We Use AI All Day. Why Are We Banning It in Interviews?

Your team uses AI to draft, code, research or analyse. A candidate opens the same kind of tool during an interview and the panel suddenly calls it cheating.

That reaction may expose an unclear assessment. Decide which capability you need to observe, choose conditions that reveal it, and tell every candidate the rules before the exercise. “AI allowed” and “AI banned” are both incomplete policies.

A recent LinkedIn discussion among engineering interviewers includes teams allowing familiar AI tools, reviewers using a fixed AI-generated diff, and others separating assisted execution from independent fundamentals. These are practitioner examples, not comparative evidence that one format predicts job performance.

Start with the thing you need to know

Different questions require different conditions:

The conditions follow the question. A polished answer produced with unlimited assistance cannot show unaided recall. An artificial memory test cannot show how someone verifies AI output during the actual job.

Use a fixed output when comparability matters

If every candidate asks a live model for help, each may receive different assistance. That variation makes comparison harder.

For a marketing coordinator, provide the same fictional AI-generated customer email containing a wrong price, an unsupported promise and a vague next step. Ask candidates to mark what they would verify and produce a corrected version using supplied source facts.

For a developer, provide the same small diff containing a missing authorization check, an unrelated change and an untested edge case. Ask what they would investigate before accepting it.

These exercises assess review and judgment under common conditions. They are still proposed exercises, not validated predictions of performance.

If candidates may use AI, observe the whole loop

Do not score prompt theatrics. Watch whether the person:

• Clarifies the task before generating material.
• Protects confidential or personal data.
• Checks claims against available sources.
• Notices when the output solves a different problem.
• Tests or revises the result.
• Can explain the final decisions in their own words.

Give candidates suitable data. Do not ask them to paste real customer records, private code or applicant information into an unapproved service.

Publish the rule with the invitation

Copy and adapt this:

This exercise has two parts.

Part 1: Please answer without AI or external assistance. We are checking [specific capability].

Part 2: You may use [permitted tools]. We are checking how you inspect, verify and improve the result. Do not enter confidential or personal data. You will be asked to explain your decisions.

We provide the same task, time and source material to every candidate. If you need an accessibility adjustment or clarification, contact [person/route].

If the role expects daily AI use, say which tools are available at work and what review responsibility remains with the employee. If tool use is restricted for security, client or regulatory reasons, explain the relevant boundary.

Record evidence instead of enforcing a vibe

After the exercise, record:

Capability being assessed:
Tools and source material available:
Candidate's approach:
Result they verified:
Material issue they missed:
Explanation or follow-up:
What remains unknown:

Our AI-skills exercise for non-experts uses a flawed draft and source facts because the owner can judge the corrected customer result. The same principle applies here: choose work whose quality you can inspect, and bring in a specialist when you cannot.

The fairest policy is the one candidates receive in advance and interviewers can connect to the job. Surprise rules mostly measure whether someone guessed the panel's private expectations.