StarApple AICanada

Tool 03

Ten questions on how these systems actually behave.

Each answer is explained as soon as you pick one, right or wrong. The explanations are the point, so take the questions you are unsure about rather than the ones you know.

1. A language model produces a confident, detailed answer citing a court case. What is the safest assumption?
2. What does a model's context window limit?
3. Your team pastes a client contract into a public chatbot to summarise it. The main risk is:
4. Retrieval-augmented generation (RAG) mainly helps by:
5. Prompt injection is best described as:
6. Fine-tuning a model on your own data primarily changes:
7. Your model scores 94 percent accuracy overall but 61 percent for one customer segment. This is:
8. Under the Treasury Board Directive on Automated Decision-Making, an Algorithmic Impact Assessment is:
9. Temperature set to zero means the model will:
10. The most reliable way to know whether an AI system is working in production is: