AI Insurance Discrimination: Can Algorithms Unfairly Change Your Premium, Coverage or Claim?
Quick Answer: Yes. Artificial intelligence can create discriminatory outcomes in insurance if its data, variables, models or implementation result in unlawful or unfair discrimination. An algorithm does not become legally neutral simply because it does not explicitly use a protected characteristic. Insurance regulators increasingly focus on fairness, actuarial justification, data quality, proxy variables, model governance and consumer outcomes.
Imagine two consumers applying for identical insurance coverage.
They have similar risk profiles.
But one receives a substantially higher premium.
The consumer asks:
“Why?”
The insurer responds:
“Our AI model calculated that you present a higher expected risk.”
The consumer then asks:
“What information did the AI use?”
Perhaps the answer includes:
- Geographic location.
- Purchasing behaviour.
- Credit-related information.
- Property characteristics.
- Claims history.
- Online activity.
- Other external consumer data.
None of those variables may explicitly state the consumer's race, ethnicity or another protected characteristic.
Yet some may correlate strongly with characteristics protected under applicable law.
This is where the problem of proxy discrimination arises.
Artificial intelligence can identify statistical relationships that humans may not notice.
That is one of its greatest strengths.
It can also be one of its greatest legal risks.
The National Association of Insurance Commissioners (NAIC) states that insurers remain responsible for complying with insurance laws, regulations, insurance standards and consumer-protection rules when using AI, including requirements concerning fairness, accuracy and avoiding unfair discrimination. State regulators may require insurers to explain AI use in underwriting, pricing, marketing and claims. ([content.naic.org](https://content.naic.org/insurance-topics/artificial-intelligence?utm_source=chatgpt.com))
The NAIC's current race-and-insurance materials make an important distinction: risk-based differentiation is fundamental to insurance, but unfair discrimination is unlawful where pricing or access is based on factors that are not actuarially justified and are legally or socially unacceptable. ([content.naic.org](https://content.naic.org/insurance-topics/race-and-insurance?utm_source=chatgpt.com))
That creates the central legal problem:
Insurance requires differentiation based on risk. Insurance law prohibits certain forms of unfair discrimination. AI can make the boundary harder to identify.
This article explains how that problem arises and how regulators are responding.
Legal disclaimer: This article provides general educational information and is not legal, insurance, actuarial, financial or regulatory advice. Insurance discrimination law varies by state, insurance product and factual circumstances.
Key Takeaways
- Insurance necessarily involves risk classification and differentiation.
- Not every difference in insurance pricing is unlawful discrimination.
- AI can nevertheless create unfair discriminatory outcomes.
- An algorithm can discriminate even without explicitly using a protected characteristic.
- Proxy variables can indirectly reproduce protected characteristics.
- Historical data can carry forward historical patterns of inequality.
- External data can create additional discrimination and accuracy risks.
- Predictive accuracy does not automatically establish legal fairness.
- Actuarial justification remains important in evaluating risk classifications.
- AI discrimination can arise in underwriting, pricing, marketing and claims.
- Insurers remain responsible for AI systems used by or on their behalf.
- State insurance regulators are increasingly developing AI examination and testing capabilities.
What Is AI Insurance Discrimination?
Quick Answer: AI insurance discrimination occurs when an AI-supported insurance process produces unlawful or unfair discriminatory treatment in areas such as pricing, underwriting, eligibility, marketing or claims.
The discrimination may be:
- Direct.
- Indirect.
- Intentional.
- Unintentional.
The precise legal standard depends on the applicable jurisdiction and insurance product.
Why Is Discrimination Different in Insurance?
Quick Answer: Insurance is fundamentally based on risk classification, meaning that insurers routinely distinguish between different risk categories.
For example, an insurer may charge different premiums because two applicants have materially different expected risks.
The NAIC recognises that risk-based pricing is essential to the insurance system. At the same time, it distinguishes legitimate risk classification from unfair discrimination. ([content.naic.org](https://content.naic.org/insurance-topics/race-and-insurance?utm_source=chatgpt.com))
Therefore:
Different price ≠ automatically discriminatory.
