AI in the Interview Room: 7 Ways to Screen Candidates Fairly

The panic in hiring right now is about candidates using AI during live interviews. Teams are buying detection tools, adding proctoring, and treating the problem as a surveillance question when it is mostly a question about what they chose to ask.

This article covers what the numbers do and do not show, why detection alone makes the problem worse, and seven changes that make a screen harder to game without making it hostile to the people you want to hire.

What the Numbers Show

An analysis of 19,368 AI-run interviews found 38.5% of candidates flagged for AI-assisted cheating, rising to 48% in technical roles. Separately, CodeSignal reports that cheating and fraud-attempt rates on proctored technical tests climbed from 16% in 2024 to 35% in 2025.

Both figures come from vendors who sell detection, which does not make them wrong and does mean they should be read as directional rather than definitive. No independent audit of these rates exists yet.

In the same dataset, using an interview score of 7.0 or above as the pass threshold, 61.1% of flagged candidates would still advance. A flag is thus not a decision, and most teams have not worked out what they would do with one.

1. Separate Cheating From Assistance

Most teams have no written definition of what they are prohibiting, which makes enforcement arbitrary and appeals impossible. A candidate who used AI to prepare is doing something different from a candidate reading generated answers off a second screen during a live round.

Write the line down and make it specific to each stage. Preparation assistance is usually fine and often unavoidable; take-home work sits in a grey area that your policy should resolve explicitly, and live generation during a timed round is what most teams object to.

What to do: Publish what is permitted clearly for every stage. Anything you have not written down cannot be fairly enforced later.

2. Stop Treating a Detection Flag as a Verdict

Detection tools report a probability, and teams keep reading that probability as a finding. The gap between those two things is where unfair rejections happen.

The false-positive burden is not distributed evenly either. Detection systems trained on fluent written or spoken English tend to flag non-native speakers at markedly higher rates, which means an aggressive detection policy becomes a filter on language background rather than on honesty.

What to do: never reject on a flag alone. Treat a flag as a trigger for a human to look at the evidence, and build an appeal route that a candidate can genuinely use.

3. Change What You Ask Before You Change How You Watch

The strongest signal in the available evidence is about supervision. Unproctored assessments produce score gains more than four times larger than proctored ones, which tells you that unobserved take-home work is the softest target in most processes.

That points somewhere more useful than buying surveillance. Generic questions with publicly available good answers are easy to generate convincingly, while questions anchored to a specific candidate's own history are not.