Question 17
IA company encodes internal policies (e.g., "never give legal advice," "do not generate discriminatory content") as explicit instructions in the system prompt for an LLM. What is this practice often called?
Correct answer: B
Explanation
This is called "Policy-as-prompt" because the organization turns internal rules into explicit prompt instructions that steer the model’s outputs. The description matches "organizational policies are translated into prompt constraints to guide model behavior," which is exactly what embedding rules like "never give legal advice" or "do not generate discriminatory content" does.
Why each option is right or wrong
A. Data poisoning.
Data poisoning means corrupting training data to manipulate model behavior, not adding runtime instructions.
B. Policy-as-prompt, where organizational policies are translated into prompt constraints to guide model behavior.
The scenario describes organizational rules being operationalized as explicit system-prompt constraints, which is the standard definition of policy-as-prompt in prompt engineering. The key diagnostic feature is that the company is not fine-tuning or hard-coding the rules; it is translating internal governance requirements into instructions that steer the model’s responses at runtime.
C. Model compression.
Model compression reduces model size or compute needs; it does not encode behavioral policies.
D. Gradient clipping.
Gradient clipping stabilizes training by limiting gradient magnitude, unrelated to prompt-based policy control.