Interview preparation support

Interview preparation starts with the same approved criteria used in the job description and screening. Decide which criteria an interview can assess and which require another method. A coding task may reveal implementation skill better than a conversation. A verified license may be checked from records. An interview can explore judgment, communication, planning, or reflection when questions and evaluation anchors are well designed.

For each criterion, define:

1. the capability being assessed;
2. a primary question;
3. permitted probes;
4. positive and negative evidence indicators;
5. what would remain unknown after the answer;
6. accommodations or alternative response formats.

Do not ask AI to derive questions from a candidate’s presumed personality, photograph, social profile, name, or protected information. Personalization should mean clarifying relevant evidence in the submitted material, not conducting hidden profiling.

A bounded prompt could say:

Using the approved criterion and role context below, draft four behavioral or situational interview questions. Each must be answerable using paid work, education, caregiving, volunteering, or personal projects. Avoid questions that solicit protected or unrelated personal information. For each question, explain the evidence it seeks, offer two neutral probes, and identify one limitation. Do not create a score or infer personality.

The output is a candidate list, not an interview pack ready for use. Review for clarity, job relevance, duplication, cultural assumptions, and unnecessary jargon. Questions about “working under pressure” can become vague invitations to reward dramatic storytelling. Specify the work condition and the decision or behavior of interest.

Situational questions ask what someone would do; behavioral questions ask what they did. Both have limits. A polished hypothetical answer is not evidence of past performance, while past opportunities are shaped by access and context. Use multiple evidence sources rather than treating one response as a complete picture.

An interview score should represent evidence against a criterion, not general charisma. Define anchors in observable terms. For a planning criterion:

• Strong evidence: identifies dependencies, prioritizes with reasons, communicates trade-offs, and revises the plan when conditions change.
• Partial evidence: describes a plan and priorities but gives little evidence of adaptation or communication.
• Insufficient evidence: answer is too general to evaluate, even after neutral probes.

“Insufficient evidence” is safer than converting uncertainty into a low capability judgment. Panelists should record concise evidence and then evaluate it. If they choose a rating, they should be able to point to the answer that supports it.

AI can format interviewer notes or map passages to criteria, but the interviewer must verify the mapping. The assistant may erase hedging, turn collective work into individual achievement, or produce a coherent narrative that the candidate did not actually give.

Structured interviews ask candidates the same core questions under comparable conditions. This improves comparability, but consistency does not require ignoring accessibility. Reasonable accommodations can alter timing, medium, breaks, or response format while preserving the capability being assessed.

Provide panelists with the question order, allowed probes, time guidance, prohibited topics, note-taking rules, and a route for unexpected disclosures. If a candidate volunteers health or family information, the panel should redirect to job-related evidence and follow approved handling procedures rather than entering sensitive detail into an AI tool.

Candidate-specific follow-ups can be appropriate when they clarify submitted evidence. “What was your contribution to the migration described in your application?” is relevant. “Your address is far away; do you have children who make travel difficult?” is not. Ask all candidates the same lawful question about an actual travel requirement.

If approved transcription or note support is used, disclose it and follow applicable consent and privacy requirements. Audio quality, accents, overlapping speech, and technical vocabulary can produce unequal transcription errors. The transcript is not an authoritative record of intent.

Ask AI to preserve quotations, mark uncertainty, and separate evidence from interpretation. Never infer emotion, honesty, enthusiasm, mental state, or personality from facial movement, voice, word choice, or response speed. Such signals are context dependent and may disadvantage disabled or neurodivergent candidates and people communicating across languages.

Panelists should complete independent evidence notes before group discussion to reduce conformity. Then they can reconcile differences against the same criteria. AI-generated summaries should not become the opening “truth” that anchors everyone else.

Interview quality also depends on what candidates receive. Use a human-reviewed template to explain format, duration, participants, broad assessment areas, technology, accommodations, and any AI-supported processing. Do not promise confidentiality or data deletion beyond the organization’s actual practice.

Give candidates a practical contact for questions and corrections. If technology fails, switch to a documented fallback rather than penalizing the candidate. A process is not fair merely because every person encounters the same broken tool.