A recruiter at a mid-size SaaS company told me something last year that stuck with me. She said she could tell when a candidate was reading off a screen not because of eye movement, but because their sentences had no filler. No “um”, no brief restart after a hard word, no slight pause before a technical term. They were too clean.
That’s the real problem with using live AI assistance during interviews. Not the tool. The behavior the tool creates.
What interviewers actually notice
Most AI-detection advice online focuses on eye tracking and suspicious pauses. Those matter, but they’re not the main tell. The bigger giveaway is unnatural fluency. A real person answering a hard behavioral question will hedge. They’ll say “I think” or “if I recall correctly” or briefly circle back to correct themselves. AI-assisted answers tend to be smooth from word one, which sounds less like a person thinking and more like a person reading.
A second tell: overly precise framing. If someone asks “tell me about a conflict you navigated” and the answer comes back with a crisp three-act structure inside 90 seconds, that raises flags. Real stories meander slightly. They have texture.
Screen-share is a separate concern. If you’re doing a technical round where the interviewer can see your desktop, any overlay UI will be visible unless the tool is explicitly built to stay off shared displays. That’s not a behavior problem, it’s a setup problem.
The setup that actually works
Camera at the top of your monitor, not on a separate screen to the side. This matters because if your AI overlay sits directly below your camera, glancing at it reads as looking slightly down while thinking, which is normal. Glancing 30 degrees to the left reads as looking at a second screen, which isn’t.
Single monitor is better than dual for this reason. The moment you have a second display, you’re fighting the geometry.
Craqly, for instance, runs as a floating overlay that sits inside your primary display and doesn’t show up in screen-share captures. I can’t speak to every interview platform’s screen-share behavior (some use OS-level capture, some don’t), but testing your setup before a real interview takes about four minutes and is worth doing every time.
How to read suggestions without sounding like you’re reading
This is the part most advice skips. The goal is to glance at a bullet point, internalize it, and then speak from memory. Not read aloud, internalize. There’s a difference.
Think of it like a teleprompter versus cue cards. Teleprompter gives you the exact sentence. Cue cards give you the topic. You want cue cards. Set whatever AI tool you’re using to produce short fragments, not full paragraphs. “discuss tradeoffs of indexing strategy” is useful. “Indexing can improve read performance but increases write overhead and storage costs” is a script, and scripts sound like scripts.
Taking a two-second pause before answering is fine. Interviewers expect people to think. What they don’t expect is someone who starts talking exactly 0.3 seconds after every question, with zero variation in response latency.
Proctored rounds are a different situation
Some companies, particularly in finance and large enterprise, use proctoring software that monitors keystrokes, active windows, and in some cases camera angles via CV. This is worth researching before the interview. Check Glassdoor reviews for the specific company. If they use something like HackerRank Proctored or Codility’s proctoring mode, the rules of the game are different and the risk of a visible overlay is real.
For standard video interviews on Zoom, Google Meet, or Teams, none of those platforms expose a floating overlay to the other participant unless you explicitly share your screen. That’s a technical fact, not a guarantee, so verify it with a test call.
The honest risk calculus
I think most people using AI assistance in interviews aren’t doing it to fake competence they don’t have. They’re doing it because behavioral questions are genuinely hard to answer under pressure, and technical specifics can slip your mind when you’re nervous. Using a tool to jog your memory is a different thing from using it to answer questions you couldn’t answer without it.
Whether that distinction matters ethically is a question I’ll leave open. But practically speaking: if you use AI suggestions to remember the name of a framework you’ve used 47 times but blanked on, that’s a lot lower risk than using it to answer a question about a technology you’ve never touched. The former requires natural delivery because you actually know the material. The latter will almost always sound like reading.
The recruiter I mentioned at the top has interviewed over 200 candidates in the past three years, according to her LinkedIn activity. She said she’s never caught anyone definitively using an AI tool. She has caught people bluffing knowledge they didn’t have, tool or no tool.
That’s probably the right frame for this whole topic.
- Camera placement beats software tricks for staying natural on video
- Use suggestions as prompts, not scripts
- Research whether the specific role uses proctoring before assuming standard video rules apply
- Fluency without filler is the main behavioral tell, not eye movement
Sources: Glassdoor interview reviews remain one of the better resources for understanding what a specific company’s technical rounds look like. The Stack Overflow Developer Survey 2024 found that 62% of developers were using AI tools in some capacity at work, which gives you a sense of how normalized this tooling has become.
What’s actually being tested in most interviews isn’t whether you memorized everything cold. It’s whether you can think, communicate under pressure, and course-correct when you’re wrong. No overlay changes that.