AI Health Insurance Discrimination: Can Algorithms Unfairly Deny Coverage or Increase Costs?
Quick Answer: Yes. Artificial intelligence can potentially produce discriminatory outcomes in health insurance even when an algorithm does not explicitly use a protected characteristic. Bias can enter through historical data, proxy variables, healthcare-access differences, model design, inaccurate information or the way an insurer defines its prediction target. Whether a particular outcome constitutes unlawful discrimination depends on the applicable federal and state law, the insurance product and the facts of the case.
Imagine two patients.
They have similar medical conditions.
They need similar treatment.
But an algorithm assigns them different risk scores.
One receives favourable treatment.
The other faces a higher premium, additional scrutiny or a coverage restriction.
The insurer says:
βThe algorithm never considered race, sex, disability or any other protected characteristic.β
Does that end the legal inquiry?
No.
An algorithm can produce discriminatory outcomes without directly asking for a protected characteristic.
It may instead use variables that are correlated with that characteristic.
It may learn from historical data containing existing inequalities.
It may rely on healthcare utilisation as a proxy for need even though utilisation also reflects access to care.
It may use geographic information that indirectly correlates with socioeconomic or demographic characteristics.
It may learn patterns from previous insurance decisions and reproduce them at scale.
This creates one of the central legal challenges of artificial intelligence in health insurance:
When does predictive risk assessment become unlawful discrimination?
The answer is not simply a technical question.
It involves civil-rights law, insurance law, healthcare regulation, actuarial methodology, data governance and consumer protection.
The issue is especially important because health insurance decisions can affect:
- Coverage.
- Premiums.
- Benefits.
- Prior authorisation.
- Medical-necessity review.
- Claims.
- Access to healthcare.
Legal disclaimer: This article provides general educational information and is not legal, medical, insurance, actuarial, financial or regulatory advice. The applicable legal framework varies according to the health plan, insurance product, jurisdiction and individual circumstances.
Key Takeaways
- AI can potentially produce discriminatory health-insurance outcomes.
- An algorithm does not have to explicitly use race, sex or another protected characteristic to create discriminatory effects.
- Proxy variables can indirectly encode protected characteristics.
- Historical healthcare data can reproduce historical inequalities.
- Healthcare utilisation is not necessarily the same as healthcare need.
- AI discrimination can arise in underwriting, pricing, claims, fraud detection and prior authorisation.
- Health-risk algorithms can be particularly sensitive because their target variables may reflect unequal access to healthcare.
- Section 1557 of the Affordable Care Act prohibits discrimination on specified grounds in covered health programmes and activities.
- The precise scope of Section 1557 and its implementing regulations has changed following litigation and regulatory developments.
- Medicare Advantage plans have additional federal requirements concerning equitable access and discrimination.
- Insurers should test AI systems for disparate outcomes and investigate unexplained differences.
- Removing protected characteristics from a dataset does not automatically eliminate discrimination risk.
- Third-party AI vendors should be included in the insurer's discrimination and model-governance programme.
What Is AI Health Insurance Discrimination?
Quick Answer: AI health insurance discrimination occurs when an artificial-intelligence system or AI-supported process produces unlawful discriminatory treatment or outcomes in a health-insurance context.
Potential applications include:
- Underwriting.
- Pricing.
- Claims.
- Prior authorisation.
- Medical-necessity review.
- Fraud detection.
- Risk scoring.
- Utilisation management.
Not every difference in outcomes is automatically unlawful discrimination.
The legal analysis depends on the applicable law, the protected characteristic, the decision, the justification and the relevant facts.
Does AI Have to Intentionally Discriminate?
Quick Answer: No. Depending on the applicable legal framework, discriminatory liability may arise from intentional discrimination or from other legally recognised forms of discriminatory treatment or effects.
An AI model does not have intentions in the human sense.
The relevant legal question is therefore often about:
What did the system do?
and:
What legal standard applies to that conduct?
What Is Algorithmic Bias?
