify a thin explanation of an adverse pricing outcome.[^4]
Running against all of this is a federal effort to narrow state authority. Executive Order 14365 (December 11, 2025) directs the DOJ to challenge state AI laws, and a March 2026 White House framework recommended that Congress legislate broad preemption.[^6] Neither preempts state law on its own, and Congress has so far declined—so the consistent advice from counsel is to keep complying with state requirements.
The emerging principle survives that uncertainty: a model that predicts well but cannot be adequately tested, governed, monitored, or explained is difficult to operate responsibly in a regulated insurance environment.
What Responsible AI Pricing Looks Like
Responsible artificial intelligence insurance pricing is not about choosing a more powerful model. A defensible program needs controls around it:
Relevant and reliable data — you can articulate why each input bears a legitimate relationship to the risk, and you understand its provenance, accuracy, and limits.
Bias testing — before deployment and on a regular cadence afterward, not as a one-time validation.
Explainability — enough understanding of material decision factors to govern the system and answer consumers and regulators.
Human oversight — meaningful review or escalation, not a person rubber-stamping the end of an unexamined process.
Documentation and vendor governance — purpose, data sources, tests, approvals, and changes for each material system, extended to externally supplied data and models.
Continuous monitoring — drift, changing populations, new disparities, complaints, and material post-deployment changes.
These parallel New York's guidance and the NAIC's governance approach.[^4] [^11] In practice they form a loop rather than a checklist: define the use case, map data and vendor dependencies, test performance and disparities, review governance and explanations, deploy with escalation paths, monitor, and return to testing whenever something material changes. New York calls for discrimination testing before production, on a regular cadence, and after material model or data changes, plus annual testing for drift.[^4]
This also draws a useful line between AI that supports insurance understanding and service delivery and AI that determines an insurer's underlying risk price. Lara sits on the first side. Carrier ratemaking, insurer-side risk classification, and pricing-fairness validation are separate functions with their own obligations.[^7]
What It Means for Consumers—and the Future of Insurance
Consumers do not need to understand machine learning to ask useful questions when a premium changes: Why did it change? What information contributed? Did any of it come from an outside source? Can errors be corrected? What you are entitled to depends on the product, the data, and the jurisdiction—New York sets detailed expectations for adverse pricing notices,[^4] but those are New York's, not a universal rule.
The longer-term question may matter even more than bias testing. Insurance works by pooling uncertain losses while still distinguishing among meaningful classes of risk, and more granular data makes it possible to move toward individual profiles. The trade-off is genuine: more precise AI risk assessment in insurance means lower-loss policyholders subsidize higher-loss ones less. But researchers have argued that individualized prediction creates real tension with pooling itself—at what point does better differentiation weaken what makes insurance socially and economically valuable?[^14] No model predicts an individual's future losses perfectly today, but the thought experiment is useful:
If insurers could eventually predict individual risk with extraordinary precision, would insurance become fairer—or would some people simply become too expensive to insure?
No statistical formula resolves that, and neither "AI is discriminatory" nor "AI will eliminate human bias" survives contact with the evidence. The regulatory direction points somewhere subtler: powerful models need institutions around them capable of deciding what should be predicted, what data should be permitted, how disparities should be read, and who remains accountable when the technology comes from outside the insurer.
The future of AI in insurance pricing will therefore depend on more than predictive accuracy. Insurers and their technology partners will need credible answers to four questions: What data is used? Why is it relevant? Does the model create disparities that cannot be justified? Can the decision be explained, monitored, challenged, and corrected?
Greater sophistication does not automatically create greater fairness—but neither does it make unfairness inevitable. The design challenge is to preserve legitimate risk-based insurance while making powerful systems accountable for the decisions they influence.
The future of insurance AI may not be decided by whether algorithms can price risk better than humans. It may be decided by whether we can build systems powerful enough to improve insurance—and transparent enough to deserve trust.
Sources
[^1]: NAIC, Artificial Intelligence insurance topic page, including AI/ML survey results by line of business. https://content.naic.org/insurance-topics/artificial-intelligence [^2]: 45 C.F.R. § 147.102, health insurance premium rating factors for the individual and small-group markets. [^3]: NAIC Big Data and Artificial Intelligence (H) Working Group, AI Systems Evaluation Tool and twelve-state pilot (March–September 2026). https://content.naic.org/committees/h/big-data-artificial-intelligence-wg [^4]: NYS Department of Financial Services, Insurance Circular Letter No. 7 (2024), "Use of Artificial Intelligence Systems and External Consumer Data and Information Sources in Insurance Underwriting and Pricing." https://www.dfs.ny.gov/ [^5]: Colorado SB21-169 (2021) and Colorado Insurance Regulation 10-1-1. https://doi.colorado.gov/for-consumers/sb21-169-protecting-consumers-from-unfair-discrimination-in-insurance-practices [^6]: Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence, December 11, 2025; White House National Policy Framework for AI, March 20, 2026. https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/ [^7]: Tigerless AI public product materials describing Lara. [Insert live product-page URL before publication.] [^8]: NAIC, Health Insurance AI/ML Survey Results, May 2025. https://content.naic.org/article/naic-survey-reveals-majority-health-insurers-embrace-ai [^9]: American Academy of Actuaries and Society of Actuaries research on bias and fairness in insurance AI. https://www.actuary.org/ [^10]: Peer-reviewed actuarial literature on direct versus proxy discrimination and the mismatch between machine-learning group-fairness metrics and actuarial or legal standards. [^11]: NAIC, Model Bulletin: Use of Artificial Intelligence Systems by Insurers, adopted December 2023. https://content.naic.org/sites/default/files/cmte-h-big-data-artificial-intelligence-wg-ai-model-bulletin.pdf.pdf [^12]: NAIC Third-Party Data and Models (H) Working Group, Regulatory Framework for Third-Party Data and Model Vendors, exposed for comment through August 5, 2026. https://content.naic.org/committees/h/third-party-data-models-wg [^13]: Colorado SB26-189 (May 2026), repealing and replacing the Colorado AI Act (SB24-205), effective January 1, 2027. [^14]: Academic literature on individualized prediction and the erosion of risk pooling in data-intensive insurance.
This article is for general information and does not constitute legal, actuarial, or compliance advice. Requirements vary by state, insurance line, and product, and continue to change.
