Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ
Dạy học sinh về bài kiểm tra thiên vị AI
Khám phá cách dữ liệu, nhãn, bối cảnh và quyết định của con người định hình thành kiến AI và cách học sinh có thể đánh giá các ví dụ trong lớp bằng bằng chứng.
Câu hỏi 1 của 8
An image classifier works well in classroom lighting but misses faces in a dim hallway. Which factor changed?
Xem trước mọi câu hỏi8 câu hỏi
Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.
- An image classifier works well in classroom lighting but misses faces in a dim hallway. Which factor changed?
- The conditions of use differ from the test setting
- The class changed the definition of a face
- The model became fair in every context
- The hallway proves all image models fail
- Why might a dataset with equal numbers of examples still be unrepresentative?
- Equal counts guarantee coverage of every relevant case
- Balance makes the task definition unnecessary
- Only the number of records matters
- It may omit important conditions or contain labels that do not match real use
- A system outputs a risk score, and a staff member treats it as a final decision. Which source of bias may contribute?
- Only training data volume
- Human interpretation and over-reliance on the output
- Only the device screen size
- The system’s filename
- What can a small classroom test establish?
- The performance of every commercial AI system
- A definitive demographic impact across the population
- A limited observation about the tested examples and conditions
- That the model will behave identically in another setting
- Students see errors concentrated in one type of example. What should they do next?
- Record the test conditions and investigate data, labels and task design
- Declare the model malicious immediately
- Ignore the examples because total accuracy may be high
- Change the labels until results look balanced
- Why does an equal aggregate accuracy score not by itself prove fairness?
- Equal accuracy means each individual receives the same outcome
- Fairness is determined only by the score average
- It confirms the training labels were correct
- It can hide different error types and does not settle whether the application is appropriate
- Which is a possible response when a classroom analysis finds a consequential risk?
- Keep it because students found a bias pattern
- Limit the system, change its design, add an appeal path or decide not to use it
- Hide the results from affected users
- Use the score for a new purpose without testing
- What question helps connect technical output to real-world impact?
- Which model has the most confident tone?
- How many labels appear on the screen?
- Who benefits, who may be harmed, and who can challenge the result?
- Can the class avoid defining the task?