AI Skills Interview Questions to Ask

Most interviews are still designed to test whether a candidate can produce work from scratch, alone, under pressure. In 2026 that tells you less than it used to, because the candidate will not be working that way once they start. They will be working with a model in the loop, and the thing you actually need to know is whether their judgment is good enough to direct it and catch it when it is wrong. This guide is about how to test AI fluency and judgment in a real interview, with example questions you can adapt. It pairs with the broader hiring guides on the AI Career Index, and with the companion piece on what AI skills to require before you even get to the interview.

Let them use AI live

The first change is the hardest for many interviewers to accept: stop banning the tools in the room. If the job involves working with AI, an interview that forbids it is testing a fiction. You learn nothing useful by watching someone do a task the slow way that they will never do slowly on the job. Worse, you select for the wrong people, the ones who memorised answers, over the ones who know how to get a model to a good answer fast and then verify it.

So open the laptop. Give the candidate a realistic task, tell them to use whatever tools they normally use, and watch how they work. The signal is not the final artifact. It is the process: how they frame the prompt, how they react when the first output is wrong, whether they check the result before they hand it over. A strong candidate treats the model like a fast, confident, occasionally wrong colleague. A weak one treats it like an oracle and pastes whatever it says.

Probe a real AI-assisted task

Generic prompts produce generic signal. The better approach is to hand the candidate a task that mirrors the real job, with a deliberate trap built in. The trap is the point: it is where you find out whether they verify or just trust.

For an engineering role, give them a function the model is likely to write subtly wrong, an off-by-one in an edge case, or a security hole in input handling, and ask them to deliver working, reviewed code. The good candidate uses AI to draft it, then reads it critically and finds the flaw. This is the same instinct you screen for when hiring a software engineer: can they review AI output as rigorously as they would a junior colleague's pull request?

For a customer-facing role, give them a tricky customer message and let them draft a reply with AI. The trap is that the literal answer is policy-correct but tone-deaf, or technically right but commercially wrong. The good candidate edits for the human reality the model missed. This maps directly to hiring a customer service representative who can use AI for speed without outsourcing their judgment about the actual person on the other end.

In both cases you are not testing whether they can use the tool. You are testing whether they can tell when the tool is wrong, which is the skill that does not commoditise.

Test judgment over output

Output is cheap now, so stop rewarding it. The candidate who produces the most polished deliverable in the room may simply be the one who trusted the model most, and that is a risk, not a strength. What you want to grade is judgment: the choices behind the output, the things they chose not to ship, the moments they decided the stakes were too high to automate.

The way to surface judgment is to ask about the reasoning, not just accept the result. After the live task, walk back through it. Where did you not trust the model, and why? What did you change in its answer, and what would have happened if you had shipped the original? When would you not use AI for this at all? The answers separate someone with real judgment from someone who got lucky that the model was right this time.

Example questions to ask

These are starting points. Adapt the specifics to your role, but keep the shape: each one forces the candidate to reveal verification habits and judgment rather than tool familiarity.

• •Walk me through a time AI gave you a confident answer that was wrong. How did you catch it, and what changed in how you work because of it?
• •Here is a task and a laptop. Use whatever tools you want. I am going to watch how you work, not just the result.
• •You used the model to produce this. How would you check it is right before you put your name on it?
• •What part of this job would you refuse to hand to AI, and why?
• •Show me a prompt you would actually write for this, then tell me what you would do if the first answer was 80 percent right.
• •When does using AI here make the work worse, not faster?

The last two are the most revealing. A candidate who can articulate where AI degrades the work, and who instinctively plans for the imperfect first answer, has internalised the way the job actually runs now.

Score the process, not the artifact

When you debrief, resist scoring the finished work. Score the four things that predict on-the-job performance: how clearly they directed the model, how fast they recovered when it was wrong, whether they verified before delivering, and whether they showed judgment about when to use it at all. A messy final artifact with strong process beats a clean one produced on blind trust, because the first candidate will keep catching errors and the second will eventually ship a confident, wrong answer into production.

If you want to calibrate what good looks like for a specific role, the workforce benchmarking tools on the platform show where the durable value sits in each job, which tells you which of the four signals to weight most. For roles built on relationships and trust, judgment dominates. For roles built on technical correctness, verification dominates.

The flip side: knowing what to avoid

Testing for the right signals is half the job. The other half is recognising the warning signs early, the candidate who cannot explain the output they produced, who never verifies, who treats the tools as a threat to be defended against. Those patterns are worth knowing before you walk into the room, and they are laid out in the companion guide to hiring red flags in the AI era. Read the two together and your interview becomes a filter that selects for judgment and screens out blind trust, which is exactly the trade you want to be making in 2026.

Run your structural assessment

The AI Career Index is free to take. Get your structural snapshot in 10 minutes, including your durability score, archetype, and projection.