Generative AI Leader Practice Question: Deploying a generative AI application that must…

Answer analysis

Option-by-option breakdown

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

• ✗

  Use Gemini 1.5 Pro with a system prompt asking for explanations

  Why it's wrong here

  A system prompt requesting explanations yields post-hoc rationalisations from an opaque model, not a faithful account of the computation, so it cannot satisfy a right to explanation. It suits flexible natural-language output, not inherent interpretability, which decision trees or linear models provide.

• ✓

  Use a smaller, simpler model that is inherently more interpretable, such as a logistic regression or decision tree

  Why this is correct

  Logistic regression and decision trees expose their decision logic directly through coefficients or split rules, so each prediction can be traced and justified to a regulator. This satisfies GDPR's right to explanation, which opaque deep generative models cannot provide inherently.

• ✗

  Use Retrieval-Augmented Generation (RAG) to ground responses in source documents

  Why it's wrong here

  RAG grounds outputs in retrieved documents but the model's own reasoning remains opaque, so it cannot explain how an answer was derived. It is tempting because it improves factual accuracy and traceability to sources, which suits citation and hallucination-reduction scenarios rather than GDPR interpretability.

• ✗

  Use PaLM 2 with prompt engineering to provide step-by-step reasoning

  Why it's wrong here

  Prompt engineering for step-by-step reasoning produces generated reasoning traces that may not reflect the model's actual internal computation, so it cannot guarantee justifiable decisions. It suits eliciting structured answers from large language models, not the inherent interpretability that inherently transparent model architectures provide.

Go deeper

Related to this question