Say What They Should Not Tell You, Then Ask What They Decided

What does a confidentiality agreement actually take off the table?

Names and specifics, most plainly. The client's identity, the unreleased product, the dataset, the internal document and the figures inside it are what a confidentiality agreement is written to protect. A sentence shaped like this one names none of them: the model drafted the segment analysis, and the candidate threw it out because the base year was wrong. Ask for that kind of sentence and almost nothing you wanted disappears.

The candidate cannot decide on the spot what the agreement covers, and should not be asked to.

They may not have read the agreement in a year, may not know whether a paraphrase counts, and the safe answer under time pressure is always to say less. So an interviewer who leaves the boundary undefined collects the most conservative version of every answer, starting with the most careful candidates.

Whatever you improvise to get around the silence is still a selection procedure.

The 1978 US federal Uniform Guidelines on Employee Selection Procedures define the term to cover the full range of assessment techniques, from paper-and-pencil tests through what the text calls informal or casual interviews and unscored application forms ¹. The Guidelines attach a validation burden only where adverse impact shows up, so an improvised follow-up with no adverse impact owes no validation study. Hold it to a plainer bar anyway: good enough to put to every candidate in the round.

The wall tells you where the candidate worked. It does not tell you the work was thin.

Name the ban out loud before the first follow-up

Say the rule yourself, in one sentence, before anything else is asked. No client, no product name, no numbers, no data. Then say what the round wants instead: the decision, what the model was allowed to do, and what got checked. A candidate who knows the shape of a permitted answer can give one without weighing every sentence against the agreement.

The announcement is also the last time confidentiality has to come up. Without it, the round turns into a negotiation about what is allowed, run by the person with the most to lose from getting it wrong. The time goes into deciding what can be said, and none of it into the work.

Six questions carry the rest, and not one of them needs a proper noun:

• What the decision was, at the level of a decision: a pricing call, a code review, a diagnosis of a broken pipeline.
• What the model was allowed to do, and where its output stopped and theirs began.
• What they kept for themselves, and the reason it was that part and not another.
• What they checked the model's claim against, and what the check turned up.
• What they refused to ship, and on what grounds.
• What changed as a result, described without a figure.

Ask the same six of everyone in the round, and the confidentiality case stops being a special one.

In a 2022 re-analysis of the selection literature, structured interviews correlate with supervisor ratings of job performance at an estimated .42, against .19 for unstructured ones, and the observed figures before any statistical correction were .32 and .13 ².

Structured, in that comparison, means a research coding of format: the same questions rated on a common scale. None of the underlying studies involved an interview about AI. The gap sits between two interviews rather than between two methods.

Running a structured AI-use interview so candidates are actually comparable is the general version of asking six fixed questions, and the confidentiality case is where improvising is most tempting.

The follow-up is where an abstraction gets tested. Push one level down on the check: what would have had to be true for the model's answer to hold, and what they would have opened next. Follow-up questions expose whether someone understands the answer they gave, and neither of those two questions needs a name.

Swap in a case you own when the abstraction stops being scorable

Hand over a case at the point where two different candidates would give the same answer. Generality is the tell. If the description would fit any analyst at any firm, the round has stopped collecting evidence about a person and started collecting fluency. Stop asking about their work, put a brief in front of them, and give it fifteen minutes.

Say why the swap is happening, so it does not land as a penalty.

The brief is short, drawn from the occupation, and comes with an AI assistant willing to do the whole thing. One thing in the source material is wrong in a way only checking catches. The anchors are the ones the six questions were aimed at.

Write down what an acceptable answer looks like before the session starts. The US Office of Personnel Management's 2008 practical guide defines a structured interview by three properties: the same questions in the same order, a common rating scale, and interviewers who agree in advance on what an acceptable answer is ³. It is a federal agency's guidance for federal hiring rather than statute, and it binds no private employer.

The third property is what lets the substitute case be scored the same way for everyone who gets it.

Build the case out of material you own and out of nothing a customer gave you. The brief goes into the assistant by design, which is the general rule made literal: anything you send a candidate goes into a model. Synthetic figures in a real shape are quick to write and remove the problem.

What if the candidate volunteers more than they should?

Stop them, and write down that you stopped them. A candidate who names the client has put something on your side of the table that you did not ask for, and the hiring record is not the place to keep it. One sentence does it: no names are needed here, describe the decision instead.

Federal trade-secret law defines the exposure narrowly, and the definition can reach the person on the receiving end.

Title 18, section 1839 of the United States Code defines misappropriation to include acquiring a trade secret of another when the person acquiring it knows or has reason to know it was acquired by improper means ⁴. The same section defines improper means to include breach or inducement of a breach of a duty to maintain secrecy.

Inducement is the word that could reach an interviewer who kept pressing after being told an agreement existed, if what came out was a trade secret. None of that is legal advice, and the question belongs with counsel.

Somebody else's material turns up in a different shape when a candidate says they pasted patient records into a chatbot. Both answers are about a boundary rather than about a tool, and the boundary is the part that transfers to your job. Holding one under pressure, in a room where breaking it would have been rewarded, is the demonstration.

Someone who ran nothing can describe a decision sequence in the abstract about as well as someone who ran it, and the six questions cannot separate them. A confidentiality agreement is also a good place to hide a thin record. The substitute case is the only part of the round that does not rest on the story.

Never write a held boundary down as a gap

Write what happened, not what it might mean. One note says the candidate declined to name the client, citing a confidentiality agreement, and answered all six decision questions. Another says they were vague on specifics. The first is a fact. The second is an inference, read months later by somebody who was not in the room.

The content of a note carries signal of its own. Shanshi Liu and colleagues text-mined the notes interviewers wrote on 7,650 candidates hired at one large Chinese internet technology company. The number of job-related capabilities a note named predicted later performance and promotions, and predicted lower turnover ⁵.

The sample is people who passed the interview and were hired, so it says nothing about the rejected side. The effect is small too, near 2 percent of a performance rating per standard deviation of match to the job analysis.

So name capabilities in the note:

Framed the problem before generating. Demanded a source for the claim the recommendation rested on. Kept the judgment that should not have been handed over. Refused an output, on stated grounds.

The candidate has the harder half of this, and the good version from their chair is worth knowing: describe the decision, not the data, and NDA work still counts.

The example they could not tell you about is still the one they reached for first. Its worth was in which decision they thought was the interesting part, and in a line they held for a document you will never read. Ask the next candidate the same six questions and find out how much of the difference was the contract.

See how it works >