But:
Different price based on an impermissible or inadequately justified classification may create a legal problem.
What Is Unfair Discrimination in Insurance?
Quick Answer: Unfair discrimination generally concerns insurance classifications or practices that are prohibited by applicable law or that lack adequate actuarial or legal justification.
The exact rules vary between states and lines of insurance.
The NAIC explains that insurance discrimination becomes unlawful where classifications are not actuarially justified and involve legally or socially unacceptable factors such as race, ethnicity or national origin. ([content.naic.org](https://content.naic.org/insurance-topics/race-and-insurance?utm_source=chatgpt.com))
What Is Algorithmic Bias?
Quick Answer: Algorithmic bias occurs when an algorithm systematically produces outcomes that reflect problematic assumptions, data patterns or design choices.
Bias can enter through:
- Training data.
- Feature selection.
- Model design.
- Target variables.
- Data collection.
- Implementation.
- Human decisions surrounding the model.
AI does not automatically eliminate human bias.
It can sometimes automate and scale it.
Can AI Discriminate Without Being Programmed to Discriminate?
Quick Answer: Yes.
A machine-learning system does not necessarily need an explicit discriminatory instruction.
Suppose historical data contains a pattern.
The algorithm learns the pattern.
The model then uses that pattern to predict risk.
If the historical pattern reflects structural or discriminatory conditions, the model can reproduce those effects even without being told to discriminate.
This is one reason regulators focus on outcomes and not merely the variables explicitly entered into the model.
What Is Proxy Discrimination?
Quick Answer: Proxy discrimination occurs when an apparently neutral variable indirectly represents or correlates with a protected characteristic.
For example, a model may not use race directly.
Instead, it may use a combination of:
- Geographic information.
- Consumer behaviour.
- Household characteristics.
- Purchasing patterns.
Those variables may collectively correlate with race or another protected characteristic.
The algorithm can therefore produce different outcomes without explicitly receiving the protected variable.
Is Proxy Discrimination Illegal in Insurance?
Quick Answer: Potentially. Whether a particular proxy constitutes unlawful discrimination depends on the applicable state law, insurance product, classification and evidence.
It is therefore incorrect to state that every correlation is automatically unlawful.
But it is equally incorrect to assume that deleting the protected variable from a dataset eliminates discrimination risk.
Can ZIP Codes Create Insurance Discrimination?
Quick Answer: Geographic information can be a legitimate insurance risk variable in some circumstances, but it can also create concerns if it functions as a proxy for protected characteristics or produces impermissible discriminatory outcomes.
Location can legitimately correlate with:
- Weather risk.
- Crime.
- Accident frequency.
- Property values.
- Natural catastrophe exposure.
But geography can also correlate with demographic characteristics.
The legal analysis therefore requires more than asking whether location predicts losses.
Can Credit-Related Data Create AI Insurance Discrimination?
Quick Answer: Credit-related information can raise legal and fairness questions depending on the insurance product, jurisdiction and permitted use of the data.
AI can amplify these concerns because a model may combine credit-related information with many other variables.
The resulting system can become difficult to assess using traditional methods.
Can Social Media Data Cause Insurance Discrimination?
Quick Answer: Potentially. Social-media or online behavioural data can create questions concerning relevance, accuracy, privacy and discriminatory effects.
For example, an insurer might use an external behavioural dataset to predict risk.
The model may discover that certain behavioural characteristics correlate with claims.
But correlation alone does not necessarily establish that the variable is legally appropriate for insurance pricing or underwriting.
Can AI Discriminate in Insurance Underwriting?
Quick Answer: Yes. AI underwriting can create discriminatory outcomes through data selection, model design or proxy variables.
Potential issues include:
- Risk classification.
- Eligibility decisions.
- Accelerated underwriting.
- External data.
- Predictive models.
