Why Does Every Candidate Give the Same Polished STAR Answer?

Why do the answers all sound the same?

Because STAR is a public, fixed, four-beat template, and a fixed template is the easiest thing in the world to generate. Paste the job description into any assistant and ask for behavioral answers; what comes back is a Situation, a Task, an Action and a quantified Result, in that order, in competent prose. Every candidate for your role pastes the same posting, so every candidate gets the same story.

Coaching is not new. Only the price is. Voluntary interview coaching was already raising situational-interview scores in police and fire promotional processes twenty-five years ago, controlling for job knowledge, motivation to do well, race and sex. The study traced the effect to a specific mechanism: coached candidates organized their answers better, and that organization predicted the score ².

So the format has always rewarded rehearsal. What changed is that the coaching session used to be a scheduled event a fraction of candidates attended, and is now free and available at 2am to all of them. The distribution didn't improve; it compressed. Your strongest candidate and your weakest one sound alike in round one, because the thing they used to have in different amounts, preparation, stopped being scarce.

State that plainly before doing anything else, because it sets the whole diagnosis: this is not a wave of dishonesty. It is an instrument that measured preparation, still measuring preparation, at the moment preparation became free. The answers aren't fake. They're no longer diagnostic. Structured interviews versus AI coaching works through which parts of the format survive that and which parts don't.

What converges, and where?

The convergence point is occupational, not universal. Marketing candidates land on the same campaign-lift narrative; data analytics candidates land on the same stakeholder-alignment story; engineers land on the outage they debugged under pressure. An assistant reaches for the most-documented task in an occupation, and the most-documented tasks are the ones with a clean arc: a problem, an intervention, a number that moved.

So build the breaking question from the occupation's real failure mode rather than from the competency you meant to test. Two worked examples, drawn from what these jobs actually do:

• Marketing, SOC 13-1161. O*NET lists the work as measuring the effectiveness of marketing, advertising and communications programs, and forecasting and tracking marketing and sales trends ³. The generated answer always attributes a lift to the campaign. The question that breaks it: what else was running that quarter, and how did you separate it from your effect? Attribution is the actual failure mode of the occupation, and no template carries it.
• Data and analytics, SOC 15-2051. O*NET lists comparing models using statistical performance metrics such as loss functions or proportion of explained variance, and delivering results to management or other end users ⁴. The generated answer is always about aligning stakeholders on a definition. The question that breaks it: which model did you compare against, and what did the losing one get right? A rehearsed narrative names the meeting. The work names the metric.

The pattern generalizes. A competency such as "influencing without authority" or "driving results in ambiguity" has a story attached, and the story is generatable because thousands of versions of it are already written down. A failure mode (attribution, leakage, a metric that improved while the thing it stood for got worse, a denial appealed that should have been conceded) has one specific answer per candidate, and specificity is what a script cannot carry into a room.

Does asking more follow-ups fix it?

The evidence runs the other way. In the study that built the standard measure of interview faking, past-behavior questions were more resistant to faking than situational ones, and follow-up questioning increased faking rather than reducing it ¹. More turns give a prepared candidate more room to embellish, unless each turn asks for something that can be checked.

The same research put the base rate where it is uncomfortable. Over 90% of the job candidates studied faked during employment interviews, with the behaviors semantically closest to lying (inventing an experience, borrowing someone else's) running between 28% and 75% depending on the specific behavior ¹. That was measured long before any of this was automated. Treat a polished answer as normal, not as evidence of anything.

A follow-up earns its place when it makes the candidate operate on their own answer instead of continuing it. "Tell me more about the stakeholder piece" continues. "You said conversion rose eleven points. What was the baseline period, and what does that number become against the prior quarter?" operates. The second one has an answer that is either there or it isn't, and the candidate finds out which in front of you.

Follow-up questions that expose understanding covers the mechanics. The short version: depth of one is the tell. A rehearsed narrative survives the first probe and dies on the second, because the preparation covered the story rather than the decisions inside it.

Ask for what the story left out

Every STAR answer is a selection: it names the action taken and silently drops the ones rejected. The discarded branch is where the judgment lives, and it is the part no template produces, because a generated story never considered anything it then dropped. Ask for it directly, and ask what the choice cost. Four probes do most of the work, and none of them need a tool.

• What did you decide against, and what did it cost? A real project has a rejected option with a price attached. A generated one has a straight line from problem to result.
• Who disagreed, and were they right? Named disagreement is specific, checkable at reference stage, and almost never in the prepared version.
• What would have made you wrong? This asks for the falsifier, which is the same thing a good analyst asks of their own work before shipping it.
• Describe the version before the good one. Every real deliverable has a bad draft. Ask what changed between them and why.

The stronger version stops asking about past work and puts material in front of the candidate. Hand them a short, confidently wrong AI output from their own occupation and ask what they'd do with it. The answer is produced on the spot, against something you wrote, and no amount of preparation reaches it. Testing whether a candidate catches AI errors covers how to build one that reads as work rather than as a trick. See how Olive measures this.

Change what the round measures

Keep the structure and change the evidence. Structured interviewing is still worth having (same questions, same order, a rubric written before the first candidate), but the rubric has to score something other than narrative quality. Score whether a specific number was defended, whether a rejected option was named, whether a claim was checked against something outside the story the candidate came in with.

Three changes carry most of the load:

• Write the rubric first, and score behaviors rather than stories. "Named a constraint before proposing an approach" is observable. "Strong communicator" is a description of polish, which is the thing that stopped varying.
• Let the assistant into the room on purpose. A round that bans AI measures a version of the job nobody has. A round where the model is open, and the candidate decides what to hand it and what to keep, produces evidence you can read and score.
• Move one round from description to work. Twenty minutes of real occupational material beats an hour of narrative. The candidate cannot rehearse a task they see for the first time in front of you.

Polish stops carrying information the moment the round stops rewarding it. The same trade appears one stage later, where every take-home comes back clean and well-formatted: grading polished take-homes is that problem, and the fix has the same shape: score the decisions the work reveals, not the finish on it.

See how it works >