不公平群體表徵與消抹
Operation & Monitoring
Risk Description
AI systems misrepresenting, underrepresenting, or completely ignoring specific social groups, leading to their marginalization in AI services. For example, image generation models defaulting to specific ethnicities for "professionals," voice assistants failing to recognize certain accents, or search recommendations ignoring minority cultural content. Such systemic representational unfairness reinforces visibility gaps and power asymmetries in society.
Framework Mappings
iso 23894R3
tw principle公平與不歧視
tw risk type社會系統性影響（Societal Systemic）
Controls English translation pending
- 表徵公平性審計
- 包容性設計原則
Implementation Steps English translation pending
- 建立表徵公平性審計流程，評估各群體在模型輸出中的能見度
- 在模型開發中採用包容性設計原則，確保多元群體參與
- 建立使用者回饋機制，收集表徵不公平的報告
- 定期進行跨文化和跨群體的表徵平衡測試