AI and safety · Question 49 of 60
A self-learning assistant in a control center was comprehensively tested before deployment. Which measures help so that no dangerous behavior creeps in during operation?
2 answers are correct. Hard
Show answer
Answer
- A Managing how the model learns and which data it uses
- D Implementing safety guards that prevent learning or decisions that compromise safety
Explanation
For self-learning systems, the behavior progressively moves away from the originally tested behavior, which undermines the safety case. Managing how the model learns and which data it uses can help. Alternatively, safety guards can prevent it from learning or deciding in ways that compromise safety, for example a content moderation component that filters prompts.