Quick Answer: Algorithmic bias occurs when an AI system systematically produces results that disadvantage particular individuals or groups because of characteristics embedded in the data, model, design or deployment process.
Bias can enter through:
- Training data.
- Labels.
- Feature selection.
- Proxy variables.
- Model architecture.
- Decision thresholds.
- Human implementation.
Is Algorithmic Bias Always Illegal?
Quick Answer: No.
Algorithmic bias and unlawful discrimination are related but not identical concepts.
A model can exhibit statistical differences without necessarily violating a particular law.
Conversely, an AI system can create legally problematic discrimination even when its overall statistical performance appears strong.
The correct legal analysis must therefore connect:
Model behaviour + protected characteristic + insurance decision + applicable law.
What Is Disparate Treatment?
Quick Answer: Disparate treatment generally involves intentionally treating similarly situated individuals differently because of a protected characteristic.
In an AI context, the challenge is determining whether protected information was:
- Explicitly used.
- Used indirectly.
- Embedded in a rule.
- Reflected through a proxy.
The precise legal standard depends on the applicable law.
What Is Disparate Impact?
Quick Answer: Disparate impact generally concerns a facially neutral practice that disproportionately affects a protected group under a legal framework that recognises effects-based discrimination.
This concept is particularly relevant to algorithmic systems because AI models frequently operate through apparently neutral variables.
However, whether disparate-impact liability applies depends on the particular statute and legal context.
Why Is the Disparate-Impact Question Important in 2026?
Quick Answer: Because the legal treatment of effects-based discrimination varies across federal statutes and has been changing in some regulatory contexts.
For example, HHS announced in July 2026 a final rule changing its Title VI regulations to eliminate disparate-impact liability under those Title VI regulations and return enforcement to intentional discrimination based on race, colour or national origin. :contentReference[oaicite:0]{index=0}
This development should not be generalised to mean that all disparate-impact theories have disappeared from U.S. law.
Different statutes and regulatory regimes can apply different standards.
What Is a Proxy Variable?
Quick Answer: A proxy variable is a seemingly neutral variable that correlates with another characteristic that may be legally or socially significant.
For example, a model may not directly use race.
But it might use:
- Geographic location.
- Healthcare utilisation.
- Provider patterns.
- Economic variables.
If those variables strongly correlate with race or another protected characteristic, they may create discrimination concerns depending on the context.
Can Geographic Location Be a Proxy for Race?
Quick Answer: Geographic variables can correlate with demographic characteristics, including race and ethnicity.
That does not mean geographic information is inherently unlawful.
It means insurers should understand:
Why is the variable being used?
What does it predict?
What consumer outcomes does it produce?
Can Healthcare Utilisation Be a Proxy for Race or Socioeconomic Status?
Quick Answer: Potentially.
Healthcare utilisation can reflect:
- Access to providers.
- Insurance coverage.
- Geographic availability.
- Income.
- Transportation.
- Historical treatment patterns.
Consequently, a model that treats utilisation as a pure measure of healthcare need may encode broader social inequalities.
Can AI Discriminate Even If Protected Characteristics Are Removed?
Quick Answer: Yes, potentially.
Removing explicit demographic variables does not automatically remove correlations from the remaining dataset.
For example:
Race removed β ZIP code retained β ZIP code correlates with race β model continues to learn the relationship.
The same principle can apply to other characteristics.
What Is Fairness Through Unawareness?
Quick Answer: Fairness through unawareness is the idea that discrimination can be reduced simply by excluding protected characteristics from the model.
In practice, this approach is limited.
Why?
Because other variables may function as proxies.
Therefore:
Protected variable removed β discrimination risk eliminated.
Can AI Discriminate Against People With Disabilities?
Quick Answer: Potentially.
Disability can affect healthcare utilisation, treatment history and medical costs.
A predictive model may therefore produce different outcomes for people with disabilities even when disability is not explicitly included as an input.
Whether that outcome violates a particular law depends on the applicable legal framework.
Federal health-program civil-rights protections can be particularly relevant where Section 504 or Section 1557 applies.
Can AI Discriminate Based on Age?
