Structured interviews: a clean screen can still be undone in the room

The argument people expect is the fairness one, so lead with the other one. Structured interviews - the same job-related questions, in the same order, scored against a defined rubric with evidence - predict job performance substantially better than unstructured conversation. That has been the consistent finding of selection research for decades, and it is the reason to do it even if nobody ever audited you.

• It is the same practice that removes the room for bias. The thing that makes an interview predictive - a defined question, a defined scale, evidence - is the same thing that leaves no gap for impressions to fill. You do not choose between fair and good. You get both from one change, exactly as the b4 levers did.
• An unstructured interview mostly measures rapport. Free conversation drifts toward whoever is easiest to talk to, and easy-to-talk-to travels with shared background. That is polish again, wearing a friendlier face.
• Order matters more than people think. If Dara asks the hardest question first to one candidate and last to another, their answers are not comparable, and the difference will be read as ability.

• The "quick extra question" for one candidate. Usually asked because the interviewer liked them and wanted to give them a chance to shine. It is a favour that makes the whole slate uncomparable.
• Scoring in the room. A number written while the person is still answering is an impression with a decimal point. Capture evidence live, score afterwards.
• The panel that debriefs before anyone writes. The loudest voice sets the anchor and the rest of the room converges on it. Everyone writes independently first, then the panel talks.
• Anchors made of adjectives. "Excellent" at level 5 and "adequate" at level 3 defines nothing, so raters go back to instinct. Anchors have to describe something you could watch happening.
• The kit that exists but nobody opens. A guide in a shared drive is not a structure. It has to be the thing on the screen during the interview, with the evidence fields being typed into.

Without calibration, two interviewers use the same 1-5 scale to mean different things. That gap is not empty. It fills with impressions - warmth, familiarity, fluency, how much the candidate sounded like the last good hire - because when a scale is undefined, something has to do the deciding, and instinct volunteers immediately.

• It is the same disease as an undefined criterion. In b4 the criterion "strong communication skills" quietly measured background because nobody defined it. An uncalibrated 1-5 scale does exactly the same thing one layer down. Undefined is never neutral.
• The failure is not usually prejudice. It is arithmetic. Two honest people applying two private scales produce one incomparable spreadsheet, and the spreadsheet is what the decision gets made from.
• Range matters as much as average. A rater who only ever uses 3 and 4 washes out every real difference on their panel. A rater who swings 1 to 5 dominates the average. Both distort a slate without either being wrong about a single candidate.
• It compounds across a wave. With 120 roles running, an uncalibrated panel is not one bad interview. It is a systematic tilt applied several hundred times.

You do not need a workshop. You need one sample, four people, and the discipline to reveal at the same time.

• Draft the guide from the scorecard. Give it the b2 capabilities and it will produce a first set of behavioural questions in an hour instead of an afternoon. You then cut half of them, which is the part only you can do.
• Generate probing follow-ups per capability. This is the one interviewers feel immediately. Two good probes per question mean a nervous candidate gets a second route into the answer, which is a fairness gain and a signal gain at the same time.
• Turn interviewer notes into evidence against each criterion. Not into a score. Take a page of scrappy notes and reorganise it so each criterion has the quoted evidence sitting under it, with "no evidence captured" written plainly where there is none. The blanks are the useful output.
• Check a guide for questions that stray into protected territory. Family plans, health, age signals, nationality, religion, "where are you really from". Interviewers rarely intend these and reliably ask a few. Screening the guide before it is used costs nothing.

• It does not score a candidate. Not out of 10, not out of 5, not "on balance a strong yes". A model that scores an individual is substantially assisting a hiring decision, and that is the trigger that pulls a whole set of obligations toward you - NYC's Local Law 144 duties apply at interview and offer stage too, not only at the screen.
• It does not summarise a person into a recommendation. "Overall an excellent culture fit" is a decision wearing the clothes of a summary. Evidence per criterion, always. The judgment is a human sentence with a human name on it.
• Nothing gets recorded without consent and a policy check. An interview recording or a live transcription tool is candidate personal data, it may be an AEDT depending on what it feeds, and in several jurisdictions recording without clear consent is its own problem. Ask before the first interview, not after the twelfth.

If a tool substantially assists a hiring decision it is in scope of the LL144 duties: an independent annual bias audit, a public summary of it, and at least 10 business days' notice to the candidate, with penalties from $500 per violation and up to $1,500 per day for continuing ones. This is education, not legal advice, and obligations vary by jurisdiction - check your own position before you deploy anything in an interview room.