179 questions
No questions match those filters.
How would you design an AI system a regulator could actually audit?
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansI’d start from the reconstruction test: months after the fact, can you show exactly why the system produced a specific output for a specific input? If the answer is no, the system isn’t auditable, however good its accuracy numbers look.
Passing that test requires recording the decision path, not just the final outcome:
- The exact inputs.
- Any retrieved context, with its sources — not just the final answer that used it.
- The model and prompt versions in effect at that moment.
- The confidence or score attached to the decision.
- Which stages were deterministic rules versus model-driven judgment.
- The human decision that followed, where a human was in the loop.
Two design principles fall out of that directly. First, make irreversible actions human-gated, so there’s always a named, accountable person behind anything that can’t be undone. Second, keep the deterministic parts of the system actually deterministic — every rule that gets moved into a model’s judgment instead of staying an explicit rule is a decision that can no longer be explained simply to an auditor, even if it performs better on average.