Safety and Guardrails GENERATIVE-AI-ENGINEER-ASSOCIATE Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   A data scientist is reviewing a set of LLM responses to identify common quality and safety issues. Which activity best describes qualitative response assessment?

   • ARunning automated metrics like BLEU or ROUGE to compute similarity scores
   • BMeasuring response latency and token usage to optimize performance
   • CManually reading responses to judge coherence, relevance, and potential harm
   • DChecking that the model's output format matches the expected JSON schema
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A legal-tech startup uses an LLM to draft contract clauses. During qualitative testing, they find that the model sometimes generates clauses that are grammatically correct but do not address the specific legal point requested. Which quality issue is this, and what guardrail would help?

   • AFactual inaccuracy; implement an input guardrail that restricts prompts to legal terms.
   • BIncoherence; implement a post-processing step that rephrases the clause.
   • CBias; implement an output guardrail that detects biased language.
   • DIrrelevance; implement an output guardrail that checks the response against the requested legal topic.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A healthcare chatbot is being developed to answer patient questions. The team is concerned about users attempting to use the chatbot to obtain medical advice that could be harmful. They want to implement a guardrail that prevents the model from even processing prompts that ask for specific dosages of prescription medications. Which approach best meets this requirement?

   • AUse a post-processing step to check the model's output for dosage information and redact it if found.
   • BConfigure the model's temperature to 0 to reduce the chance of generating harmful content.
   • CImplement a pre-processing step that detects dosage-related keywords and blocks the request before it reaches the model.
   • DAdd a system prompt instructing the model to refuse dosage requests.
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   Which of the following is an example of an input guardrail?

   • APost-processing the model's response to remove offensive language
   • BDetecting and redacting personally identifiable information (PII) from the prompt
   • CValidating that the response matches a JSON schema
   • DLogging the response for audit purposes
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A team has implemented an input guardrail that blocks prompts containing certain keywords. After deployment, they notice that the guardrail is blocking many legitimate prompts that use those keywords in a benign context. They want to reduce false positives without compromising safety. Which approach is most effective?

   • AUse a more sophisticated input guardrail that considers context, such as a classifier trained on labeled examples.
   • BRemove the keyword-based guardrail and rely on the output guardrail.
   • CIncrease the model's temperature to make it more creative.
   • DExpand the keyword list to include more variations.
   Reveal Solution Discussion Select an answer first