The NAIC's accelerated-underwriting guidance specifically identifies potential unfair discrimination as a regulatory consideration when insurers use external data, predictive models and algorithmic or machine-learning techniques. ([content.naic.org](https://content.naic.org/insurance-topics/accelerated-underwriting?utm_source=chatgpt.com))
Can AI Discriminate in Insurance Pricing?
Quick Answer: Yes, potentially.
AI pricing models may analyse large numbers of variables to predict expected losses.
The model may therefore produce highly differentiated prices.
The legal question is whether those classifications are permissible under the applicable insurance framework.
What Is Actuarial Justification?
Quick Answer: Actuarial justification generally concerns whether an insurance classification has a legitimate relationship to expected risk and is supported by appropriate actuarial analysis.
The concept is important because insurance cannot function if every difference between consumers is treated as discrimination.
Risk differentiation is fundamental to insurance.
The regulatory challenge is identifying when differentiation becomes unfair discrimination.
Is Actuarial Fairness the Same as Legal Fairness?
Quick Answer: No.
A model may be actuarially predictive while still raising broader legal or fairness concerns.
For example, a variable might statistically improve prediction but also create a problematic relationship with a protected characteristic.
Therefore:
Predictive value is not necessarily the same thing as legal permissibility.
Can Historical Data Make AI Discriminatory?
Quick Answer: Yes.
Historical insurance data reflects historical underwriting and pricing practices.
If those historical practices contained bias, the data may encode those patterns.
A machine-learning model trained on the data may then learn them.
This produces an important AI governance problem:
Historical data is not automatically neutral simply because it is historical.
What Is Data Bias in Insurance AI?
Quick Answer: Data bias occurs when the data used by an AI system does not appropriately represent the population, phenomenon or risk the model is intended to evaluate.
Examples include:
- Incomplete datasets.
- Historical bias.
- Measurement errors.
- Sampling problems.
- Incorrect labels.
- Missing information.
Can Poor Data Cause Discrimination?
Quick Answer: Yes.
Suppose an insurer's historical claims dataset disproportionately represents one group.
The AI model learns from that data.
The resulting predictions may be less reliable for other groups.
This can create unequal outcomes even when the model itself contains no explicit discriminatory instruction.
What Is Disparate Impact in AI Insurance?
Quick Answer: Disparate impact generally refers to a facially neutral practice producing disproportionately adverse outcomes for a particular group.
The legal significance of disparate impact varies depending on the applicable law.
Insurance regulators and policymakers have nevertheless paid increasing attention to outcome-based analysis of AI systems.
Does AI Insurance Discrimination Have to Be Intentional?
Quick Answer: Not necessarily.
The legal treatment of discriminatory outcomes depends on the specific law involved.
From an AI governance perspective, however, unintentional discrimination can still create significant regulatory and consumer-protection risk.
The NAIC's AI framework emphasises fair and accurate outcomes and compliance with applicable unfair trade-practice laws. ([content.naic.org](https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers?utm_source=chatgpt.com))
Can an AI Model Be Fair If Its Outcome Is Unequal?
Quick Answer: Unequal outcomes do not automatically establish unlawful discrimination.
Insurance inherently produces different outcomes because risk differs.
The important questions include:
- Why are the outcomes different?
- Is the classification legally permitted?
- Is it actuarially justified?
- Is the data appropriate?
- Does the model use a prohibited proxy?
- Does applicable law prohibit the practice?
Can AI Discriminate in Insurance Claims?
Quick Answer: Potentially. AI claims systems can create discriminatory risks if certain groups are disproportionately flagged, delayed, investigated or denied without lawful justification.
For example, a fraud-detection model might identify a particular pattern as suspicious.
If the pattern correlates with a protected group, regulators may need to examine whether the system produces unfair outcomes.
Can AI Discriminate in Insurance Marketing?
Quick Answer: Potentially.
AI can be used to target advertisements and identify potential customers.
This creates questions about:
- Who receives insurance offers.
- Who is excluded from marketing.
- Which customers are targeted with particular products.
- Whether certain groups receive systematically different opportunities.
AI discrimination therefore does not begin only when a premium is calculated.
