Does Practicing With an AI Interviewer Actually Prepare You for a Human One?

What the AI mock-interview study actually found

The paper — "Virtual Interviewers, Real Results," by Nathalia Gomez, Sue Batham, Tiffany Do (Drexel University), and Matias Volonte (Clemson University) — put 20 computer-science students through a 50-minute session with a conversational AI interviewer built on GPT-4o, paired with a monitored code editor and real-time feedback. Each student solved the same LeetCode-style problem out loud while the AI asked questions, the way a real technical interviewer would.

The self-reported results were genuinely good: 80% found the AI's conversational style realistic, 80% found the session useful, and 65% said they'd use it again. Most tellingly, 60% said the absence of a judging human made the experience less frightening than a real interview — which is exactly the kind of low-stakes rehearsal space deliberate-practice research says is valuable for building fluency before you take a skill somewhere that counts.

• What people said it helped with: practicing articulating their thought process out loud, and getting used to being watched and timed while solving a problem.
• What people wanted that it didn't yet have: natural conversational timing (85% flagged lag or awkward pacing), visual/interactive elements like running code, and adjustable difficulty.
• What the study did not measure at all: whether any of this changed how someone actually performed in a subsequent real interview. There was no before/after comparison and no outcome data — the authors call the work "preliminary" and "formative," not conclusive.

“80% said the AI session felt useful. 0% of that usefulness was ever checked against a real interview outcome — the study didn't measure that at all.”

The actual question: does practice on one task transfer to a different one?

This isn't a new problem invented by AI. Psychologists have been studying when a skill learned in one setting carries over to another since 1901, when Edward Thorndike and Robert Woodworth proposed what's still called the identical elements theory of transfer: practice on task A improves performance on task B to the extent that A and B share actual common elements — the same mental operations, the same stimulus-response patterns — not just a surface resemblance.

That theory predicts, directly, which parts of AI interview practice should transfer and which shouldn't. Structuring an answer, retrieving a relevant example under time pressure, narrating your reasoning instead of going silent while you think, staying composed while someone is evaluating you — those are mental operations the AI session shares with a real interview, so practicing them against an AI should genuinely help. Reading a human interviewer's non-verbal reactions, adjusting your pitch in real time when you sense you've lost them, building actual rapport with a specific person in the room — those elements are mostly absent from a text-and-voice AI exchange, so there's no theoretical reason to expect practice against an AI to improve them.

This is the same reason repeated, structured mock-interview practice works at all — the benefit comes from rehearsing the specific operations a real interview demands, not from the mock session merely existing.

When someone actually tested transfer, it worked — for a narrower tool than you'd think

There is a body of research that did test real-world transfer, not just self-reported confidence — it's just a different system than a general-purpose chatbot interviewer. Virtual Reality Job Interview Training (VR-JIT), developed at Vanderbilt and tested in multiple randomized controlled trials, is a structured simulation where a user practices answering a fixed set of interview questions posed by an on-screen virtual character, with coded feedback on specific response behaviors.

In trials with young adults on the autism spectrum, published in Psychiatric Services and the Journal of Vocational Rehabilitation, VR-JIT training was linked to measurably better job-interview role-play performance scored by independent raters, and to real differences in vocational outcomes months later. That's a meaningfully stronger result than "participants said they felt more confident" — it's outcome data, collected after the fact, from people who weren't just asked how the practice felt.

The honest caveat: VR-JIT is a narrower, far more structured tool than an open-ended GPT-4o conversation, it was tested specifically with autistic young adults rather than the general population, and its gains showed up on a fixed, rehearsed question set rather than free-flowing conversation. It's evidence that AI-mediated interview practice can transfer to real-world results under the right conditions — not proof that any AI mock interview automatically will.

“The tool with proof it transfers is narrower and more rehearsed than most AI mock interviews. The tool that feels the most realistic is the one nobody has actually tested for transfer.”

Where practicing with an AI could plausibly teach you the wrong habit

Identical elements theory cuts both ways. If an AI interviewer never interrupts, never visibly loses interest, and responds to literally everything in a flat, patient register, a candidate who trains exclusively against it could reasonably get worse — not better — at reading and adjusting to a real human's actual reactions, simply because that skill never gets exercised.

The participants in the Drexel/Clemson study noticed a version of this themselves: the conversational lag and unnatural pacing were the single most common complaint (85% of participants), which means the AI's rhythm was noticeably different from a real conversation's rhythm — and rhythm is one of the "identical elements" a live interview actually tests.

None of this is a reason to avoid AI practice. It's a reason to treat it as one input, not the whole preparation plan — the same way a batting cage helps a hitter groove mechanics but doesn't replace ever facing a live pitcher.

What this actually means for how you should practice

Put together, the research points to a specific, usable split — practice the elements that transfer against an AI, and deliberately get at least some of the elements that don't from a human.

• Use AI practice for what it's good for: structuring answers, rehearsing your reasoning out loud, getting repetitions in on common questions until retrieval is fast, and getting comfortable with being watched and timed — all identical elements that a real interview also demands.
• Don't rely on AI practice alone for reading a room, adjusting tone to a specific interviewer's reactions, or recovering gracefully when a human visibly disagrees with you — get at least one or two live run-throughs with a person, even an informal one, before anything that matters.
• Treat your own felt confidence after an AI session the way the Drexel/Clemson study's authors did: as a real, useful signal about anxiety reduction, not as proof you've improved at the thing you'll actually be judged on.
• If a tool gives you scored, specific feedback on structure and content rather than just letting you talk, lean on that harder — feedback quality is one of the better-evidenced levers in all of deliberate-practice research, independent of who or what is delivering it.

Whether practice helps at all, independent of the AI question, is its own debate worth understanding — repetition with feedback is the mechanism that does the actual work, and that holds whether the feedback comes from a human, an AI, or your own review of a recording.

“Practice the parts an AI and a human interview actually share. Don't assume the parts they don't share get trained for free.”

Key Takeaways

If you remember only a few things, remember these:

• A 2025 study (Gomez, Batham, Do & Volonte, Drexel University and Clemson University) found a GPT-4o mock technical interviewer reduced anxiety and felt useful to 80% of 20 student participants — but never measured whether that translated into better performance in a real interview.
• Whether a practiced skill transfers to a new context is governed by identical elements theory (Thorndike & Woodworth, 1901): practice helps to the extent the training task and the real task share actual mental operations, not just surface similarity.
• VR-JIT, a more structured AI-based interview simulation, has randomized controlled trial evidence of real transfer to interview performance and vocational outcomes — but it was tested with a narrower population (autistic young adults) and a more rehearsed, structured format than an open-ended chatbot interview.
• Elements that plausibly transfer from AI practice: structuring answers, rehearsing reasoning aloud, speeding up retrieval, tolerating being watched. Elements that plausibly don't: reading a human's real-time reactions, adjusting tone to a specific person, recovering from genuine disagreement.
• The practical fix isn't choosing AI or human practice — it's using AI for repetitions on the first group and getting at least one real human run-through for the second.