Candidates rehearse standard behavioral questions with AI and arrive polished. These eight resist coaching because they demand a specific real event and reward follow-up that generic prep can’t survive. Pair them with the 15-minute structured phone screen. For the strategy, see the pillar guide.

Quick Comparison

1. “Describe a judgment call you made with incomplete information.”

This is the anchor question. There’s no fixed right answer to reverse-engineer, and the follow-up “what did you decide and why” exposes whether they lived it. Run it on a real candidate and watch the shape of the answer. A manager describes deciding whether to ship a release with a known minor bug or slip the date a week, with the data on the bug’s blast radius still incomplete. You ask what they were missing, and they tell you they did not yet know how many customers hit that code path. You ask what tipped them, and they say they shipped because the workaround was one click and the slip risked a contract deadline. The mechanism is that incomplete information has no answer key, so a candidate cannot retrieve the “right” response from a tool. They can only reconstruct what they actually weighed, and reconstruction under follow-up is exactly what rehearsal cannot fake.

• Listen for the specific constraint.
• Probe the alternative they rejected.
• Ask what information they wished they had had, and why they moved anyway.

Verdict: The single best AI-resistant question.

2. “Tell me about a decision you got wrong.”

Coached answers default to humble-brags. A real one names a genuine mistake and what changed after. The coached version is easy to spot because it always resolves in the candidate’s favor: “I worked too hard and burned out, so I learned to delegate.” A real answer has a cost the candidate actually paid. Someone says they kept a struggling hire three months too long because they liked the person, the team’s output suffered, and they now run a thirty-day checkpoint they cannot talk themselves out of. The mechanism is that admitting a real mistake carries social risk in an interview, and a candidate willing to pay that risk is showing you something a generated answer will not: the self-awareness to name a genuine failure and the discipline to have changed because of it. Follow up by asking what specifically changed in how they work, because that is where the learning either is or is not.

• Reward honest error over polished spin.
• Ask what they’d do differently now.
• Push past mistakes that conveniently flatter the candidate.

Verdict: Surfaces integrity fast.

3. “Walk me through a tradeoff where both options were bad.”

Forces prioritization under real constraints. The reasoning is the deliverable. The point is the word “both” — you are removing the escape hatch where a candidate finds a clever third option that makes everyone happy. A real answer commits to a loss. Someone tells you they had to choose between missing a customer deadline or shipping without full test coverage, and they shipped, accepting the support load because the relationship mattered more than the risk that quarter. The mechanism is that prioritization only becomes visible when something has to be given up. Anyone can list priorities in the abstract, but forcing a sacrifice reveals the candidate’s actual ranking under pressure. Accept any direction they defend well, because you are scoring the quality of the reasoning, not whether you would have made the same call.

• Score what they chose to sacrifice.
• Accept multiple defensible answers.
• Reject answers that dodge the tradeoff by inventing a painless option.

Verdict: Strong judgment signal.

4. “How did you discover a problem nobody flagged?”

Reveals analytical instinct. Fabricated answers lack the diagnostic chain. A strong answer has a sequence to it: something looked off, the candidate pulled a thread, the thread led somewhere, and they confirmed it before raising the alarm. Someone notices that a report’s totals drifted from the dashboard, traces it to a timezone mismatch in one data source, and verifies it by re-running a single day’s numbers by hand. A fabricated answer skips the middle — it states the problem and the resolution but cannot reconstruct the steps between, because there were no steps, only a story. The mechanism is that discovery is a chain of small, specific observations, and a chain is hard to invent convincingly when you ask “and how did you know that was the real cause and not a coincidence?”

• Ask how they confirmed it was real.
• Probe for the specific observation that first tipped them off.

Verdict: High depth signal.

5. “What did you change your mind about on this team?”

Tests intellectual honesty. Generic prep struggles with specificity here. The question demands a before and an after that belong to one real context, which is hard to generate because it requires a specific prior belief, a specific piece of evidence, and a specific reversal. A candidate says they used to push for synchronous standups every day, watched the remote half of the team lose two hours to it, and switched to async written updates after the data showed velocity rose. The mechanism is that changing your mind on evidence is the opposite of what a model produces — generated answers tend to defend a position fluently rather than abandon one. Probe the evidence: a real reversal can tell you exactly what moved it, while a manufactured one waves at “feedback” without a concrete trigger.

• Probe what evidence moved them.
• Ask what they believed before and what specifically broke that belief.

Verdict: Reliable.

6. “Describe work you’re proud of that looked unimpressive on paper.”

Directly targets the gap between presentation and substance. This question rewards exactly what a homogenized resume hides. Someone tells you the work they are proudest of was deleting a feature — pulling a rarely-used module that three customers loved but that caused a third of all support tickets. It reads as a negative on a resume, a thing removed rather than built, but it was the highest-leverage call they made that year. The mechanism is that the question inverts the resume’s incentive. A resume optimizes for impressive-looking accomplishments, so asking for valuable-but-unimpressive work surfaces judgment that the resume actively suppresses. Listen for why the candidate thought it mattered despite looking small, because that reasoning is the signal.

• Listen for why it mattered despite looking small.
• Reward judgment that a resume would have buried or omitted.

Verdict: Cuts through homogenized resumes.

7. “Where were you wrong about a person or a plan?”

Hard to coach because it requires a real, slightly uncomfortable story. The discomfort is the feature. A candidate admits they wrote off a quiet junior hire as disengaged, only to learn the person was solving the team’s hardest bug on their own time and never mentioned it. There is a small sting in telling that story, and the sting is what marks it as real — a generated answer files the rough edge off until nothing is left but a tidy lesson. The mechanism is that genuine reflection includes a moment the candidate would rather not have had, and rehearsed answers sand those moments away. Reward the candidate who lets the discomfort show and tells you what it changed, and discount the one whose error somehow reflects well on them.

• Reward candor and reflection.
• Distrust a “mistake” that leaves the candidate looking better for having made it.

Verdict: Good integrity check.

8. “Explain a past project to me like I’ll ask three follow-ups.”

Sets the expectation of depth up front. Borrowed answers buckle by the third question. Naming the follow-ups in advance does real work: it tells the candidate the conversation is going deep, which steadies honest people and unsettles rehearsed ones. Then you actually go deep. They describe the project, and you ask why they chose that architecture, then why not the obvious alternative, then what broke first when it scaled. By the third question you have left every rehearsed boundary, and the candidate is either still supplying specifics or visibly improvising. The mechanism is that any single question can be prepped once, but depth is unbounded — there is always another “why” — and only lived experience keeps producing real answers that far down. The follow-up is not a supplement to the question; it is the test.

• Actually ask the three follow-ups.
• Go one question deeper than feels polite, because that is where the signal lives.

Verdict: The follow-up is the test.

How We Evaluated

Questions were rated on resistance to generic AI prep and on how much real signal the follow-ups surface. For where these fit in the funnel, see screening signals HR can trust and the pillar guide.