Can AI Discriminate in Insurance Eligibility?
Quick Answer: Potentially.
If an algorithm determines whether an applicant receives a particular product or underwriting pathway, discriminatory model outputs can affect access to insurance.
This is especially important where automated systems determine:
- Acceptance.
- Rejection.
- Referral.
- Additional underwriting requirements.
What Is a Protected Characteristic?
Quick Answer: A protected characteristic is a characteristic that receives legal protection under applicable anti-discrimination or insurance law.
The precise list varies by jurisdiction.
Examples may include:
- Race.
- Ethnicity.
- National origin.
- Sex.
- Other legally protected characteristics.
Insurers should therefore conduct state-specific legal analysis rather than assuming a single nationwide list applies identically to every insurance product.
Why Is Race Particularly Important in AI Insurance?
Quick Answer: Race has a long and complex history in insurance regulation and risk classification.
The NAIC's Race and Insurance materials explain that risk-based differentiation is legitimate, but unfair discrimination based on race, ethnicity or national origin is not. ([content.naic.org](https://content.naic.org/insurance-topics/race-and-insurance?utm_source=chatgpt.com))
This history makes algorithmic use of demographic or geographic data particularly sensitive.
Can an Insurer Simply Remove Race From Its AI Model?
Quick Answer: Removing race from the input variables does not necessarily eliminate discrimination risk.
Other variables may operate as proxies.
For example:
Race removed → geography retained → geography correlates with race → model learns geographic pattern.
The model can therefore produce differentiated outcomes without explicitly using race.
How Can Insurers Test AI for Discrimination?
Quick Answer: Insurers can use a combination of statistical, actuarial, legal and governance testing.
Potential approaches include:
- Outcome testing.
- Variable analysis.
- Proxy analysis.
- Performance testing across groups.
- Historical-data review.
- Scenario testing.
- Human review.
The appropriate methodology depends on the model and applicable law.
What Is Fairness Testing?
Quick Answer: Fairness testing evaluates whether an AI system produces materially different or potentially problematic outcomes across relevant groups.
Testing may examine:
- Approval rates.
- Premium differences.
- Referral rates.
- Claim investigation rates.
- Claim-denial rates.
But statistical differences do not automatically establish unlawful discrimination.
They are signals requiring further legal and actuarial analysis.
Should Insurers Test for Proxy Variables?
Quick Answer: Proxy analysis can be an important part of AI governance because removing an explicitly protected variable does not eliminate indirect relationships.
Insurers should understand which variables:
- Are strongly correlated.
- Drive model outputs.
- Change risk classifications.
- Produce materially different outcomes.
What Is Outcome Testing?
Quick Answer: Outcome testing examines what actually happens after an AI system is deployed.
For example:
Input → Model → Premium → Consumer outcome
Instead of examining only the model's design, regulators can examine the resulting consumer outcomes.
The NAIC has emphasised the importance of responsible AI governance and fair, accurate consumer outcomes. ([content.naic.org](https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers?utm_source=chatgpt.com))
Why Is Outcome Testing Important?
Quick Answer: A model can appear neutral on paper while producing problematic outcomes in practice.
For example, an insurer may say:
“Our model does not use race.”
But the regulator may ask:
“What outcomes does the model produce?”
That is a fundamentally different question.
Can AI Discrimination Be Caused by Third-Party Vendors?
Quick Answer: Yes.
An insurer may obtain:
- External consumer data.
- Predictive models.
- Risk scores.
- AI software.
from third-party providers.
If the external system creates discriminatory outcomes, the insurer may still face regulatory questions concerning how it selected, validated and governed the technology.
Can an Insurer Blame the AI Vendor for Discrimination?
Quick Answer: Not automatically.
The NAIC states that existing insurance laws apply regardless of whether decisions are made by humans, algorithms or third-party vendors. ([content.naic.org](https://content.naic.org/sites/default/files/ai-issue-brief.pdf?utm_source=chatgpt.com))
This means vendor outsourcing does not automatically become regulatory outsourcing.
