Candidates are using AI in interviews: how to design for it

Candidates are using AI in interviews, and in every step around them. They draft résumés with it, rehearse with it, and a growing number keep it open during the conversation itself. Most hiring teams have answered with rules and suspicion: bans, disqualification warnings, a return to the office. Some of that is reasonable. The larger problem sits in the questions themselves. If a chatbot can produce a strong answer to your question, the question was never measuring much. The durable fix is to ask questions that only lived experience can answer.

How candidates are using AI in interviews today

The numbers vary by survey and by what counts as “use,” but every survey points the same direction.

• Applications. In a Gartner survey released in July 2025, about four in 10 candidates said they used AI during the application process, mostly to write résumés, cover letters, writing samples or assessment answers (HR Dive).
• Volume. LinkedIn was handling an average of 11,000 applications a minute in mid-2025, up 45% in a year, according to The New York Times (eWeek).
• The interview itself. In Resume Genius’s 2026 survey of 1,000 U.S. job seekers, 22% said they already use AI during real-time interviews (Newsweek). That figure is self-reported and comes from a single survey, so treat it as a rough signal.
• What recruiters see. In Greenhouse’s November 2025 survey of U.S. recruiters and hiring managers, 32% cited AI scripts during job interviews, and 91% said they had caught or suspected AI-driven candidate misrepresentation (HR Dive).
• The tools. Interview copilots are a product category now. One, Interview Coder, used a “hidden in-browser window that can’t be viewed by the interviewer or test giver.” Its founder was suspended by Columbia University and, in April 2025, raised $5.3 million for a startup pitched as a way to “cheat on everything” (TechCrunch).

That use falls along a spectrum:

1. Preparation: researching the company, rehearsing, tightening a résumé the candidate wrote.
2. Real-time assistance: a copilot listening to the call and suggesting answers.
3. Misrepresentation: claiming experience the candidate does not have, or having someone or something else answer in their place.

The first is ordinary. The third is fraud, and we cover proxies and deepfakes separately in candidate fraud and deepfake interviews. Most of the anxiety lives in the middle, which is also where interview design does the most work.

Policing AI use is expensive and unreliable

Large employers have mostly reached for enforcement. Amazon’s interview guidelines, as reported by Business Insider in early 2025, ask candidates: “please do not use GenAI tools during your interview unless explicitly permitted,” and warn that not adhering “may result in disqualification from the recruitment process” (ITPro). By August 2025, Google had banned AI tools in virtual interviews, and Google, Cisco and McKinsey had re-instituted in-person interviews for some candidates. In a Gartner survey cited in the same report, 72.4% of recruiting leaders said they were conducting interviews in person to combat fraud (Computerworld). Greenhouse found 39% of U.S. hiring managers running more in-person interviews to verify that candidates are authentic (HR Dive).

Clear rules help, because candidates deserve to know what is allowed. As a strategy, though, policing has three problems.

Detection does not work well. In an early-2024 experiment by interviewing.io, candidates secretly used ChatGPT across 32 usable technical interviews. Not a single interviewer raised a concern about cheating (interviewing.io). The behavioral tells interviewers are trained to watch for (typing during a question, eyes drifting, an answer that sounds read) also describe a nervous candidate taking notes. Score suspicion and you end up scoring nerves.

In-person rounds bring back the bottleneck. An onsite is the right tool for a shortlist. As a first round for two hundred applicants, it reintroduces exactly the scheduling cost that made round one the bottleneck in the first place. We ran those numbers in the math of high-volume screening.

Catching a copilot tells you nothing about competence. Even perfect enforcement leaves you with an unassisted candidate answering the same generic question, and that question still cannot separate the person who did the work from the person who read about it.

Questions a model can answer carry little signal

The interviewing.io experiment is more useful for what it found about questions than for what it found about cheating. With ChatGPT, candidates passed 73% of verbatim LeetCode problems and 67% of lightly modified ones, against a 53% pass rate for candidates on the platform working without AI. On custom questions written by the interviewers, the pass rate fell to 25%. The cheaters did worse than people who were not cheating at all, and the authors’ advice was blunt: start asking custom questions immediately.

