Kuis Kalibrasi Kepercayaan AI

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. What does statistical calibration describe?
   • Whether users like the interface
   • Whether the model ranks every example correctly
   • How often the model explains itself
   • Whether predicted probabilities match observed frequencies across comparable cases
2. Can a statistically calibrated model still perform poorly?
   • No, calibration guarantees high accuracy
   • Yes; calibration and task quality are different properties
   • Only if users distrust it
   • No, if its confidence is displayed
3. What does a model’s verbal statement “I am 95% sure” establish?
   • Only that it produced those words; the confidence needs empirical evaluation
   • That its output is calibrated
   • That the source evidence is correct
   • That a human should not review it
4. A model is calibrated on one dataset. What remains uncertain after a domain shift?
   • Whether its old test score exists
   • Whether all users agree with the system
   • Whether the calibration still holds under changed language, domain, or task
   • Whether the interface color is accessible
5. Which interface feature can support independent checking?
   • Use a decorative confidence badge without measurement
   • Show sources and relevant limitations beside the answer
   • Hide the source to reduce clutter
   • State that the model is always reliable
6. How should teams evaluate human trust in the system?
   • Measure only whether users click “accept”
   • Assume explanations always improve trust
   • Count the model’s confident words
   • Test whether people notice errors and use overrides appropriately
7. Why distinguish accuracy from calibration?
   • Accuracy and calibration are synonyms
   • Calibration only measures latency
   • A model may rank cases correctly while its probabilities are systematically too high or low
   • A calibrated score cannot be wrong
8. What should happen when measured calibration changes over time?
   • Update user guidance and keep a route to report errors or complete the task another way
   • Leave the interface unchanged indefinitely
   • Remove all sources
   • Display a more precise confidence value