3.1 Describe How to Proactively Overcome Foundation Model Limitations. GENERATIVE-AI-LEADER Practice Questions (Page 3)

Questions 11–15

1. 11expert · hard

   A company has a generative AI application that classifies customer support tickets. The model was evaluated at launch with 95% accuracy. Six months later, the team notices accuracy has dropped to 80% even though the model and its configuration are unchanged. The team needs to identify the cause and implement a monitoring solution. What is the most appropriate first step?

   • ADisable the model and switch to a rule-based system to guarantee accuracy.
   • BRoll back to the previous model version because the current one is faulty.
   • CSet up continuous evaluation with a held-out dataset and track performance metrics to detect drift.
   • DImmediately fine-tune the model on the most recent customer tickets to restore accuracy.
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A retail company has deployed a generative AI model to write product descriptions. After three months, the marketing team notices the descriptions have become repetitive and less aligned with the current brand voice. The model's underlying version has not changed. What is the most likely cause and the appropriate monitoring practice to detect this?

   • AThe temperature setting has degraded; re-deploy the model with the original configuration.
   • BThe model has reached its knowledge cutoff; schedule automatic model upgrades.
   • CThe model is overfitting to the training data; apply more aggressive fine-tuning.
   • DData drift in the input product catalog; track output distribution and performance metrics over time.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A tech company has a generative AI model in production that generates code snippets. The model was fine-tuned on internal codebases. The team wants to ensure the model remains effective as the internal codebase evolves. Which practice should they implement?

   • AUse a larger model to handle the growing codebase automatically.
   • BSet up a pipeline to periodically re-evaluate the model against new internal code samples and retrain if performance drops.
   • CIncrease the model's temperature to generate more varied code snippets.
   • DFreeze the model version to ensure consistent output for all users.
   Reveal Solution Discussion Select an answer first
5. 14foundation · easy

   A team is deploying a generative AI model for a medical diagnosis assistant. They want to reduce the risk of the model giving harmful advice without human oversight. Which Google-recommended practice is most directly aimed at this concern?

   • AGrounding with public web data
   • BPrompt engineering
   • CFine-tuning on general medical textbooks
   • DHuman-in-the-loop (HITL)
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   A legal firm wants to use a foundation model to draft contract clauses. The firm has a large corpus of approved clauses and strict requirements: (1) the output must match the firm's established legal style, and (2) the model must not cite outdated or superseded clauses. The team has limited ML engineering resources and needs a solution that is fast to implement. Which approach best balances these constraints?

   • AFine-tune the model on the approved clause corpus and deploy it without additional review.
   • BUse prompt engineering with a few examples of approved clauses and no external data source.
   • CUse RAG with the approved clause corpus and add a human-in-the-loop review step for final approval.
   • DDeploy the model with a low temperature and rely on its pre-trained legal knowledge.
   Reveal Solution Discussion Select an answer first