An insurance company uses an AI model to set premium rates.

AAIA guidance states thatfairness evaluations begin with the training data, because bias is most commonly introduced through data selection, sampling imbalances, labeling inconsistencies, or historical discrimination embedded in source data.

Thus, thefirstcourse of action is to review training data (option B) to identify:

Skewed demographic distributions

Missing or underrepresented populations

Inappropriate use of sensitive attributes

Incorrect labels or improperly encoded variables

Historical decisions that may propagate discrimination

Only after identifying the existence and nature of bias can an organization move on to remediation steps such as retraining (C) or comparing alternate model behavior (A).

Allowing customers to contest premiums (D) is a post-decision remedy, not a fairness evaluation step.

AAIA emphasizes addressing bias at theroot—the training data—before adjusting the model or outputs.