When a Candidate Reads Answers in an AI Interview

A candidate reading answers in an AI interview once walked away with the only Strong Go in their batch. The answers were structured, specific and fluent. On paper, nobody else came close.

A human reviewer caught it later by listening. The candidate had been reading prepared answers off a screen, and to the reviewer it sounded like someone reading an essay aloud. Our screen had seen the words. It had not listened for how they were delivered.

That miss is where our “possibly scripted” flag came from. It went live on production on July 22, 2026. Below: what it listens for, why it errs on the side of silence, and the check to run once it fires. It is the cleanest example of the gap between what an interview transcript records and what the voice reveals.

Why a Transcript Cannot Catch Someone Reading

A transcript of a read answer is indistinguishable from a transcript of a brilliant one. Both are grammatical, organised and on-topic. Anything you can check in text, keywords, structure, relevance, the reader passes.

A written check makes it worse. If your screen rewards structured, complete answers, a reader with good notes is the ideal candidate on paper. The screen ends up rewarding preparation of the wrong kind.

What a transcript throws away is the texture of live thinking. People who compose an answer as they speak restart sentences, repeat a word while they find the next one, and drop small fillers into the gaps. Those fillers are not sloppiness. Clark and Fox Tree argued that uh and um work as signals of a coming delay, the audible sound of someone still building the sentence. Someone reading does almost none of that.

This is not our theory. Mark Liberman at the Linguistic Data Consortium compared the same speaker giving read remarks and spontaneous remarks to the same audience. The overall speaking rate was similar. The spontaneous speech was full of filled pauses and quick word repetitions. The read speech had essentially none.

The Three Signals Behind the Possibly Scripted Flag

The flag checks three things from the candidate’s own audio, and it fires only when all three are true at once:

1. Near-zero hesitation. At most 1.0 hesitation per 100 words, counting fillers, repeated words and hedges. Natural spontaneous speech usually runs somewhere around 3 to 8 per 100.
2. Fluent pace. 140 words a minute or faster. Readers tend to be quick and even. Nervous candidates are rarely both quick and clean.
3. No self-corrections at all. Zero repeated words across the whole call. Live speakers repeat themselves while they think.

On top of those, there is a sample-size rule. The candidate must have spoken at least 150 words. Under that, the flag stays silent, because a short call cannot tell a reader from someone who simply answered briefly.

When the sample is too small, the flag does not stay quietly blank either. It records that the sample was too small and how many words there were, so nobody mistakes “not checked” for “checked and clean.” It sounds pedantic until someone upstairs wants to know why a reader slipped through, and the true answer is that the call held only 90 words.

The AND matters more than any single threshold. Each signal alone is common in honest candidates. Across 1,447 finished production interviews we checked, 8.7% of candidates had zero self-repetitions and the least hesitant tenth sat at about 1.5 hesitations per 100 words. Flag on any one of those and you would be accusing a lot of fluent, honest people. Requiring all three together is what keeps the flag rare.

Three Candidates Run Through the Flag

The rules are easier to trust once you see them sort real-looking cases. These three are illustrations, with numbers chosen to show where the lines sit.

Candidate A clears every condition. Almost no hesitation, quick even delivery, not one repeated word across more than 400 words. That is the pattern of text being read, and the chip appears.

Candidate B looks almost identical, but repeated themselves twice. Two small self-corrections are enough to show live composition somewhere in the call, so no flag. This is deliberate. A very polished speaker who stumbles even once or twice is far more likely to be fluent than to be reading.

Candidate C is the uncomfortable one. Near-zero hesitation, no repeats, but slow. They might be a careful reader. They might be a calm, deliberate speaker. The pace rule cannot tell those apart, so it stays silent rather than guess.

If Candidate C bothers you, that is the right instinct, and the recording is where to settle it. The flag exists to point you at the obvious cases, not to rule on the ambiguous ones.

It Also Checked Calls From Before It Existed

When the flag shipped on July 22, it did not start from a blank slate. The check is a pure calculation over speech measurements every call already stores, so it runs in two places: when a new call is analysed, and whenever an older call is opened in the review queue or on the candidate’s page.

That meant months of past interviews gained the chip on the day it went live, with no reprocessing job and no change to any stored score. For a team with a long queue of past Strong Go candidates, that is useful. You can scroll back through the shortlist you already sent to hiring managers and see whether any of them carry the chip.

