AI Insurance Underwriting Bias: Can Algorithms Discriminate Against Policyholders?
Quick Answer: Yes. AI underwriting systems can produce discriminatory outcomes even when developers do not explicitly instruct the algorithm to discriminate. Bias can enter through historical data, proxy variables, model design, geographic information, behavioural data or other features. Whether a particular outcome is unlawful depends on the applicable insurance, anti-discrimination and consumer-protection framework.
Insurance underwriting has always involved classification.
An insurer asks:
How risky is this applicant?
Traditionally, underwriters relied on actuarial information, historical claims experience, applicant information and other risk indicators.
Artificial intelligence expands the amount and complexity of information that can potentially be analysed.
An AI underwriting system may evaluate hundreds of variables simultaneously.
That can produce significant benefits.
It can identify correlations that traditional underwriting methods might miss.
It can process applications faster.
It can improve consistency.
It can potentially improve risk prediction.
But the same capability creates a fundamental legal question:
What happens when the algorithm's correlations reproduce or create discriminatory outcomes?
This is where AI underwriting bias becomes important.
The algorithm may never receive an instruction saying:
โTreat Group A differently.โ
Instead, it may learn from historical data that certain characteristics are associated with particular outcomes.
Some of those characteristics may be legitimate risk indicators.
Others may function as proxies for protected characteristics or otherwise produce legally problematic outcomes.
The result can be:
Neutral-looking input โ algorithmic correlation โ unequal outcome.
That is why AI underwriting cannot be evaluated solely by asking whether the code contains discriminatory instructions.
The outcomes and underlying data also matter.
Legal disclaimer: This article provides general educational information and is not legal, insurance, actuarial, financial or regulatory advice. Anti-discrimination and insurance requirements vary by product, jurisdiction and factual circumstances.
Key Takeaways
- AI underwriting can create bias even without explicit discriminatory instructions.
- Historical underwriting data can reproduce historical patterns.
- Proxy variables can indirectly correlate with protected characteristics.
- Geographic and behavioural data require careful assessment.
- Different outcomes do not automatically establish unlawful discrimination.
- AI systems should be evaluated using both technical and legal criteria.
- Fairness testing should examine outcomes across relevant consumer groups.
- Model accuracy alone does not establish fairness.
- Human oversight can provide an important safeguard.
- Insurers should document model purpose, data sources and validation.
- Third-party AI underwriting systems require appropriate vendor oversight.
- State insurance regulation remains particularly important in the United States.
What Is AI Insurance Underwriting?
Quick Answer: AI insurance underwriting is the use of artificial intelligence, machine learning or related computational techniques to assist or automate the assessment of insurance risk.
AI can potentially assist with:
- Risk classification.
- Application screening.
- Risk scoring.
- Document analysis.
- Fraud indicators.
- Claims-history analysis.
- Pricing inputs.
What Is AI Underwriting Bias?
Quick Answer: AI underwriting bias occurs when an AI system systematically produces problematic or unequal outcomes because of characteristics of its data, variables, model design or deployment.
Bias can occur at several stages:
Data โ Model โ Decision โ Outcome.
Each stage requires examination.
Can an AI Algorithm Be Biased Without Intending to Be?
Quick Answer: Yes.
AI systems learn patterns from data.
If the data contains historical inequalities or problematic correlations, the model can reproduce them.
This is sometimes described as algorithmic bias.
How Does Historical Data Create AI Bias?
Quick Answer: Historical datasets can contain patterns generated by previous human decisions, institutional practices or unequal access to opportunities.
Suppose an insurer's historical data shows:
Variable X โ historically associated with higher claim costs.
The AI system may learn that Variable X is highly predictive.
But the insurer must still ask:
Why was Variable X associated with higher claims?
Correlation does not automatically establish that the variable is legally or actuarially appropriate.
What Is Proxy Discrimination in AI Underwriting?
Quick Answer: Proxy discrimination occurs when a variable indirectly captures information associated with a protected characteristic or otherwise contributes to a discriminatory outcome.
A model might not use a protected characteristic directly.
Instead:
Protected characteristic โ correlated variable โ AI model โ outcome.
The indirect relationship can still require scrutiny.
What Are Examples of Potential Proxy Variables?
