At MIRAI, we believe that fairness in AI is a shared responsibility. In our latest challenge with a partner in the insurance sector, we applied our proprietary BRIO framework to evaluate the fairness of a fraud prediction algorithm used to triage insurance claims for potential fraud.

The Challenge

The task: Insurance Company X asked us to assess whether their proprietary AI model for fraud-risk assessment was behaving fairly across sensitive attributes — policyholder gender, policyholder age, and the geographic region where the claim occurred. They provided us with a dataset of about 33,000 insurance claims recorded over the 2020-2026 period. 

The following features were considered sensitive: 

• Policyholder Gender (2 classes) 
• Policyholder Age (binned into 5 classes) 
• Italian Region of Claim Occurrence (20 classes)

Problem: The dataset carried an intrinsic selection bias: it only contained claims already flagged for inspection — either by the algorithm’s automated triage or by a human investigator acting independently — so the analysis could not observe what happens below the triage threshold.

Our Approach

Using the BRIO (Bias-RIsk-Opacity) tool, we conducted three complementary fairness analyses:

– False Positive Rate Disparity: to check whether the algorithm misclassifies non-fraudulent claims uniformly across sensitive groups. 

– Calibration Disparity: to check whether the same score corresponds to the same actual fraud rate across sensitive groups. 

– Conditional Scoring Disparity: to check whether, at parity of true label, the score distribution is the same across sensitive groups.

Each disparity was quantified using MIRAI’s proprietary R-Index which, unlike standard analyses of fairness, combines together three crucial dimensions of algorithmic discrimination:

• Hazard (i.e., severity of discrimination), 

• Exposure (i.e., width of discrimination), 
• Vulnerability (i.e., concentration of distinct discrimination effects on the same individuals) 

The R-Index combines these dimensions into a single, comparable risk measure aligned with the AI Act’s requirements for discrimination risk assessment:

Key Findings

– The algorithm is fully fair with respect to policyholder gender, across all three analyses — no risk detected. 

– A low-risk bias against policyholders aged 56–70 emerges consistently. In particular, this group shows a higher false positive rate (R-Index 14%) and is systematically under-calibrated (R-Index 13%), meaning the model overestimates the real risk of fraud for these categories. 

– The most significant issue regards geography: claims occurring in Campania are the most penalized overall: the model was found to be systematically “harsher” in evaluating claims occurred in Campania at parity of true labels. This disparity appeared to be particularly stark in the group of genuine frauds, where we observed an R-Index of 23%, the highest of the entire report. 

– Overall, the algorithm proves to be a valid tool for fraud triage, but the geographic dimension requires attention: the moderate-risk disparity in conditional scoring is concentrated in claims ultimately confirmed as fraud, where Campania accounts for over half of all cases in that group.

Why It Matters

Fraud-detection models shape which policyholders get flagged, investigated, and delayed — with real consequences for people who filed a genuine claim. A model that systematically treats claims from one region, or one age group, as more suspicious than the evidence supports does not constitute just a fairness problem: it’s a compliance risk that insurers are required to identify, document, and mitigate under the AI Act. Multi-component metrics like BRIO’s R-Index make that risk measurable, interpretable, and trackable over time — rather than a single aggregate accuracy score that hides where the harm actually falls.

What’s Next

We recommended a deeper investigation into the conditional scoring disparity affecting confirmed-fraud claims, and evaluated the case for recalibrating score thresholds by age band and region to compensate for the false-positive-rate gaps identified. We also recommended periodic re-running of the three analyses to monitor bias trends over time, and the adoption of a dedicated platform — such as MIRAI Toolbox — to track this continuously as part of ongoing AI Act data-governance obligations.

Ready to Level Up Your AI?

With BRIO, MIRAI empowers organizations to detect, quantify, and mitigate unfairness in AI-enabled fraud detection. It balances ethical responsibility with business needs — fueling trust, compliance, and smarter risk triage.

Curious to see BRIO in action on your systems? Reach out to MIRAI to learn how fairness-aware AI can transform your risk analysis practice.

References:

Sass et al. (2026). Quantifying Values: The Problem of AI Risk.

Novelli et al. (2024). AI Risk Assessment: A Scenario-Based, Proportional Methodology for the AI Act.

Mantelero (2024). The Fundamental Rights Impact Assessment (FRIA) in the AI Act: Roots, legal obligations and key elements for a model template.

Castelnovo et al. (2020). A clarification of the nuances in the fairness metrics landscape.

Coraglia et al., Evaluating AI fairness in credit scoring with the BRIO tool, arXiv (June 2024).

Buda et al., Bias Amplification Chains in ML-based Systems with an Application to Credit Scoring, BEWARE 2024.