Quick Answer: Potentially.
Age can be a legitimate actuarial or medical variable in certain contexts, but the use of age remains subject to applicable legal restrictions.
Health insurers should therefore distinguish between:
Lawful risk classification
and:
Unlawful discriminatory treatment.
Can AI Discriminate Based on Sex?
Quick Answer: Potentially.
Sex-related variables may correlate with legitimate health risks, but using them in a health-insurance system can implicate applicable anti-discrimination requirements.
Section 1557 prohibits discrimination based on sex in covered health programmes and activities. HHS describes its protections as covering race, colour, national origin, sex, age and disability in applicable health programmes and activities. :contentReference[oaicite:1]{index=1}
Can AI Discriminate Based on Race?
Quick Answer: Potentially.
Race can enter an AI system:
- Directly.
- Through proxy variables.
- Through historical datasets.
- Through labels.
- Through healthcare-access patterns.
The legal analysis depends on the specific health programme and applicable civil-rights law.
What Is Section 1557?
Quick Answer: Section 1557 of the Affordable Care Act is a federal civil-rights provision prohibiting discrimination on specified grounds in covered health programmes and activities.
HHS identifies the protected grounds as including:
- Race.
- Colour.
- National origin.
- Sex.
- Age.
- Disability.
HHS states that Section 1557 can apply to health insurance marketplaces and issuers participating in those marketplaces, among other covered health programmes and activities. :contentReference[oaicite:2]{index=2}
Does Section 1557 Apply to Health Insurance?
Quick Answer: It can apply to certain health insurance programmes and issuers, including covered health programmes and activities within the statute's scope.
The precise coverage must be assessed based on the entity, programme, funding and applicable regulation.
Has Section 1557 Changed in 2026?
Quick Answer: Yes. The regulatory landscape has changed following litigation and subsequent HHS action.
In June 2026, HHS announced that portions of the 2024 Section 1557 final rule had been vacated following Tennessee v. Kennedy, while stating that core protections remained in effect. :contentReference[oaicite:3]{index=3}
In July 2026, HHS announced a separate final rule changing its Title VI regulations to eliminate disparate-impact liability under those regulations. :contentReference[oaicite:4]{index=4}
Accordingly, legal analysis of AI discrimination should be based on the current statute, applicable regulations and judicial decisions rather than assuming that the 2024 regulatory framework remains unchanged.
Does Section 1557 Specifically Regulate AI?
Quick Answer: Section 1557 is a civil-rights provision, not a general AI statute.
The legal principle is nevertheless important:
Using an algorithm does not automatically exempt a covered health programme from applicable anti-discrimination law.
The technology may change.
The underlying legal obligation can remain.
Can Medicare Advantage Plans Discriminate Through AI?
Quick Answer: Medicare Advantage plans remain subject to applicable federal requirements concerning equitable access and discrimination.
CMS has specifically addressed AI and automated systems in its Medicare Advantage regulatory work.
CMS proposed requirements stating that when Medicare Advantage organisations use AI or automated systems, they must comply with applicable requirements and provide equitable access to services without discrimination based on factors related to an enrollee's health status. :contentReference[oaicite:5]{index=5}
This is significant because it makes clear that automation does not place an AI-supported process outside Medicare Advantage regulation.
Can Health-Risk Algorithms Discriminate?
Quick Answer: Yes.
Health-risk algorithms can create discrimination concerns when their target variable or input data reflects unequal healthcare access or historical differences in treatment.
For example:
Healthcare spending β predicted healthcare need
may appear reasonable.
But spending can also depend on:
- Access.
- Provider availability.
- Insurance coverage.
- Geography.
- Socioeconomic circumstances.
Therefore:
Healthcare spending is not always a neutral proxy for healthcare need.
Can AI Underwriting Discriminate?
Quick Answer: Potentially.
AI underwriting can use large datasets to predict risk.
Potential inputs can include:
- Medical information.
- Claims history.
- Demographic information.
- Healthcare utilisation.
- External data.
The legal risk depends on what the model predicts, how the information is used and what laws govern the insurance product.