Quick Answer: Potential proxy variables can include geographic, behavioural, economic or demographic characteristics that correlate with other characteristics.
Examples may include:
- Location.
- Purchasing behaviour.
- Digital activity.
- Economic indicators.
- Network characteristics.
A variable is not automatically unlawful simply because it correlates with a protected characteristic.
The legal question depends on the applicable law and use.
Can Location Data Create Insurance Bias?
Quick Answer: Potentially.
Geographic information can be actuarially relevant for many insurance products.
For example, location can correlate with:
- Weather risk.
- Crime rates.
- Traffic patterns.
- Property characteristics.
- Healthcare availability.
But location can also correlate with demographic and socioeconomic characteristics.
This creates a difficult regulatory question:
When is geographic information legitimate risk information, and when can its use produce impermissible discrimination?
Can AI Use Credit-Related Data in Insurance Underwriting?
Quick Answer: The use of credit-related information in insurance varies according to state law, insurance product and applicable regulatory requirements.
Where such information is permitted, insurers should still assess:
- Data accuracy.
- Relevance.
- Consumer impact.
- Applicable legal restrictions.
Does AI Discrimination Require Intent?
Quick Answer: Not necessarily. The applicable legal standard depends on the specific discrimination claim and governing law. Some legal frameworks focus on discriminatory treatment or outcomes rather than requiring proof that an algorithm was deliberately designed to discriminate.
This is one reason insurers should not rely solely on the defence:
โWe never told the AI to discriminate.โ
Is Every Difference in Insurance Treatment Discrimination?
Quick Answer: No.
Insurance is based on risk classification.
Two consumers can legitimately receive different:
- Premiums.
- Coverage terms.
- Underwriting decisions.
The critical question is whether the difference is legally permissible and appropriately related to the relevant insurance risk.
What Is the Difference Between Risk Classification and Discrimination?
Quick Answer: Risk classification seeks to distinguish consumers based on legally permissible and actuarially relevant differences, while unlawful discrimination involves treatment prohibited by applicable law.
Therefore:
Different treatment โ automatically discriminatory treatment.
Can AI Make Underwriting More Fair?
Quick Answer: Potentially.
AI is not inherently discriminatory.
A properly designed system may:
- Reduce inconsistent human decisions.
- Identify relevant risk factors.
- Improve statistical accuracy.
- Apply documented criteria consistently.
The important issue is governance.
Can AI Also Make Discrimination Worse?
Quick Answer: Yes.
Automation can scale a problematic rule across a large population.
Consider:
Human bias โ 100 decisions.
versus:
AI bias โ 1,000,000 decisions.
The second problem can have dramatically greater consequences.
What Is Disparate Impact?
Quick Answer: Disparate impact generally refers to a situation where a seemingly neutral practice disproportionately affects a particular group, subject to the specific legal framework governing the claim.
Whether disparate impact is legally actionable in a particular insurance context requires jurisdiction-specific analysis.
What Is Disparate Treatment?
Quick Answer: Disparate treatment generally involves treating similarly situated individuals differently because of a protected characteristic or other legally prohibited basis.
Again, the precise legal standard depends on the applicable law.
How Can Insurers Test AI for Bias?
Quick Answer: Insurers can use statistical and governance techniques to examine whether model outcomes differ materially across relevant groups.
Testing may examine:
- Approval rates.
- Decline rates.
- Premium outcomes.
- Coverage outcomes.
- Referral rates.
- Error rates.
What Is Fairness Testing?
Quick Answer: Fairness testing evaluates whether an AI model produces materially different outcomes across relevant groups and whether those differences require investigation under applicable legal or organisational standards.
Fairness testing should not be treated as a single mathematical test.
Different fairness metrics can produce different results.
Why Is Model Accuracy Not Enough?
Quick Answer: A model can be highly accurate overall while producing problematic outcomes for a particular population.
Example:
Overall accuracy = 96%.
That figure does not tell us whether:
Group A accuracy = 98%
while:
Group B accuracy = 82%.
Therefore, aggregate performance can conceal subgroup performance.
What Is Subgroup Testing?
Quick Answer: Subgroup testing examines model performance and outcomes across relevant consumer groups.
This can help identify:
- Unequal error rates.
- Unequal approval rates.
- Unequal referral rates.
- Unexpected outcome differences.
