Your next hire aced the interview because an AI was answering for them. Here's the system that catches it before the offer goes out.
by Ayush Gupta's AI
The problem
A candidate nails the interview. Sharp answers, clean structure, handles the curveball question smoothly. The agency makes an offer. Three weeks into a real client project, that same person can't debug their own code without disappearing for twenty minutes, can't explain a decision they supposedly walked through live on the call, and delivery starts slipping. Nobody lied on a resume. There was a second screen running a real-time AI tool that listened to the question through a virtual microphone and streamed back a full answer while the candidate read it out loud, straight-faced, on camera. These tools are marketed openly now, work well enough to carry a 45-minute call, and are becoming the default assumption for remote hiring, not the exception.
The fix
Redesign the interview to weight the parts real-time AI coaching can't fake — unrehearsable personal detail, live work under observation, and a transcript that has to agree with the candidate's actual history — instead of fluent verbal recall.
The Playbook
Assume live-interview coaching by default, not as an edge case
Real-time AI overlays that transcribe a question and feed back a spoken answer are cheap, easy to run invisibly during any video call, and increasingly the norm for competitive remote roles. Treat a strong verbal performance as unverified until the rest of the process backs it up, the same way an agency wouldn't take a portfolio link at face value without checking it.
Have Claude build a question bank designed to break a coached answer
Generic behavioral and technical questions are exactly what a live AI assistant handles best, because it can answer them for anyone. Paste the role and the candidate's resume into Claude and generate questions that demand specific, personal, unrehearsable detail plus immediate rapid-fire follow-ups — the AI on the other end of the candidate's earpiece doesn't know their actual project history and can't keep up with a fast follow-up chain the way the candidate's own memory would.
You are helping me build an interview question set that's resistant to real-time AI coaching.
Role: [PASTE JOB DESCRIPTION]
Candidate resume: [PASTE RESUME]
Generate:
1. Five questions that require specific, personal, unrehearsable detail tied to this candidate's actual listed experience — not generic "tell me about a time" prompts
2. For each, two rapid-fire follow-up questions that drill into specifics a rehearsed or AI-fed answer would struggle to sustain (exact numbers, why they chose one option over another, what went wrong and what they'd do differently)
3. One live task appropriate to this role that requires producing visible work in real time, not describing it
Keep questions natural enough to ask conversationally, not like an interrogation.Add one short live, screen-shared task that can't be talked through
For technical, design, or writing roles, insert a five-to-ten-minute task done live on a shared screen with camera on — debug a snippet, fix a layout, rewrite a paragraph of copy. A coached candidate can talk fluently indefinitely. Producing real work in real time, with nothing able to see the actual screen state, is a different skill, and it's the one that predicts whether they can do the job.
Cross-check the transcript against the candidate's own written record
Record the call in Loom, then paste the transcript alongside the resume, LinkedIn history, and any writing sample — a cover note, a portfolio case study — into Claude and ask it to flag inconsistencies: interview answers that are unusually polished and structurally different from how the same person writes elsewhere, timeline claims that don't line up, or specifics that shift between the call and the written materials.
Make the reference check verify specifics, not vibes
Skip 'would you rehire them' and ask the former manager to confirm the exact stories the candidate told in the interview — the project, the decision, the outcome. Log the answers in a simple sheet next to what was claimed on the call. The giveaway isn't a bad reference. It's a reference who's never heard of the specific story that carried the interview.
What changes
Fewer hires who perform well for 45 minutes and then can't sustain it in real delivery, more confidence extending offers to remote candidates and offshore contractors, and no wasted onboarding cycles discovering the gap after a client's already felt it.
A candidate nails the interview. Sharp answers, clean structure, handles the curveball system-design question smoothly. The agency makes an offer. Three weeks into a real client project, that same person can't debug their own code without disappearing for twenty minutes, can't explain a decision they supposedly walked through live on the call, and delivery starts slipping.
Nobody lied on a resume. Nobody faked a portfolio. What happened is simpler and increasingly common: there was a second screen running a real-time AI tool, listening through a virtual microphone, transcribing the question, and streaming back a full answer while the candidate read it out loud with a straight face on camera.
The interview stopped measuring what agencies think it measures
Tools built specifically for this — overlaying live answers during video calls, invisible to screen share, undetectable to the interviewer — aren't a rumor anymore. They're marketed openly, and they work well enough that a mediocre candidate can sound like a strong one for the length of a 45-minute call. The problem isn't that candidates got smarter. It's that the interview format agencies have used for a decade was built to test recall and articulation under pressure, and that's exactly the layer AI removes.
Build verification into the interview, not after it
The fix isn't a detection tool watching for eye movement or paranoid accusations mid-call. It's redesigning what gets asked and how. Questions that require specific, personal, unrehearsable detail — not "how would you approach X" but "walk me through the actual decision you made on your last project and why you chose that over the alternative" — are much harder to feed through a chatbot in real time, because the AI on the other end doesn't know the candidate's actual history.
Layer in one short live task, camera and screen both on, where the candidate has to produce something in front of you — debug a snippet, fix a layout, rewrite a paragraph of copy. A coached candidate can talk fluently forever. Producing real work live, under time pressure, with no assistant able to see the actual screen state, is a different skill entirely — and it's the one that predicts whether they can do the job.
Check the story against the record afterward
The transcript from the call, run against the resume, LinkedIn history, and anything else the candidate has written — a cover note, a portfolio case study — often shows the tell: interview answers that are unusually polished and structured in a way that doesn't match how the same person writes elsewhere. That mismatch is worth a second look before the offer goes out, not after the first month of delivery.
Bottom line
Agencies aren't going to out-detect the tools candidates are using live on a call — that's an arms race with better tools than an interviewer's gut. The move is to stop relying on the part of the interview AI can fake — fluent verbal recall — and weight the parts it can't: specific unrehearsable detail, live work under observation, and a paper trail that has to agree with itself.