Can AI Pricing Discriminate?
Quick Answer: Potentially.
Pricing models can create different premiums or benefit structures for different consumers.
Actuarial justification does not automatically answer every civil-rights question.
Insurers should therefore evaluate:
- Variables.
- Data sources.
- Model outputs.
- Protected-group impacts.
- Applicable state and federal law.
Can AI Claims Processing Discriminate?
Quick Answer: Potentially.
A claims algorithm could produce different outcomes because of:
- Incomplete records.
- Historical claims patterns.
- Provider characteristics.
- Geographic variables.
- Proxy variables.
Claims AI should therefore be monitored for systematic differences in outcomes.
Can AI Prior Authorisation Discriminate?
Quick Answer: Yes, potentially.
Prior-authorisation algorithms can discriminate if they:
- Use biased historical data.
- Misclassify particular populations.
- Rely on inappropriate proxies.
- Ignore individual clinical circumstances.
This is particularly important because prior authorisation can affect access to treatment.
Can AI Fraud Detection Discriminate?
Quick Answer: Potentially.
A fraud-detection system may identify certain populations or providers as βhigh riskβ based on historical patterns.
If those patterns reflect biased enforcement or historical disparities, the algorithm may reproduce them.
Therefore:
Fraud prediction should be tested for false-positive disparities.
What Is a False Positive in AI Insurance Discrimination?
Quick Answer: A false positive occurs when the system identifies a legitimate case as suspicious, high-risk or otherwise problematic.
Examples include:
- Legitimate claim flagged as fraud.
- Necessary treatment flagged as unnecessary.
- Low-risk patient classified as high-risk.
False-positive rates should be evaluated across relevant populations.
What Is Disparate Error Rate?
Quick Answer: A disparate error rate occurs when an AI system makes errors at materially different rates for different groups.
For example:
| Group | False Denial Rate |
|---|---|
| Group A | 5% |
| Group B | 14% |
The difference does not automatically prove unlawful discrimination.
But it should trigger investigation.
Should Insurers Test AI for Disparate Outcomes?
Quick Answer: As a governance practice, insurers should evaluate whether material AI systems produce unexplained or potentially problematic differences across relevant populations.
Testing can examine:
- Approval rates.
- Denial rates.
- Premium outcomes.
- Fraud flags.
- Prior-authorisation outcomes.
- Appeal reversals.
- Processing times.
Can Fairness Testing Guarantee Legal Compliance?
Quick Answer: No.
Fairness metrics are technical tools.
Legal discrimination standards depend on the applicable law.
A model can pass one fairness metric and still create legal concerns under another framework.
Therefore:
Fairness testing supports legal compliance; it does not replace legal analysis.
What Is the Problem With a Single Fairness Metric?
Quick Answer: Different fairness metrics can conflict.
For example, an insurer might evaluate:
- Equal approval rates.
- Equal false-positive rates.
- Equal false-negative rates.
- Calibration.
A model can perform well under one metric and poorly under another.
The insurer therefore needs to identify which metric is relevant to the actual legal and business purpose.
Can Insurers Simply Remove Race From AI Models?
Quick Answer: No. Removing race from the input dataset does not necessarily eliminate racial bias.
Other variables may operate as proxies.
In some contexts, protected-characteristic information may also be necessary for testing whether a model produces disparate outcomes.
This creates an important distinction:
Using protected data to discriminate
versus:
Using appropriately governed protected data to test for discrimination.
Why Is This Distinction Important?
Quick Answer: If an insurer never measures demographic outcomes, it may not discover that its model performs differently across populations.
Therefore, governance can require controlled analysis of protected characteristics even where those characteristics are not used as direct decision variables.
Can Third-Party AI Vendors Cause Insurance Discrimination?
Quick Answer: Yes, potentially.
An insurer may purchase a predictive model from a vendor.
The insurer may not have built the model itself.
But the vendor's model can still influence:
- Underwriting.
- Pricing.
- Claims.
- Fraud detection.
- Prior authorisation.