Should Insurers Remove All Correlated Variables?
Quick Answer: Not necessarily.
Correlation alone does not establish that a variable should be removed.
Some correlations may reflect genuine insurance risk.
The better approach is to assess:
- Predictive relevance.
- Actuarial justification.
- Legal permissibility.
- Consumer impact.
What Is Actuarial Relevance?
Quick Answer: Actuarial relevance concerns whether information meaningfully contributes to the assessment of insurance risk under the applicable actuarial and regulatory framework.
A variable can be statistically predictive without automatically being legally appropriate.
That distinction is critical.
Can AI Underwriting Use Behavioural Data?
Quick Answer: Behavioural data can potentially be used in some insurance contexts, but insurers must consider the product, jurisdiction, data source, consumer expectations, privacy implications and applicable legal restrictions.
Examples can include:
- Driving behaviour.
- Transaction patterns.
- Digital interactions.
- Application behaviour.
What Is Black-Box Underwriting?
Quick Answer: Black-box underwriting refers to situations where an AI model's internal decision-making process is difficult for users or reviewers to understand.
This can create governance problems.
An underwriter may know:
โThe system rejected the applicant.โ
but not:
โWhy?โ
Why Is Explainability Important in Underwriting?
Quick Answer: Explainability can help insurers understand model behaviour, identify errors and provide appropriate explanations where required.
It can also support:
- Internal audit.
- Regulatory review.
- Consumer complaints.
- Model validation.
Can Human Underwriters Override AI?
Quick Answer: Where appropriate, human underwriters can provide an important safeguard by reviewing unusual or disputed AI recommendations.
But an override system is meaningful only if humans can genuinely question the AI output.
What Is Automation Bias in Underwriting?
Quick Answer: Automation bias occurs when human decision-makers place excessive confidence in automated recommendations.
For example:
AI: High risk.
Underwriter: Automatically accepts the classification.
That is not necessarily meaningful human oversight.
Can Third-Party AI Underwriting Vendors Create Liability?
Quick Answer: Potentially.
Insurers should conduct appropriate due diligence concerning:
- Training data.
- Model methodology.
- Validation.
- Bias testing.
- Model updates.
- Security.
A vendor's claim that its model is โAI-poweredโ should not substitute for governance.
What Should an AI Underwriting Vendor Contract Include?
Quick Answer: Material AI contracts should address relevant responsibilities and risks.
Potential provisions include:
- Model documentation.
- Performance standards.
- Audit rights.
- Data restrictions.
- Security.
- Incident reporting.
- Material model-change notification.
- Indemnification.
AI Underwriting Bias Risk Matrix
| Risk | Example | Potential Control |
|---|---|---|
| Historical bias | Model learns problematic past decisions | Training-data assessment |
| Proxy discrimination | Variable correlates with protected characteristic | Proxy analysis |
| Geographic bias | Location produces disproportionate outcomes | Outcome testing |
| Data error | Incorrect consumer information | Data correction process |
| Model opacity | Underwriter cannot understand decision | Explainability |
| Automation bias | Human blindly accepts model | Independent review |
| Model drift | Outcome patterns change over time | Continuous monitoring |
| Vendor risk | External model not adequately assessed | Vendor governance |
AI Insurance Underwriting Bias Compliance Checklist
- Identify every AI underwriting system.
- Define the purpose of each model.
- Identify all input variables.
- Document data sources.
- Assess historical data quality.
- Identify potential proxy variables.
- Assess legal restrictions on relevant variables.
- Conduct fairness testing.
- Conduct subgroup performance testing.
- Validate actuarial relevance.
- Document model methodology.
- Establish human-review procedures.
- Monitor model outcomes.
- Monitor model drift.
- Review consumer complaints.
- Review third-party vendors.
- Maintain model-version records.
- Document material changes.
- Periodically reassess discrimination risks.
- Retire or modify models that create unacceptable risk.
Frequently Asked Questions
Can AI underwriting discriminate?
Yes. AI can produce discriminatory outcomes through historical data, proxy variables, model design or other mechanisms.
Can AI be biased without using race or another protected characteristic?
Yes. Other variables can potentially act as proxies for protected characteristics or otherwise produce unequal outcomes.
Is every different insurance premium discriminatory?
No. Insurance inherently involves risk classification. Different treatment is not automatically unlawful discrimination.
