Signs a Candidate Is Reading a Script on Zoom

Before "AI interview cheating" was the phrase everyone searched, recruiters were already dealing with a simpler version of the same problem: candidates reading answers off a script, a second monitor, or someone else in the room feeding them lines. AI just made the script better and easier to generate on the fly. The visual and behavioral tells are largely the same either way — which is useful, because it means you don't need to become an AI expert to notice most of them.

What to actually look for on the call

• Eyes moving sideways or down, repeatedly, at the start of an answer. A brief upward glance while thinking is normal — most people do it. Consistent sideways or downward eye movement right before speaking, especially if it happens on every question rather than just the hard ones, is a real pattern worth noting rather than a one-off.
• A delay before speaking that doesn't match the question's difficulty. A two-second pause before "tell me about yourself" — a question everyone expects and has usually rehearsed — is unusual. If it happens consistently on questions that shouldn't require thinking time, something is being read or generated in that gap.
• Answers that are suspiciously well-structured for how they're being delivered. A real off-the-cuff answer usually has some looseness — a restart, a "wait, let me put that better," a sentence that trails off and gets picked back up. An answer delivered in clean, complete paragraphs with no self-correction, every single time, is a pattern that stands out once you notice it.
• A flat, reading cadence rather than a speaking cadence. People speaking their own thoughts vary their pace and emphasis — they speed up on things they're excited about, slow down when they're choosing a word carefully. Reading, whether from a physical script or a screen showing generated text, tends to produce a more even, monotone delivery that doesn't track the content's actual emotional weight.
• Typing sounds, or a slight lag, right before a detailed answer. If your setup allows you to notice it, the gap between a question landing and a long, specific answer starting is sometimes filled with the faint sound of someone typing a prompt, or a half-second of dead air that doesn't match natural speech timing.
• A camera angle that never quite looks at you. Reading from a second monitor placed beside the laptop camera produces a subtly consistent off-axis gaze — not looking away occasionally, but never quite landing on the camera even during supposedly personal or emotional answers.

What this looked like before AI, and why the signals are the same

Long before stealth AI copilots existed, some candidates prepared cheat sheets — sticky notes around the monitor, a printed page of talking points, a friend on a second phone line feeding answers. The tells were identical to the ones above: sideways glances, unnatural pauses, oddly polished delivery. AI changed what's being read (a generated, contextually relevant answer instead of a static, generic one) but it didn't change how reading looks on camera. That's actually good news for recruiters — you don't need new detection skills, you need to trust the ones that already applied.

A realistic example

A candidate gives a strong, structured answer to "tell me about a time you disagreed with a manager" — clear situation, action, result, delivered smoothly with almost no filler words. On its own that's just a well-rehearsed answer, which isn't suspicious by itself. What's worth noticing is what happens on the natural follow-up: "what did your manager say when you brought it up?" If the candidate answers that immediately and specifically, they were probably speaking from real memory. If there's a longer pause, a vaguer answer, or a subtle restart of the same general point without adding a new detail, the first answer may have been read rather than recalled.

How to word the follow-up that actually tests it

The follow-up is what turns a hunch into something you can act on, and it works best when it asks for a detail that couldn't have been anticipated in the original script:

• "What was the exact number/date/name involved in that?"
• "What would you do differently if it happened again tomorrow?"
• "Who else was in the room, and what was their reaction?"
• "Walk me through the ten minutes right before that conversation."

Someone who was speaking from real experience adjusts to these easily, often adding a detail that makes the story more vivid. Someone reading a prepared or generated answer often can't extend past what was already written for them, and the gap between the fluent first answer and the halting follow-up is the tell.

The honest limitation of watching for this yourself

All of the signals above are real, but they're also things a tired interviewer misses on the fifth call of the day, and none of them are proof on their own — some candidates are just naturally more measured speakers, and some have genuinely rehearsed a true story until it sounds smooth. That's exactly why a visual read should be a prompt to ask a sharper follow-up question, not a basis for a decision by itself. Treat it as a reason to dig one level deeper, not a verdict.

Where the transcript does this work for you

Candit reads the interview transcript in the background the whole call and checks for exactly the cadence pattern behind a read or generated answer — missing the disfluencies a live thinking process normally produces, a sudden shift into unnaturally clean phrasing — cross-checked against how real stealth AI copilots actually render answers. When it crosses the threshold, it's flagged live, per speaker, with the specific line that triggered it, so the visual read you might have caught (or missed) has evidence backing it up either way. A flag also suggests a follow-up to ask on the spot (Counter AI), and after the call the AI use report compares how each answer was timed, alongside the flagged lines.

It's also worth being clear about why this still works against a tool that hasn't been named yet: Candit doesn't try to lock down a candidate's device or browser the way older proctoring tools do — there's no environment to escape, because the signal isn't what's on a screen, it's how the answer was actually delivered, read live from the transcript and audio as the interview happens. That's ongoing work, not a one-time build: reverse-engineering how candidates actually use these tools is retested on a roughly weekly basis as new copilots appear.

First session free. No card required.