AI Interview Questions That Reveal Judgment (With a Scoring Guide)

Most AI interview questions can be answered by reading a blog post on the way to the interview. "How do you use AI in your work?" is a prompt for a rehearsed paragraph. The questions that work have one property in common: they ask for a specific story with a mistake in it, because you cannot rehearse a story about a failure you didn't have.

This is a question bank for hiring anyone into a role where AI is in the workflow — engineer, accountant, recruiter, marketer, analyst. Each question lists what it exposes, what a strong answer sounds like, the follow-up that breaks a prepared answer, and what a weak answer looks like. Scoring guidance and a 45-minute structure are at the end. If you want the underlying model, it's the six pillars of AI-first talent; if you want the practical test version, see assessing AI skills in non-technical roles.

1. Verification: "When did AI give you a confident, wrong answer? How did you catch it?"

What it exposes: whether they actually use these systems on consequential work, and whether checking is a habit or an afterthought.

Strong: a specific story with a mechanism — "it cited a case that didn't exist, and I checked because I'd been burned before," or "the totals didn't tie to the source file, so I re-ran the query by hand." The strong answer always contains what the check is now.

Follow-up: "What's the check you added to your workflow after that?" — a real story produces a changed process. A rehearsed one produces a philosophy.

Weak: "I always double-check everything" with no example. Or a story where the error was caught by someone else.

2. Delegation: "Walk me through a task you handed to AI completely, and one you'd never hand over."

What it exposes: whether they can split work deliberately, and whether they understand the boundary between drafting and deciding.

Strong: the pair is specific and asymmetric — "I hand over first drafts of internal summaries and meeting notes; I never hand over the accrual estimate or the message to a customer about a delay." They can articulate why the line sits there: accountability, judgment, or irreversible consequences.

Follow-up: "What would have to be true for you to move that second one across the line?" — you're testing whether the boundary is reasoned or superstitious.

Weak: "I let it write everything, I just edit it" (no boundary), or "I don't let it touch anything important" (no leverage).

3. Judgment: "Tell me about a time you overruled what the model said."

What it exposes: whether they own the decision, and whether they can defend it with evidence rather than instinct.

Strong: a concrete override with a reason — a candidate the ranking dropped who they pushed forward, a forecast the model produced that ignored a known one-off, a diagnosis that contradicted the data. They name what they used that the model didn't have.

Follow-up: "Were you wrong that time?" — the genuinely thoughtful candidate can produce the reverse case. Anyone who has never been wrong hasn't been deciding much.

Weak: "I trust it mostly" — the machine is doing the thinking.

4. Ethics: "How do you keep AI-assisted screening fair?"

What it exposes: whether fairness is a designed control or a slogan.

Strong: talks about rubric design before ranking, auditing false negatives by sampling rejected candidates, checking adverse impact, and keeping humans on consequential decisions. They may name what they explicitly exclude from the model's view.

Follow-up: "Show me the check you'd run on a shortlist of 30 before it goes to a hiring manager." — you want a procedure, not a value statement.

Weak: a prepared sentence about bias with no procedure attached, or "the model is trained on lots of data so it's neutral."

5. Teaching: "Explain a technique you use to someone who's never tried it. I'm that person."

What it exposes: whether the capability is transferable or personal. This is the single best predictor of whether they'll raise the team or just be fast alone.

Strong: starts with the problem, not the tool; checks for understanding; gives a worked example; makes you leave with one thing to try.

Follow-up: "How would you teach this to five people at once?" — you're looking for a template, a checklist, or a short session, not a repeat of the explanation.

Weak: jargon, or an explanation that only makes sense if you already know what they do.

6. Workflow design: "Show me something you built that other people now use."

What it exposes: whether they've built a workflow or just become good at a tool. This is the difference between Level 2 and Level 3 in our AI fluency levels.

Strong: a template, checklist, prompt chain, automation or rubric with a name and users. They can say what broke when they first shipped it and how they fixed it.

Follow-up: "What did it replace, and what did it make worse?" — every real workflow has a cost. Someone who can't name one hasn't owned one.

Weak: "I have a really good prompt" — useful to them, invisible to everyone else.

7. Learning: "What changed in the last six months that made you change how you work?"

What it exposes: whether they're actually paying attention, and whether they discard what they know.

Strong: a specific change — a model release, a pricing shift, a capability that arrived — and a workflow they abandoned because of it.

Follow-up: "What did you have to unlearn?" — the hardest part of this field is deleting what stopped working.

Weak: nothing has changed, or their setup is identical to a year ago.

8. Craft first: "What's the hardest thing about this job that AI doesn't touch?"

What it exposes: whether they understand the role deeply enough to name its irreducible core — and whether they'd be lost without the machine.

Strong: a precise answer specific to the craft: for a recruiter, closing and market counsel; for a controller, the estimate and the audit relationship; for an analyst, deciding which question is worth asking.

Follow-up: "How would you know you'd gotten bad at that?" — self-assessment discipline.

Weak: a generic answer about "the human touch," or an answer that implies the machine does the hard part.

Scoring

Score each answer 1–5 against written anchors, and don't let one strong answer carry the interview. The bar we use for AI-first roles:

• Verification and Judgment both 3+ — non-negotiable. Below that, the person is a speed risk on consequential work.
• Delegation 3+ — otherwise you're hiring a slow worker with good intentions.
• Teaching 4+ for anything senior or team-lead — you're buying the multiplier, not just the output.
• No dimension at 1.

Use the candidate scorecard builder to lock the anchors before the interview, and the screening call script for the 20-minute version. The full structure — including how to run the work sample — is in the AI-first hiring framework.

A 45-minute structure

Two interviewers, scored separately before you compare notes. And wherever the role allows, put a real artefact in front of them for twenty minutes — a messy email thread, a bad report, a vague brief. The work sample catches what the conversation can't.

Frequently asked questions

Should I ask candidates to use AI during the interview?

Yes, where the role uses it. Give them a real task with a tool available and watch the process: what they hand over, what they check, what they do when the output looks wrong. It's the most honest version of the job.

How do I stop candidates from reciting answers they found online?

Ask for the follow-up, not the headline. "What check did you add?" and "Were you wrong that time?" are hard to prepare for because they require a real history. Rehearsed answers collapse on the second question.

What's the single best question if you only get one?

"When did AI give you a confident, wrong answer — and what's the check you run now?" It simultaneously proves real usage, reveals a verification habit, and shows whether experience changed their process.

Do these questions work for senior and executive candidates?

They work better, because senior candidates have more history to draw on. For executives, swap question 6 for "what workflow did you build for your team, and what did it make worse?" — ownership and trade-off thinking is what you're buying at that level.