CodeSignal published a report in February 2026 that should make every hiring manager uncomfortable: AI cheating on technical assessments went from 16% of candidates to 35% in a single year. More than one in three people you’re interviewing may not be answering your questions themselves.
That’s not a niche problem anymore. That’s your pipeline.
The frustrating part is that most companies haven’t changed how they interview. They’re running the same video calls and coding tests from 2021 while candidates have ChatGPT, Claude, and real-time coding assistants running on a second monitor. Or a second phone. Or just earbuds connected to an AI reading the screen.
Why This Got So Bad So Fast
A year ago, using AI during a live interview felt risky. The tools were slow, obvious to use, and easy to spot if you knew what to look for. That’s changed. AI answers now come back in under two seconds. Voice-to-text piped into a prompt and back into a hidden earpiece is a real thing people are doing. Screen-sharing doesn’t catch it because the AI interface is on a separate device.
The candidates doing this aren’t necessarily bad people. A lot of them are desperate job seekers who know the system is broken — that they’ll get filtered out by keyword scanners before a human even reads their resume — and they’ve decided to fight fire with fire. That’s a separate conversation. Your job right now is to make sure the person you hire can actually do the work.
What AI-Assisted Answers Actually Look Like
If you’ve interviewed enough people, you develop an instinct for when something’s off. Here are the specific patterns that show up when someone is feeding questions to an AI and reading back the response.
The pause-burst pattern. The candidate goes quiet for 4-8 seconds after you ask a question, then delivers a surprisingly complete and structured answer all at once. Normal human thinking is messier — people start talking before they’ve fully formed their answer, backtrack, add things. AI output comes pre-organized.
Perfect structure, no texture. Every answer has an intro, numbered points, and a clean conclusion. Real interview answers are looser. People say “um” and “like” and “actually, wait.” They trail off. When every response sounds like a polished LinkedIn post, something’s generating it.
Can’t defend the answer. Ask them to go deeper on something specific they just said. “You mentioned X — can you walk me through how that works under the hood?” A candidate who answered genuinely can elaborate. A candidate who read an AI response often can’t go any further than what they already said.
Eyes that aren’t on the camera. This is subtle but consistent. When someone is reading from a second source, their gaze shifts slightly and regularly. Not dramatically, just a micro-pattern of looking slightly left or down every 15-20 seconds.
Vocabulary mismatch with the resume. The resume is casual and straightforward. The interview answers are dense and technical. Or the reverse. Pay attention when the voice doesn’t match the writing.
What to Actually Do About It
Detection is useful, but process changes are better.
Ask opinion questions, not knowledge questions. “What’s the difference between X and Y” is easy for AI to answer. “Tell me about a time you had to choose between X and Y in a real project, and why you went the way you did” is much harder to fake. Personal experience questions require specifics that AI can’t manufacture.
Go off-script mid-answer. If a candidate is reading AI output, interrupting them mid-answer with a contextual follow-up question throws them off. “Wait, back up — when you said [phrase], what specifically did you mean?” Genuine answers handle detours. Scripted ones don’t.
Do a live debug session instead of a clean problem. Give candidates broken code and ask them to talk through what’s wrong while they fix it. Narrating while debugging is genuinely hard to outsource in real-time.
Add a short no-aid screen after the initial interview. A 20-minute video call where they share their screen and work through a problem in real time — no AI help, no second monitor — is enough to verify they can actually do what they claimed.
Use AI detection tools where appropriate. Tools like Fabric and others are built to catch behavioral signals of AI assistance — response latency analysis, keystroke patterns, behavioral flags. They won’t catch everything, but they raise the cost of cheating substantially.
A Note on the Hard Cases
Some candidates use AI not to cheat but to compensate for real disadvantages — English as a second language, interview anxiety that tanks their verbal performance even when they’re technically strong, or other factors that affect how they communicate in high-pressure settings.
The candidates you want to screen out are the ones who genuinely can’t do the work at all. If someone’s interview performance and their on-the-job performance match, you don’t have a problem worth worrying about. Keep that in mind when building detection into your process.
The Bottom Line
35% is high enough that you should assume some portion of every hiring cohort used AI assistance during the interview. Adjusting your process to verify genuine competence isn’t paranoid — it’s practical. The fixes aren’t complicated: better questions, more follow-up, one real-time hands-on task. That’s enough to catch the gap between what someone claims in an interview and what they can actually do.