What Should AI Insurance Vendor Due Diligence Include?
Quick Answer: Insurers should investigate the data, methodology, validation, governance and performance of material third-party models.
Questions can include:
- What data is used?
- Where did the data originate?
- How is the model validated?
- How is bias assessed?
- How are material changes communicated?
- Can the insurer audit the system?
- Can regulators obtain relevant information?
What Does the NAIC Say About AI Discrimination?
Quick Answer: The NAIC has repeatedly identified fairness and unfair discrimination as important considerations in insurance AI.
The NAIC adopted its AI Principles in 2020 and its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023. The bulletin states that AI-supported consumer decisions must comply with applicable insurance laws and regulations and establishes expectations concerning governance, risk management, testing and documentation. ([content.naic.org](https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers?utm_source=chatgpt.com))
Is the NAIC AI Model Bulletin a Law?
Quick Answer: No. The Model Bulletin is not itself a model law or regulation.
The NAIC explains that it functions as guidance for state insurance regulators and establishes expectations concerning responsible AI use. ([content.naic.org](https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers?utm_source=chatgpt.com))
Its legal effect therefore depends on how states adopt or use the guidance.
How Are Regulators Approaching AI Insurance Discrimination in 2026?
Quick Answer: Regulators are moving from broad principles toward implementation, examination and model testing.
The NAIC's 2026 materials state that regulators are developing methods to review AI systems and test models for accuracy, fairness and potential bias. ([content.naic.org](https://content.naic.org/insurance-topics/insurtech?utm_source=chatgpt.com))
The NAIC has also described 2026 as a phase of moving from principles and guidance toward implementation and supervisory scrutiny. ([content.naic.org](https://content.naic.org/article/evolving-marketplace-continued-state-leadership-naic-president-white-2026-spring-national-meeting?utm_source=chatgpt.com))
Does AI Have to Be Perfectly Fair?
Quick Answer: No.
There is no simple mathematical definition of “perfect fairness” that automatically resolves every insurance decision.
Insurance requires risk differentiation.
Different models can produce different statistical outcomes.
The legal question is therefore contextual:
Is the classification permissible, appropriately supported and consistent with applicable insurance law?
AI Insurance Discrimination Risk Matrix
| Source of Risk | Example | Potential Control |
|---|---|---|
| Historical data | Past discriminatory patterns | Data review |
| Proxy variable | Geography correlating with protected traits | Proxy analysis |
| External data | Incorrect consumer information | Data validation |
| Model design | Problematic target variable | Model governance |
| Pricing | Unequal premium outcomes | Outcome testing |
| Underwriting | Unequal eligibility | Fairness testing |
| Claims | Unequal fraud referrals | Monitoring |
| Third-party vendor | Opaque external score | Vendor due diligence |
AI Insurance Discrimination Compliance Checklist
- Identify AI systems affecting consumers.
- Identify every material data source.
- Document the purpose of each model.
- Identify protected characteristics relevant to the applicable legal framework.
- Analyse potential proxy variables.
- Review historical data for material bias.
- Validate predictive performance.
- Conduct appropriate fairness and outcome testing.
- Assess whether classifications are actuarially justified where required.
- Monitor post-deployment outcomes.
- Review third-party data and model vendors.
- Document material model changes.
- Establish escalation procedures for identified discrimination risks.
- Maintain documentation for regulatory examination.
Frequently Asked Questions
Can AI discriminate in insurance?
Yes. AI can produce unlawful or unfair discriminatory outcomes through data, model design, proxy variables or implementation.
Is all AI insurance discrimination intentional?
No. Discriminatory outcomes can potentially arise from historical data, proxy variables or model behaviour without an explicit instruction to discriminate.
What is proxy discrimination in insurance?
Proxy discrimination occurs when apparently neutral variables indirectly correlate with protected characteristics and contribute to discriminatory outcomes.
Can ZIP codes be discriminatory in insurance?
Geographic information can be relevant to insurance risk, but it can also create proxy-discrimination concerns depending on how it is used and the applicable law.
