Gemini planned their hike. Rescue teams had to finish it. The habit that would have caught it.
On September 5, 2026, TechCrunch reported that a group of hikers had to be rescued after planning their trip with Google Gemini. The detail that traveled came from the sheriff's office, which said the hikers "were advised by Gemini to bring far less food and water than their group required."
It would be easy to file this under trusting AI too much, but that framing lets the useful failure escape. Using a model to draft the plan was not the mistake; a draft is a reasonable place to start. A trip plan, though, is a stack of numbers wearing prose, and one of those numbers, water per person against distance and conditions, is load-bearing. It is also checkable: a published requirements table and one multiplication will confirm it or destroy it. The rescue tells you that check never won the argument. The plan read well right up until it met the terrain.
That habit of going underneath, past the fluent summary to the one number it stands on, is what the dive-deep question is there to find.
Why they ask it
Management runs on summaries by design. You cannot read every ticket, sit in every incident review, and audit every dashboard, and nobody wants you to. Your information diet is compressed by other people — dashboards, weekly updates, a thumbs-up in standup — and compression is where problems hide.
So the question, long associated with Amazon's Dive Deep principle though versions of it have been reported well beyond Amazon, tests two things. First, whether you can still make contact with raw reality when it matters: the actual logs, the actual tickets, the actual rows, not a tidier chart. Second, and this is the senior half, whether you know when it matters. The interviewer is listening for the moment a summary said one thing, something in you disagreed, and you went down to find out.
The trap
The common weak answer is a diligence story. "The numbers looked off, so I investigated thoroughly and found the issue." Thoroughness is a work ethic, not a finding. Nothing in that sentence proves you went a single layer below where anyone else would have stopped.
The subtler trap is depth as a lifestyle. Candidates keen to prove rigor describe diving deep on everything, which at the manager level reads as an inability to delegate or to prioritize. Depth is expensive; that is the point of it. Spent everywhere, it signals you cannot tell where it is needed.
A gradeable answer has three parts. A specific doubt that triggered the descent. The raw layer you went to — not a better dashboard, but the thing underneath the dashboards. And the one fact you brought back that no summary could have shown you.
Applying STAR-T
Situation. Open with the disagreement between an aggregate and a human signal. "Our on-call dashboard showed load flat quarter over quarter, but two strong engineers resigned within a month, and both mentioned burnout in their exit conversations." One summary, one anecdote, pointing in opposite directions.
Task. State that arbitrating between them was yours to do. "Both could not be true, and hiring backfills without knowing which was real meant burning out whoever came next."
Action. Describe the descent concretely. "I exported six months of raw pager events and rebuilt the picture per person, per hour, ignoring the team-level averages entirely." Then the finding, because depth is only worth the trip if it returns something: "The mean was flat, but the distribution was brutal — an escalation shortcut nobody had reviewed was sending most after-midnight pages to the same two people." Name the layer, name the method, name the fact.
Result. Give two results, one for the system and one for the truth. "We rewrote the escalation policy and the rotation; after-midnight pages for the worst-hit engineers fell by more than half, and nobody else resigned that year. The dashboard also gained a distribution view, so the average could never hide that shape again."
Trade-off. Depth costs something; say what. "I spent two days of a planning week on forensics, and I had to tell the platform team their metric had hidden a problem for two quarters. I accepted the friction, because the alternative was pricing a team's health off an average."
The follow-up that breaks weak answers
Expect the meta-question: "how do you decide when to dive deep and when to trust the summary?" The weak answers are "instinct," which cannot be examined, and "always," which cannot be afforded.
A strong answer states a trigger rule and a boundary. Dive when the number is load-bearing and unverified — when a decision of real size rests on a figure nobody has touched. Dive when two sources disagree, especially when one is an aggregate and the other is a person. Dive when the explanation is suspiciously tidy. Then name what you deliberately did not dive on, because the interviewer is also checking that your depth is allocated, not compulsive.
Score your answer against the director’s bar
Q: Tell me about a time you dove deep.
Bank the story where you went under the dashboard and came back with the number, and rehearse it in L8 Loop until the descent sounds routine. Try it free →