The caveats are real: it was a small study run by a company that sells interview services, and models have improved a lot since early 2024. The lesson about questions holds anyway: the more a question resembles something already on the internet, the more AI help is worth.

The same holds far outside engineering. “What’s your greatest weakness?” “Tell me about a time you handled conflict.” A model has read thousands of good answers to both. A fluent, well-structured story is now free to produce, which means polish has stopped being evidence. Polish was a weak signal to begin with; it tracked practice, privilege and first language more than ability.

Follow-ups that need a life behind them

The first answer to any question is cheap. The signal is in the follow-up, especially one that depends on what the candidate just said:

• “Walk me through what you personally did.”
• “What went wrong?”
• “What would you do differently next time?”
• “What did you cut, and what did that cost?”
• “Earlier you said the team was four people. Who owned the migration?”

These work because they are about the candidate’s own history, asked in real time, built on their last answer. A copilot can suggest plausible text, but the candidate still has to own it, extend it two levels down, and keep it consistent with what they said five minutes ago. Someone who did the work finds that easy, and someone reading a script runs out of material fast.

Adaptive interviews still need structure, because the strongest evidence on interviews favors it: in Sackett and colleagues’ 2022 re-analysis of selection research, structured interviews had the highest mean operational validity of any predictor studied, r = .42 (SIOP). To keep that structure while following the conversation, fix the techniques in advance (which follow-ups apply to which competency) and let each answer decide where to use them. Every candidate faces the same pressure, even though no two conversations are word-for-word identical.

Then score the evidence. Every rating should cite the moment in the transcript that earned it. A fluent answer with no specifics is a thin answer, however good it sounds. We wrote up that discipline in evidence-based scorecards.

Draw the line at whose experience it is

Some AI use is harmless. A candidate who uses a chatbot to research your company, rehearse, or tighten a résumé they wrote is preparing. It is the modern version of a friend reading a draft. Some employers now say so explicitly. Anthropic’s published candidate guidance, last updated July 10, 2025, asks applicants to write their own first draft and then refine it with AI, encourages AI for research and practice, and says of live interviews: “This is all you–no AI assistance unless we indicate otherwise” (Anthropic).

That points to a clean rule. AI that helps a candidate describe their own experience is preparation. AI (or a person) that supplies experience they do not have, or answers in their place, is misrepresentation.

Say the rule out loud, in advance, stage by stage. Most candidates want to know where the line is. Then design so the line mostly enforces itself: an interview that goes deep into the candidate’s own history leaves a copilot very little to contribute.

A checklist for hiring teams

1. Publish an AI policy for candidates, stage by stage: what is encouraged, what is allowed, what is off-limits.
2. Run your own questions through a chatbot. If its answer would score well on your rubric, rewrite the question.
3. Retire generic prompts: strengths and weaknesses, textbook definitions, famous puzzles.
4. Ask for one specific instance from the candidate’s own work.
5. Pre-approve follow-up techniques per competency: a specific example, personal ownership, the failure case, what they would change.
6. Go two follow-ups deep before moving on.
7. Return to earlier threads. Record contradictions as quotes for a human to raise and judge.
8. Score evidence against a versioned rubric, with every rating citing a transcript moment.
9. Never score polish, fluency or suspected AI use. If something looks off, flag it for a person.
10. Save in-person time for the shortlist, where it is affordable and where identity can be checked too.

Where Hure lands

Hure was designed on the assumption that a polished answer costs nothing. Each candidate gets a natural 8–10 minute voice-to-voice interview in their own language. The interviewer asks follow-ups using only the techniques the hiring team approved for each competency: a specific example, the candidate’s own role, a failure case. It returns to threads the candidate opened earlier and notices when two answers do not line up, always quoting the candidate verbatim. Every score is anchored to a sealed, versioned rubric and cites the transcript moment behind it, so fluency without substance earns nothing.

Hure does not try to catch AI use. It has no eye tracking, no voice or face analysis and no emotion inference, and it scores the content of what candidates say, never how they look or sound. None of this guarantees against a well-coached candidate, and we would not claim it does. You get a scorecard that shows exactly which evidence each rating rests on, and a human makes the final call. You can see how the interview runs on the interviewer page.