Vendor involvement therefore creates additional governance requirements.
Can Insurers Blame Vendors for Algorithmic Bias?
Quick Answer: Outsourcing a model does not automatically eliminate the insurer's responsibilities.
An insurer should understand:
- What the vendor model predicts.
- What data it uses.
- How it was validated.
- What populations were tested.
- What limitations exist.
The contractual relationship should also address auditability and regulatory cooperation.
What Should an AI Discrimination Audit Examine?
Quick Answer: An AI discrimination audit should examine the model, data, outcomes and decision process.
Key questions include:
- What is the model's purpose?
- What variables does it use?
- Where did the data come from?
- Which variables are potential proxies?
- What are the error rates?
- Do outcomes differ across populations?
- Are differences explainable?
- Are differences legally permissible?
- What corrective action is available?
AI Health Insurance Discrimination Risk Matrix
| AI Application | Potential Discrimination Risk | Governance Response |
|---|---|---|
| Underwriting | Biased risk classification | Model and outcome testing |
| Pricing | Unequal premium outcomes | Actuarial and legal review |
| Prior authorisation | Unequal treatment access | Clinical and fairness review |
| Claims | Different approval/denial rates | Outcome monitoring |
| Fraud detection | Disproportionate false positives | Error-rate analysis |
| Risk scoring | Historical inequality encoded | Target-variable assessment |
| External data | Hidden proxies | Data provenance review |
AI Health Insurance Discrimination Compliance Checklist
- Create an inventory of material AI systems.
- Identify which systems affect consumers.
- Identify all direct model variables.
- Identify potential proxy variables.
- Document data provenance.
- Identify the model's target variable.
- Assess whether the target variable is an appropriate measure of the intended concept.
- Test model accuracy.
- Test false-positive rates.
- Test false-negative rates.
- Evaluate relevant population-level outcomes.
- Investigate unexplained disparities.
- Review applicable federal civil-rights requirements.
- Review applicable state insurance requirements.
- Review Medicare Advantage requirements where applicable.
- Review Medicaid requirements where applicable.
- Review third-party vendor documentation.
- Document corrective actions.
- Monitor the model after deployment.
- Repeat testing when the model or data materially changes.
Frequently Asked Questions
Can AI discriminate in health insurance?
Yes. AI can potentially produce discriminatory outcomes through biased data, proxy variables, model design or deployment.
Does an AI system have to use race to discriminate?
No. Other variables can potentially operate as proxies for race or other characteristics.
What is algorithmic bias in health insurance?
Algorithmic bias occurs when an AI system systematically produces outcomes that disadvantage particular groups because of its data, design or deployment.
Is algorithmic bias automatically illegal?
No. The legal question depends on the applicable statute, regulation, insurance product and facts.
What is disparate impact?
Disparate impact generally concerns a facially neutral practice that disproportionately affects a protected group under a legal framework that recognises effects-based discrimination.
Does disparate impact apply to every health-insurance AI system?
No. The availability and scope of disparate-impact theories depend on the specific legal framework. The law in this area is not uniform across all federal statutes and regulatory programmes.
What is Section 1557?
Section 1557 of the Affordable Care Act prohibits discrimination on specified grounds in covered health programmes and activities.
Does Section 1557 apply to health insurance?
It can apply to certain health insurance programmes and issuers within the statute's scope, including covered marketplace programmes.
Can AI discriminate against people with disabilities?
Potentially. Disability-related healthcare patterns and data can influence algorithmic outputs, creating potential discrimination concerns under applicable civil-rights law.
Can AI discriminate based on age?
Potentially. Age can be relevant to healthcare risk, but its use remains subject to applicable legal restrictions.
Can AI discriminate based on sex?
Potentially. Sex-related variables can create discrimination concerns under applicable health-program civil-rights requirements.
Can removing race from an AI model eliminate discrimination?
No. Proxy variables can continue to reproduce correlations with race or other protected characteristics.
Can health-risk scores be discriminatory?
Potentially. Risk scores can reflect historical healthcare-access disparities, particularly when healthcare spending or utilisation is used as a proxy for healthcare need.
