7.3. Mixed-Topic Practice Questions
These mirror the exam's format and span all domains. Try each before reading the answer.
1. A bank's loan-approval model receives no race or gender field, yet approves one demographic far more often. Which responsible-AI principle is most directly at stake?
- A. Transparency
- B. Fairness
- C. Reliability and safety
- D. Inclusiveness
Answer
B — Fairness. Fairness is about equitable outcomes across groups. Removing the sensitive attribute doesn't ensure fairness because proxy features (like zip code) reintroduce bias. Transparency (A) is about understandability; inclusiveness (D) is about access for all abilities/backgrounds.
2. You need to classify 100,000 short product reviews as positive or negative, as cheaply and quickly as possible. Which model choice best fits?
- A. The largest available frontier reasoning model
- B. A small, lower-cost model capable of sentiment classification
- C. A multimodal image model
- D. An agent with multiple tools
Answer
B. Match the model to the task's actual needs. Sentiment classification is simple and high-volume, so the smallest capable model minimizes cost and latency. The frontier model (A) is overkill; a multimodal image model (C) is the wrong modality; an agent (D) is unnecessary when no actions/tools are required.
3. A Python snippet authenticates with AIProjectClient.from_connection_string(...). What does this most likely indicate?
- A. The current, recommended Foundry pattern
- B. An image-generation workflow
- C. The older Classic / hub-based project pattern
- D. A speech-recognition application
Answer
C. Current Foundry projects use an endpoint plus DefaultAzureCredential. A connection string indicates the older Classic / hub-based approach. The pattern says nothing about modality (B, D).
4. A team needs to pull the merchant name, date, and total from photographed receipts into a database. Which capability fits best?
- A. Image classification
- B. Speech recognition
- C. Image generation
- D. Information extraction with Content Understanding
Answer
D. They need specific structured fields from the image, which is information extraction. Classification (A) only labels the image; speech recognition (B) is the wrong modality; image generation (C) creates images rather than reading them.
5. Which scenario requires an agent rather than a single model call?
- A. Summarize a provided article
- B. Translate a sentence
- C. Check live flight status and rebook if delayed
- D. Classify an email's sentiment
Answer
C. Only this requires taking actions (calling a status tool, performing a rebooking) toward a goal — the hallmark of an agent. The others are single generation/analysis tasks a plain model call handles.
6. A summarization feature returns a different summary every time it runs on the same document. Which single change most improves consistency?
- A. Raise the temperature
- B. Lower the temperature
- C. Increase max tokens
- D. Switch to a multimodal model
Answer
B. Temperature controls randomness; lowering it makes output more focused and consistent. Raising it (A) increases variation; max tokens (C) affects length, not consistency; modality (D) is irrelevant here.
7. Which best distinguishes transparency from accountability?
- A. They are the same principle
- B. Transparency makes the system understandable; accountability keeps humans answerable
- C. Transparency is about data protection; accountability is about fairness
- D. Transparency applies only to code; accountability only to data
Answer
B. Transparency = people understand how/when AI is used. Accountability = humans remain answerable for the system. A system can be transparent yet still lack a responsible owner.
8. An app must let users speak a question and hear a spoken answer. Which capabilities does it use, in order?
- A. Speech synthesis, then speech recognition
- B. Image generation, then OCR
- C. Speech recognition, then speech synthesis
- D. Entity detection, then summarization
Answer
C. Recognition (speech-to-text) captures the question; the app reasons; synthesis (text-to-speech) speaks the answer. A multimodal audio model could collapse the first step, but among these options C is correct.
