引導式測驗 · 中等 等級
AI 測驗中的交叉偏差
使用來自性別陰影的範圍有限的證據,測試組合的組成員身份如何揭示單軸平均值隱藏的錯誤。
問題 1 的 8
What does intersectional bias describe in an AI audit?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What does intersectional bias describe in an AI audit?
- Only a low overall score
- A model that runs slowly
- Worse outcomes for a combined subgroup hidden by broad averages
- A missing database index
- Which group had the highest errors in the Gender Shades study’s tested systems?
- Lighter-skinned men
- Darker-skinned women
- All groups equally
- People using wheelchairs
- What limitation should accompany the Gender Shades result?
- It proved all current AI systems are biased
- It studied specific systems and a binary gender-classification task at that time
- It evaluated only loan decisions
- It included every demographic group
- Why can separate race and gender metrics miss a problem?
- They always use invalid mathematics
- They remove the need for testing
- A disparity at the race-by-gender intersection can average out in broad groups
- They reveal every individual prediction
- What should be reported alongside intersectional performance rates?
- Counts and uncertainty
- Only the best subgroup score
- The model’s marketing slogan
- No information about sample size
- Why may intersectional testing require privacy safeguards?
- It can involve sensitive attributes and small cells
- It always reveals model weights
- It prevents evaluation
- It changes the protected class definitions
- What should a team avoid concluding from sparse subgroup data?
- That uncertainty should be monitored
- That more evidence may be needed
- That a group is unaffected solely because the estimate is imprecise
- That privacy controls matter
- When should intersectional evaluation be repeated?
- Only before the first release
- After updates and when deployment expands to new groups
- Never if the model passed one audit
- Only when overall accuracy falls