How to Spot "AI-Scripted" Candidates
ChatGPT is writing résumés, drafting cover letters, and whispering answers from a second screen.
Every recruiter I've spoken to in 2026 says the same thing: the funnel looks better than ever, and the hires are worse than ever. Polished résumés. Articulate intro calls. Frameworks dropped on cue. Then the person starts the job and you realize you hired a prompt, not a professional.
This isn't a moral failing of candidates. It's a system failure. We built interviews that reward verbal fluency, structured answers, and recall — exactly the things an LLM is best at. If you don't redesign the interview, you will keep hiring AI.
The classic behavioral interview — "tell me about a time you led a team through change" — is the perfect prompt for an LLM. It's open-ended, structured (STAR), and rewards a polished story. A weak candidate with a second monitor will beat a strong candidate who's nervous and authentic, every single time.
- Behavioral questions a candidate can prepare in advance.
- Long monologue answers (3+ minutes).
- Generic case studies anyone can google.
- Treating the interview as a Q&A exam.
- Letting the candidate control screen-share and camera angle.
- Live, collaborative problem-solving on a shared doc.
- Sudden scenario shifts mid-answer ("now the budget is cut in half").
- Real-time critique of a sample artifact from your company.
- Working sessions, not interrogations.
- Requiring full-frame camera and asking to see the room once.
1. Anchor the conversation in their actual artifacts
Open with: "Walk me through this specific bullet on your résumé — what was the input, who pushed back, what did you change your mind about?" AI-generated bullets collapse under three layers of follow-up. Real experience gets richer.
2. Use scenario shifts mid-answer
Halfway through their response, change a variable. "Now imagine the head of sales disagrees publicly in the same meeting — what do you do in the next 60 seconds?" Scripted candidates restart from the top. Real practitioners pivot.
3. Make them critique, not just produce
Share a real (sanitized) artifact from your company — a job description, a policy draft, a one-pager — and ask them to critique it live. LLMs are trained to please. Strong candidates push back with specifics.
4. Collaborate, don't quiz
Spend 30 minutes co-writing something in a shared doc with them. You'll learn more in that half hour than in three rounds of behavioral questions. Watch how they handle ambiguity, how they ask clarifying questions, and what they do when stuck.
Benjamin Bloom's 1956 taxonomy ranks cognitive work from recall (lowest) to creation (highest). LLMs are extraordinarily good at the bottom two layers — recall and comprehension — and meaningfully worse at evaluation and creation. Almost every interview question that LLMs can ace is a recall question dressed up as a 'reasoning' question. The fix isn't to detect cheating; it's to redesign the interview so the question sits in the top half of Bloom's pyramid, where AI's help is marginal.
Layer on Karl Popper's falsifiability principle: a great interview question is one where a candidate's answer can be wrong in a specific, identifiable way. LLM answers are usually plausible-but-generic, which is exactly the failure mode a falsification-friendly question surfaces. 'Tell me about a time you disagreed with your manager' is unfalsifiable. 'Walk me through a specific decision you reversed last quarter and what changed your mind' is not.
A YC-backed dev tools company, as one HR director recounted, in 2024 was seeing inflated technical interview scores and disappointing 90-day performance. They replaced the 'design a URL shortener' question (perfect for ChatGPT) with a live pair-debug of their actual codebase, with the interviewer evolving the bug mid-session. Hire-quality match-to-expectation scores jumped from 51% to 79% in two quarters. They didn't add an AI-detector. They removed AI's leverage.
- Replace recall questions with live scenarios that evolve as the candidate responds.
- Use your actual codebase, real customer email, real anonymized data — not a generic prompt.
- Require the candidate to think out loud and respond to interviewer perturbations.
- Score 'depth of follow-up' not 'first answer.' AI gets the first answer; humans handle the third follow-up.
- Include one question with no clean answer. AI fabricates; humans say 'I don't know, but here's how I'd find out.'
- Train interviewers to push for specifics: 'which Monday,' 'who was in the room,' 'what was the dollar number.'