We do not ban AI. We test the judgment it still requires.

Faced with AI-assisted candidates, many teams reach for a rule: no AI during the interview. It is an understandable instinct and a losing one. The ban is unenforceable without surveillance you should not deploy, and even if it worked, it would test the wrong thing. The job these candidates are interviewing for involves AI every day. An interview that forbids it is measuring a skill the role does not require: working without the tools the role provides.

Abstinence is not the skill

Think about what a strong engineer actually does with AI now. They use it to draft, then they read the draft critically. They notice the plausible-looking function that has a subtle concurrency bug. They reject the suggested approach because it does not fit a constraint the model did not know about. They ask it to defend a choice and recognise when the defence is hand-waving. The valuable skill is not avoiding the tool. It is directing it, inspecting its output, and owning the result.

A candidate who says, I would use a model for the first draft, then verify the boundary conditions and the failure modes myself, is showing you exactly the judgment the job needs. Under an AI ban, that candidate looks identical to one who cannot work without a model at all. You have erased the distinction that matters most.

How to test judgment directly

If the skill is judgment over AI output, then test that, openly. One of the sharpest interview mechanics available now is to hand the candidate a plausible but flawed AI-generated solution and ask them to critique it. Not to write it from scratch under a webcam, but to do the thing the job actually requires: look at machine-generated work and find what is wrong with it.

1. 1Show a solution that looks right and contains a real flaw: a wrong assumption, a missed edge case, a subtle correctness bug, a trade-off that does not fit the stated constraints.
2. 2Ask the candidate to evaluate it as if a teammate had opened the pull request.
3. 3Watch whether they find the flaw, explain why it matters, and propose a fix that holds up.
4. 4Change a constraint and ask whether their critique still stands.

This is impossible to game with more AI, because the task is already about AI output. The candidate cannot ask a model to tell them what is wrong with the model's work without exercising exactly the judgment you are trying to measure. The interview and the job have converged: both are about steering and checking, not about producing from nothing.

Why this is the durable position

Banning AI puts you in permanent conflict with reality, and reality keeps winning. Detecting AI puts you in an arms race you cannot finish. Testing judgment over AI puts you in step with where the work is going. As models get better, the value of a person who can direct and correct them goes up, not down, and an interview built around that skill gets more predictive over time, not less.

A candidate who uses AI well is a strong signal, not a disqualified one.

So we do not ban AI, and we do not try to catch it. We assume it is in the room, because it is in the job, and we build probes that test the one thing it cannot do for the candidate: exercise judgment about its own output. That is the skill you are hiring for. It is also the only honest thing left to test.