Why the Flag Never Touches the Score

This was the argument inside the team, and I think we landed on the right side of it.

A reader who gets through the screen costs you one wasted interview round. An honest, articulate candidate who gets marked down because they happen to speak smoothly costs you a hire, and it is a quiet unfairness they will never know happened. The harm is not symmetric, so the design is not either.

So the flag informs and nothing more. Communication, the overall verdict and your auto-reject band all ignore it. In the /inbox review queue it shows as a small “possibly scripted” chip in the AI column. On the candidate’s Speech Analysis panel the same chip adds “verify recording,” and hovering shows which conditions held.

That wording is deliberate. The flag is a reason to spend ninety seconds listening. It is not a finding, and nobody on your team should treat it as one.

The same principle runs through how we handle accent risk in voice scoring: a delivery signal should earn a closer look, never a silent penalty.

How to Check a Flagged Candidate in 90 Seconds

When the chip appears, play the call back from the candidate’s record in the HR dashboard and do three things.

First, listen to one prepared-sounding answer. Does the rhythm stay perfectly even from start to finish? Readers keep a steady cadence through long, complex sentences that live speakers would break up. You will often hear a slight sing-song quality, the same rise and fall on every clause.

Second, listen to a follow-up. The interviewer builds follow-ups on the candidate’s previous answer, and nobody can script those in advance. A reader usually changes sharply here. The pace drops, the fillers come back, or the answer stops matching the confidence of the one before it.

Third, compare the two. If the follow-ups sound like the same person thinking at the same level, clear the flag in your head and move on. You have probably found a fluent speaker who prepared well, which is a good sign, not a bad one.

A short checklist for the review:

That third row is a different pattern with a different cause, and we covered it in chatbot help during a live screen. If the flag shows up on a call that also carries audio warnings, read what happens when the line is bad first, because noise can hide the small hesitations the flag depends on.

Where Preparation Ends and Reading Begins

It is worth being honest about the grey area, because candidates are told to prepare and they should.

A candidate who has practised telling the story of their last project three times will sound fluent telling it a fourth. That is preparation. We want it. The flag’s conditions are strict enough that most well-prepared candidates still hesitate and self-correct at a normal rate, because they are recalling, not reading.

Some roles blur the line more. For voice process and customer support hiring, fluent delivery of a standard answer is close to the job itself, and for BPO hiring the voice screen is the job test. If your role involves reading scripts to customers, a smooth reader is not a red flag, and your team should know to read the chip in that light.

And some candidates read because they are anxious, not dishonest. A fresher who wrote out answers to calm their nerves is not the same as someone feeding off a second screen. We cover what to tell those candidates beforehand in preparing candidates for an AI interview. Telling people up front that the call is a conversation, with follow-up questions, removes most of the incentive to script.

What the Flag Will Miss

A flag this cautious misses things, and you should know which ones.

A slow reader will not trip it. Someone reading at 120 words a minute fails the pace condition, so the flag stays silent. A reader who adds fake hesitations would get past it too, though in practice that is harder than it sounds over a fifteen to twenty minute call. And any candidate with under 150 words of speech is never judged at all.

We accepted those misses on purpose. Tightening the thresholds to catch slow readers would start catching slow, careful honest speakers, which is exactly the harm the design is built to avoid. Our earlier write-up on detecting rehearsed answers described the idea. This flag is the narrower, measured version that actually shipped.

What This Changes for Your Screening Queue

The practical change is small and worth making: give flagged candidates a listen before you advance them, not before you reject them. The chip sits next to the communication score, and reading that score’s two halves tells you whether delivery or content carried the candidate to the top. Since a reader’s transcript is usually strong, the flag matters most at the top of your queue, where a false Strong Go is expensive.

It also changes what you should ask any AI screening vendor. Not “can you detect cheating,” which every vendor will say yes to. Ask what the signal is based on, how often it fires on honest candidates, and whether it can change a score without a human listening. If the answers are “our model,” “we don’t know” and “yes,” you are buying an accusation engine.

For a feel of the delivery read on an actual call, try the HireQwik interview as a candidate. It is short, free and signup-free, and the scorecard in your inbox carries the same speech line.