Generative AI Leader Practice Question: A research team uses a generative AI model to…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Chain-of-thought reasoning

  Why this is correct

  Chain-of-thought prompting makes the model emit intermediate reasoning steps before its conclusion, exposing the inferential path users can inspect. This satisfies the stem's explainability requirement by revealing how historical texts were interpreted, unlike post-hoc methods that only approximate feature importance.

• ✗

  Grounding

  Why it's wrong here

  Grounding ties generated statements to retrieved source passages, which supports citation and factuality rather than exposing the model's internal reasoning steps. It is tempting because it is the standard technique for retrieval-augmented historical analysis, but the stem asks how the model reached its conclusion, not which sources support it.

• ✗

  Confidence indicators

  Why it's wrong here

  Confidence indicators expose only the model's certainty score for its output, not the chain of reasoning that produced it. They are tempting because they cheaply flag unreliable answers in classification or QA deployments, but the team needs the derivation itself, which chain-of-thought or attention-based explanation supplies.

• ✗

  Safety filters

  Why it's wrong here

  Safety filters block or flag harmful, toxic, or policy-violating outputs; they reveal nothing about how the model derived an answer. They are tempting because they are a standard part of responsible generative AI deployment, but they operate on output acceptability, whereas the team wants visibility into the reasoning process itself.

Go deeper

Related to this question