Can AI fraud detection discriminate?
Potentially. Historical fraud investigations and claims data can contain patterns that produce disproportionate false positives.
Should insurers conduct AI fairness testing?
Fairness and outcome testing can be an important component of AI governance, particularly for systems that materially affect consumers.
Can a third-party AI vendor be responsible for discrimination?
Potentially, depending on the contractual relationship, applicable law and facts. Insurers should not assume that outsourcing removes their own regulatory responsibilities.
Does fairness testing guarantee legal compliance?
No. Technical fairness metrics do not replace legal analysis under applicable civil-rights and insurance laws.
Conclusion
Artificial intelligence can make health insurance more predictive.
It can process more data.
It can identify patterns that would be difficult for humans to detect manually.
But predictive power creates a paradox.
The more information an algorithm can analyse, the greater the possibility that it will discover and use correlations that humans did not intend to encode.
Some of those correlations may be useful.
Others may reproduce historical inequality.
Some may operate as proxies for protected characteristics.
Others may simply reflect differences in healthcare access rather than differences in medical need.
This is why AI discrimination cannot be reduced to one question:
βDid the model use race?β
The better questions are:
What data did the model use?
What does each variable represent?
What is the model predicting?
How are the predictions used?
Do outcomes differ across populations?
Why do those differences exist?
Are those differences legally permissible?
This distinction is especially important in health insurance.
A model that predicts healthcare expenditure may appear actuarially sophisticated.
But expenditure can reflect healthcare access.
Healthcare access can reflect socioeconomic circumstances.
Socioeconomic circumstances can correlate with protected characteristics.
The algorithm may therefore reproduce a chain of relationships that was never explicitly programmed into it.
This is the problem of proxy discrimination.
Historical data creates another challenge.
If historical healthcare decisions were unequal, an AI system trained on those decisions can learn the inequality as if it were a legitimate predictive signal.
The model does not know that the historical pattern was problematic.
It simply learns:
βThis happened frequently in the past.β
That creates a dangerous feedback loop:
Historical practice β training data β algorithm β new decision β new data.
Without intervention, historical patterns can become algorithmic patterns.
That does not mean insurers should stop using AI.
It means AI should be governed differently when it affects consumers' rights and access to healthcare.
Insurers should know:
- What their models do.
- What data they use.
- What variables act as proxies.
- How the models perform.
- Whether errors differ across populations.
- How adverse outcomes are reviewed.
Third-party vendors should not be treated as black boxes.
And fairness testing should not be treated as a one-time certification.
Models change.
Data changes.
Healthcare changes.
Consumer behaviour changes.
Therefore, discrimination testing must be capable of continuing after deployment.
The 2026 legal landscape also demonstrates why precision matters.
Section 1557 continues to prohibit specified forms of discrimination in covered health programmes and activities, while the regulatory framework surrounding the statute has changed through litigation and HHS rulemaking. :contentReference[oaicite:6]{index=6}
HHS's July 2026 Title VI rulemaking also demonstrates that the legal treatment of disparate-impact theories can differ depending on the statutory and regulatory framework. :contentReference[oaicite:7]{index=7}
Accordingly, insurers should not rely on a generic statement such as:
βOur algorithm is fair.β
They should be able to demonstrate:
How fairness was evaluated, under which legal framework, using which data, and with what corrective safeguards.
The central principle is simple:
AI may automate a decision, but it does not automate away civil-rights obligations.
For health insurers, responsible AI therefore requires a combination of:
Actuarial discipline + technical validation + civil-rights analysis + clinical context + regulatory governance.
That combination will become increasingly important as algorithms move from administrative functions into the decisions that determine who receives healthcare, what it costs and how quickly it can be accessed.
Legal Disclaimer
This article is provided for general educational and informational purposes only. It is not legal, medical, insurance, actuarial, financial or regulatory advice and does not create an attorney-client relationship. The legal treatment of algorithmic discrimination varies by statute, insurance product, jurisdiction and individual circumstances.