AI has made it genuinely difficult to know whether a video interview is telling you anything real about a candidate. That's the short version. The longer version is that it's not just about candidates who might be bending the rules. The tools available now are good enough that a fully coached, AI-assisted candidate can look indistinguishable from someone who actually knows their stuff. And the detection tools trying to catch that aren't keeping pace. So if video interviews are carrying real weight in your hiring process, it's worth thinking hard about whether that weight is still justified.
AI-assisted cheating runs on a spectrum from preparation to impersonation
Not all AI-assisted cheating in video interviews is the same, and the different levels present very different problems.
At one end, you've got candidates who've used AI to help prepare. They've fed a job description into ChatGPT, generated a list of likely questions, and used it to craft polished answers they've then rehearsed.
This has always happened in some form.
Before AI, candidates wrote notes on sticky pads, propped revision cards against their monitor, or rehearsed scripted answers with a mate. The effort was higher, but the intent was identical. The candidate doing this is still, broadly, themselves but they've just their very-prepared self. Their answers are more considered, more structured, and probably more impressive than they'd be off the cuff. Whether that inflates your assessment of them is a real question, but it's not a new one.
The middle tier is where things start to get more serious. Real-time AI coaching tools, products like Parakeet AI and Interview Coder, listen to the conversation as it happens, generate a structured answer to whatever question the interviewer just asked, and display that answer on screen fast enough for the candidate to read it back naturally. The candidate isn't drawing on their own thinking at all. They're reading. The effort barrier that existed with pre-prepared notes has essentially gone, and what you're now assessing is something closer to literacy than capability.
At the far end, you've got full identity substitution. Not coaching, not prompting, but a completely different person sitting the interview using deepfake technology to alter their appearance. We'll come back to that tier separately because it's its own category of problem and is only really an issue for freelance and remote work.
Real-time AI coaching tools have made the interview a reading exercise
The middle tier, live coaching, is the one that most directly undermines what a video interview is supposed to do.
The way these tools work is fairly straightforward. The software listens through the candidate's microphone, transcribes the interviewer's question in near-real-time, generates a structured answer, and displays it on the candidate's screen. On the interviewer's side, you see someone looking at you, pausing briefly (like someone naturally would), then delivering a clear and well-structured response. It looks like someone thinking. It isn't.
What you're now measuring is their ability to read fluently under mild pressure, and perhaps their composure while doing it. Which, to be fair, is a skill, but it's not what you were trying to assess.
Around 83% of candidates say they would use AI assistance in a job interview if they believed they wouldn't be caught. And the tech and people's abilities to use it are only going to increase.
Deepfakes have made it possible for someone else entirely to take the interview
The third tier is stranger, and in some ways the hardest to get your head around. Deepfake technology has improved to the point where building a convincing real-time fake video feed takes just over an hour, and deepfake fraud attempts in hiring increased by around 1,300% between 2023 and 2024. In the most organised version of this, a fabricated identity is used to secure a remote role while the actual work is done by someone else entirely. The person you hired doesn't exist.
It's worth being clear that this tier is mostly a risk for remote and freelance hiring. But it does illustrate how far the validity question has shifted: it's no longer just about getting a misleading read on a candidate, it's about whether the candidate you interviewed was real at all (spooky!).
Detection tools are a patch on a structural crack
The industry response to all of this has been, predictably, a collection of detection measures. Asking candidates to turn sideways to check for artefacts in the video feed. Deepfake detection software. Platform-level restrictions on screen-sharing or third-party audio. Proctoring tools that flag unusual behaviour.
These measures are okay but it's better off to use the selection methods that are much less vulnerable to AI (like reserving your interviews for in-person only).
The honest conclusion is that video interviews have moved from being a moderate selection tool to a fairly unreliable one in the age of AI. Using them as a meaningful evaluation stage, one that carries real weight in your hiring decision, is a much harder case to make now than it was pre-2022.
The evaluative weight has to go somewhere else:
- AI-resistant cognitive assessments
- In-person work samples and task-based assessments (during interviews or as part of an assessment day)
These methods don't carry the same fraud surface that a live video call now does. Read our breakdown of which selection methods are affected by AI.
Conclusion and next steps
Video interviews haven't been destroyed by AI. But they've been compromised enough that treating them as a primary evaluation tool is now a decision worth reconsidering because the cheating is so effective and hard to fully catch.
The structural issue is that interviews have always measured a mix of actual ability and performance skill. What AI has done is make it possible to inflate the performance skill layer dramatically, at every tier from preparation to identity fraud, while the effort required to do so has dropped close to zero. The thing that made coaching a limited problem before was that it took time and effort. That's no longer the case.
If you want to keep something predictive in your early process, our MindmetriQ assessments are specifically designed to be resistant to AI-assisted inflation, testing the actual reasoning and problem-solving ability that no tool can read out for you. For more info read our breakdown of which selection methods are vulnerable to AI and why.
Ben Schwencke
Chief psychologist at Test Partnership. MSc in Organisational Psychology with over ten years experience in psychometric testing.