Can AI discriminate in insurance pricing?
Potentially. AI pricing models can produce materially different premiums, which must be assessed under the applicable insurance regulatory framework.
Can AI discriminate in underwriting?
Yes. AI underwriting can create discrimination risks through predictive models, external data and proxy variables.
Can AI discriminate in claims?
Potentially. Claims models can create unequal investigation, referral, delay or denial outcomes.
Is actuarial fairness enough to make an AI model lawful?
No. Predictive or actuarial validity does not necessarily resolve every legal or consumer-protection issue.
Can an insurer simply remove race from its AI model?
No. Other variables may operate as proxies for race or other protected characteristics.
Can historical data make AI discriminatory?
Yes. Historical data can encode historical patterns of unequal treatment or structural disadvantage.
Can third-party AI vendors create discrimination risk?
Yes. External data and models can introduce additional bias, accuracy and governance risks.
Can insurers blame AI vendors for discrimination?
Not automatically. Insurers remain responsible for complying with applicable insurance laws when using AI systems and third-party technologies.
Does the NAIC regulate AI insurance discrimination?
The NAIC develops principles, model guidance and regulatory tools that support state insurance regulators. Binding requirements depend on applicable state law and regulatory action.
What is fairness testing in AI insurance?
Fairness testing examines whether an AI system produces materially different or potentially problematic outcomes across relevant groups.
What is outcome testing?
Outcome testing examines the real-world results produced by an AI system rather than looking only at its design.
Conclusion
Artificial intelligence creates a difficult paradox for insurance.
Insurance requires differentiation.
An insurer cannot charge every customer the same premium regardless of risk.
Risk classification is fundamental to the insurance business.
But not every method of differentiating between consumers is legally permissible.
This is where AI creates a new layer of complexity.
An algorithm may analyse thousands of variables.
It may discover relationships that human underwriters never intentionally programmed.
It may identify highly predictive correlations.
And those correlations may sometimes interact with protected characteristics in ways that are difficult to detect.
The critical mistake would be to assume:
“The algorithm did not use race, therefore the algorithm cannot discriminate.”
That is not necessarily true.
Geography can act as a proxy.
Consumer behaviour can act as a proxy.
Historical data can encode historical inequalities.
External datasets can introduce hidden assumptions.
And a model can produce unequal outcomes even when no developer intended discrimination.
This is why AI insurance governance must look beyond the model's technical accuracy.
The NAIC's current AI framework emphasises fairness, accuracy, accountability, transparency and compliance with applicable insurance law. ([content.naic.org](https://content.naic.org/insurance-topics/big-data?field_committee_term_target_id=All&utm_source=chatgpt.com))
The NAIC has also specifically identified unfair discrimination as a concern in accelerated underwriting involving external data and predictive models. ([content.naic.org](https://content.naic.org/insurance-topics/accelerated-underwriting?utm_source=chatgpt.com))
And in 2026, regulators are increasingly moving toward actual examination and model-testing capabilities, including assessment of fairness and potential bias. ([content.naic.org](https://content.naic.org/insurance-topics/insurtech?utm_source=chatgpt.com))
The emerging regulatory approach can therefore be summarised as:
Do not ask only what the algorithm was designed to do. Ask what the algorithm actually does.
For insurers, that means:
- Review the data.
- Understand the variables.
- Identify proxies.
- Test outcomes.
- Validate models.
- Monitor performance.
- Govern third-party vendors.
- Document decisions.
For regulators, it means increasingly looking at the relationship between:
Data → Model → Risk Classification → Insurance Decision → Consumer Outcome.
And for consumers, the most important point is this:
An AI-generated insurance decision is not automatically lawful merely because it was generated by a sophisticated algorithm.
The technology may be new.
The legal responsibility is not.
Legal Disclaimer
This article is provided for general educational and informational purposes only. It is not legal, insurance, financial, actuarial or regulatory advice and does not create an attorney-client relationship. Insurance discrimination law varies by state, insurance product and individual circumstances.