What is proxy discrimination in insurance?
Proxy discrimination occurs when a variable indirectly captures information associated with another characteristic and contributes to an impermissible discriminatory outcome.
Can location be an insurance AI proxy?
Potentially. Location can reflect legitimate insurance risk but may also correlate with demographic or socioeconomic characteristics. Its legal treatment depends on the context and applicable law.
Can insurers use AI for underwriting?
Yes, subject to applicable insurance, anti-discrimination, consumer-protection and other legal requirements.
Does model accuracy prove an AI underwriting system is fair?
No. Overall accuracy can conceal significant differences in performance or outcomes among consumer groups.
What is fairness testing?
Fairness testing examines whether an AI system produces materially different outcomes or error rates across relevant groups.
Should humans review AI underwriting decisions?
Human oversight can be an important safeguard, particularly for unusual, disputed or consequential decisions.
Can an AI vendor be responsible for discriminatory underwriting?
Potentially. Responsibility depends on the vendor's role, contractual obligations, applicable law and the facts of the case.
Why is historical underwriting data important?
Historical data can contain patterns generated by earlier practices. AI models can reproduce those patterns if they are not properly assessed.
What should insurers do if an AI model shows bias?
The insurer should investigate the source of the bias, assess the legal and consumer impact, consider mitigation and document the corrective action.
Conclusion
Artificial intelligence is transforming insurance underwriting.
The technology can evaluate enormous quantities of information and identify relationships that traditional systems may not detect.
That capability can improve risk assessment.
But it also creates a fundamental governance challenge.
The more variables an AI system can analyse, the more opportunities there are for problematic correlations to influence its decisions.
The central mistake would be to assume that an algorithm is neutral simply because it is mathematical.
Mathematics can reproduce the assumptions contained in its data.
A model can therefore be technically sophisticated and still produce problematic outcomes.
This is particularly important in insurance because underwriting decisions can affect whether consumers obtain coverage and how much they pay for it.
But there is another important point.
Different treatment does not automatically equal discrimination.
Insurance depends on risk classification.
An insurer may legitimately distinguish between different levels of risk where permitted by applicable law.
The relevant question is not simply:
โDid the AI treat consumers differently?โ
It is:
โWhy did the AI treat them differently, and is that distinction legally permissible?โ
This requires insurers to examine the entire model.
They should understand:
- Where the data came from.
- What variables are used.
- Why those variables are predictive.
- Whether proxies exist.
- How outcomes differ across groups.
- Whether the model remains appropriate over time.
Fairness testing should therefore become part of ordinary model governance.
It should not happen only after a regulator or consumer raises a complaint.
The same principle applies to third-party AI systems.
An insurer purchasing an underwriting model from a vendor should not simply assume that the vendor's model is compliant.
The insurer should ask:
What was the model trained on?
What variables does it use?
How was it validated?
How was bias assessed?
How does the vendor handle model changes?
These are basic governance questions.
Another important safeguard is meaningful human oversight.
Humans should be capable of challenging the algorithm.
If an underwriter cannot question a model recommendation, human involvement may become little more than a procedural formality.
The better framework is:
AI recommends โ Human evaluates โ Evidence supports โ Decision is documented.
AI should improve underwriting judgment, not eliminate accountability.
The future of AI underwriting will likely involve increasingly sophisticated data sources, including behavioural, geographic and other non-traditional information.
That makes governance even more important.
Insurers should distinguish carefully between:
Prediction
and:
Permission.
A variable may predict insurance risk.
That does not automatically mean the insurer is legally permitted to use it in every context.
This distinction is one of the most important principles in AI insurance law.
The central lesson is therefore:
An AI underwriting system should not be judged only by how accurately it predicts risk. It should also be evaluated for how it obtains that prediction, how it affects consumers and whether its use complies with applicable law.
Responsible AI underwriting ultimately requires three things:
Good data.
Good models.
Good governance.
Without all three, algorithmic efficiency can become algorithmic unfairness.
Legal Disclaimer
This article is provided for general educational and informational purposes only. It is not legal, insurance, actuarial, financial, privacy or regulatory advice and does not create an attorney-client relationship. Insurance underwriting and anti-discrimination requirements vary according to jurisdiction, insurance product, model design and specific facts.