Should Candidates Be Allowed to Use AI in Interviews?

By Victor Hoang · Updated Oct 10, 2026

Most leadership teams have a policy on this, whether or not they wrote it down. Some say "no AI in the interview." Others say nothing and hope. Both positions share a flaw: they treat AI use as a compliance question when it's really a design question.

Our view is simple. A ban is unenforceable at any reasonable cost, and it measures a version of the job that no longer exists. The better design lets candidates use AI openly on a realistic task, then evaluates what AI can't do for them: judgment, verification, and reasoning under questioning. That's a harder interview to build. It's also a more honest one.

Key Facts

• Structured interviews had the highest mean operational validity (r = .42) in the 2022 Sackett et al. re-analysis of selection methods, though with wide spread around that mean.
• In an April 2025 Blind survey of 3,617 professionals, 20% of those who answered the question said they had secretly used AI tools during interviews, and 55% agreed that using AI in interviews has become the new norm.
• In a Checkr survey of 3,000 US managers, only 19% were extremely confident their current hiring process would catch a fraudulent applicant.
• Canva told engineering candidates in 2025 that it would expect them to use AI assistants in technical interviews, because policing covert use was increasingly difficult.
• Meta confirmed in July 2025 that it was developing coding interviews where candidates can use an AI assistant, saying the format is more representative of the developer environment and makes LLM-based cheating less effective.

Why the ban is the weaker position

Start with enforcement. The Blind figure is a self-selected sample of its own app users, so it's not a census of the workforce. But it points in a direction hiring teams already feel. If one in five respondents will admit to secret use in an anonymous poll, the real rate in a given interview pool is unknown and probably not zero.

Then look at the detection side. The Checkr managers aren't confident either: only 19% said they were extremely confident their process would catch a fraudulent applicant, and 59% said they'd suspected a candidate of using AI to misrepresent themselves at some stage. Suspicion without reliable detection is the worst combination. It makes interviewers feel like investigators, and it punishes the honest candidate who follows the rule while a less scrupulous one doesn't.

So a ban creates a selection effect nobody intended. You filter for rule-followers over capable people, and you can't verify the rule is being followed.

The second problem is validity. Ask what the interview is supposed to predict. If the job involves working with AI tools every day, then an interview that forbids them predicts performance in a setting the new hire will never work in. Canva's engineering team reached this conclusion in 2025. They wrote that dismissing AI tools meant they weren't truly evaluating how candidates would perform in the actual role, and that AI assistants could trivially solve their traditional coding questions.

What the evidence on interviews says to keep

None of this means "anything goes." The strongest research we have on interviews points to what to protect.

The 2022 paper by Sackett, Zhang, Berry and Lievens, published in the Journal of Applied Psychology, re-examined how earlier meta-analyses corrected for restriction of range. Their conclusion, as summarized by HumRRO, was that structured interviews came out as the strongest single predictor of job performance, with a mean operational validity of .42. Cognitive ability tests, long treated as the benchmark, no longer held that position.

Two lessons follow, and they pull in the same direction.

First, structure is what does the work. Same questions, anchored rating scales, multiple independent raters. That doesn't change when a candidate has an AI assistant open. It becomes more important, because structure is how you stop a polished AI answer from reading as a strong signal.

Second, the same summary notes that structured-interview validities vary quite a bit around that mean. A structured interview aimed at the wrong construct isn't automatically valid. If your scoring rubric rewards fluent, well-organized answers, you're now measuring access to a model. If it rewards decisions, trade-offs, and error-catching, you're measuring the person.

Three positions, compared

Employers have taken three public positions, and each makes a different bet.

We're not saying the employers who ban are foolish. A ban is defensible for early-stage screens where you want to hear a person think from scratch, and it's reasonable to ask for unaided work in a specific exercise if you say so clearly. Anthropic's own page, for instance, encourages AI for interview preparation and asks for transparency about use. The problem is the unstated, blanket ban, applied to roles where the daily work is AI-assisted.

The Open-Task Interview: a design you can run

Here's the mental model we'd use. Call it the Open-Task Interview. It has four parts, and each one has to be present or the design collapses into theater.

1. Declare the rule up front. Tell candidates in the invitation which tools are allowed and which exercises are unaided. Ambiguity is where unfairness lives. Canva tells candidates in advance and recommends they practice with the tools beforehand.

2. Use a task AI can start but can't finish. The task should be realistic and a little underspecified. An illustrative example: give a finance candidate a messy vendor contract and a two-page brief, and ask for a recommendation to the CFO. A model will produce a tidy summary in seconds. It won't know which clause your company has been burned by, and it may state a wrong figure with total confidence.

3. Score the process, not just the output. Ask the candidate to talk through what they asked the model, what they rejected, and what they checked. Canva's post frames the goal as seeing how well candidates collaborate with AI and says they must own the result, whoever wrote it.

4. Probe with follow-ups the model can't answer for them. Change a constraint mid-interview. Ask "what would make you change this recommendation?" The candidate who understands the work adapts. The one who pasted an answer stalls.

What to score

A rubric for an open-task interview should have four dimensions, each with anchored levels (for example, 1 to 5, with a written description of what a 1, 3 and 5 look like).

Notice what's absent: speed of typing, polish of prose, and recall of facts. Those are the things AI hands out for free. The scarce skill is knowing when the output is wrong, which connects to a theme we've covered in keeping human judgment sharp and to the argument that complementary skills are where the premium sits.

Where an open-AI interview goes wrong

Three failure modes show up when teams try this.

The task is too easy. If the model solves it in one prompt, you're measuring nothing. Canva's own experience was that AI assistants could trivially solve traditional questions, so they had to write harder, more open-ended ones. Expect to rewrite your question bank.

The rubric drifts back to polish. Interviewers are human. A smooth, confident presentation still wins unless the rubric forces attention onto verification and trade-offs. Calibrate raters on recorded examples before the first live round.

Unequal access. A candidate who has used a given tool for a year has an edge over one who hasn't. Either supply the tool in the session or allow a short warm-up. Otherwise you've replaced one unfair advantage with another.

There's also identity. Using AI to answer is one thing. Using it to impersonate is another, and the Checkr survey shows managers worry about fake identities: 31% had interviewed someone later revealed to be using one. Open AI use doesn't remove the need to verify who you're talking to. Keep an identity check and, for senior roles, at least one in-person or live-video conversation.

What to do next

If you run hiring or sit above it, here's a practical sequence.

CEO. Decide the principle: for roles that use AI daily, the interview should reflect the job. Ask your leadership team to state the current policy in writing. If nobody can, that's your answer about consistency.

COO. Pick one role family, ideally one with high volume and clear daily AI use, and pilot an open-task interview. Run it alongside the existing process for a quarter and compare outcomes at 90 days. Keep the task bank confidential and rotate it.

CHRO. Rewrite the rubric around the four dimensions above, train raters with recorded examples, and publish the rule in candidate communications. Keep unaided exercises where you have a real reason, and say so plainly. Pair this with your approach to AI-native new hire expectations so the interview matches what you'll ask of people after they join.

Everyone. Treat the first pilot as a measurement exercise, not a verdict. Structure is what carries the research-backed validity, so don't loosen the standardization while you loosen the tool rules.

For the screening side of the same question, see whether AI should screen job candidates, and for what happens to the written application signal, the resume problem with AI-written applications.

Learn More

• AI-Native New Hire Expectations
• Sales and Marketing Hiring in 2026: AI Fluency
• The Executive Decision Framework for AI Workforce Strategy
• Keeping Human Judgment Sharp
• Skills-Based Talent Strategy