IA-03 Fairness & Non-discrimination

4 sub-dimensions · 19 risk scenarios

Evaluates whether the AI system outputs unfairly favor or discriminate against specific individuals, groups, or parts of society, avoiding harm to protected and vulnerable populations caused by algorithmic bias, unfair treatment, or cultural barriers.

🎯 BenefitsAchieves fair and unbiased allocation of resources and opportunities through debiased data and fairness audits, expanding diverse customer coverage and social inclusivity.

⚠️ Harms / RisksModels inherit or amplify historical data bias, producing unfair treatment, systematic exclusion, or reinforcing societal stereotypes against specific demographic groups (such as gender, race, or age).

Disparities in Quality of Service3 risk scenarios · ISO 23894 A.6

DefinitionEvaluates whether the AI system's performance differs across specific demographic groups (age, gender, ethnicity, etc.).

🎯 BenefitsEnsures the AI system delivers equivalent, high-quality service across demographic groups (accents, cultures, genders, ages), avoiding tiered performance gaps.

⚠️ Harms / RisksThe system exhibits markedly higher failure rates when recognizing certain minority groups, accents, or persons with disabilities, imposing asymmetric processing time and service barriers on those groups.

Representative ScenariosS0080 Disparate Accuracy in Biometric Recognition S0082 Linguistic and Cultural Understanding Bias S0104 Service Quality Disparity Across Groups

Inequitable Allocation of Opportunities6 risk scenarios · ISO 23894 A.6

DefinitionEvaluates whether the AI system's outputs lead to an inequitable allocation of opportunities across demographic groups.

🎯 BenefitsEliminates algorithmic and historical dataset bias, ensuring equal opportunity for all groups in critical decisions such as credit, recruitment, education, and resource allocation.

⚠️ Harms / RisksThe model inherits historical social biases or produces indirect discrimination through proxy variables, systematically lowering the scores of particular genders or ethnic groups.

Representative ScenariosS0026 Societal Bias Embedded in Historical Data S0079 Group Bias in Credit Scoring S0081 Gender Bias in Recruitment Screening S0083 Indirect Discrimination Through Proxy Variables S0085 Unequal Outcomes in Resource Allocation S0086 Amplified Discrimination at Feature Intersections

Stereotype Reinforcement6 risk scenarios · ISO 23894 A.6

DefinitionEvaluates whether the AI system's outputs reinforce, erase, or demean stereotypes of certain demographic groups.

🎯 BenefitsApplies diversity audits to generated content and recommendation models, preventing the system from entrenching, erasing, or demeaning stereotypes of minority or disadvantaged groups.

⚠️ Harms / RisksGenerative systems persistently depict a single ethnicity or gender when portraying occupations or roles, or synthesize marketing and educational content laden with bias and demeaning tendencies.

Representative ScenariosS0027 Insufficient Representation in Training Data S0029 Synthetic Data Amplifying Existing Bias S0030 Imbalanced Multilingual Data Distribution S0031 Stereotypes Learned and Reinforced S0084 Unfair Group Representation and Erasure S0197 Impact on Cultural Expression and Identity

Accessibility Exclusion4 risk scenarios · ISO 23894 A.6

DefinitionEvaluates whether interested parties are excluded due to disability or a lack of digital skills.

🎯 BenefitsFollows accessibility and inclusive design principles, preventing certain users from bearing asymmetric operational burdens due to limited digital skills or interface barriers.

⚠️ Harms / RisksProcess designs lacking accessibility support force the elderly or digitally disadvantaged to spend several times longer on repeated attempts, or even abandon their entitlements because of these barriers.

Representative ScenariosS0131 Service Logs Exposing Sensitive Inputs S0139 Unequal User Burden S0187 Recommendation Mechanisms Affecting Autonomy of Choice S0196 Obstruction of Capability Development