AI & Integrity6 min read

What AI integrity scores actually catch (and what they don't)

Every AI interview platform worth using now ships some version of an integrity score — a number that claims to tell you how "authentic" a candidate's responses were. But what does that number actually measure? When should you trust it? When should you ignore it?

This is an honest breakdown, including the things we'd rather not admit.

What the score is actually measuring

A suspicion score isn't a lie detector. There's no microphone analysis, no voice stress detection, no webcam feed being processed (in Round1's case, there's no video at all by design). What the score actually measures is a combination of:

Response latency patterns
Unusually long pauses before answering (suggesting external lookups), or suspiciously instant responses to complex technical questions.
Linguistic consistency
Sudden shifts in vocabulary, sentence complexity, or technical precision mid-interview. If someone explains basics in simple terms but then delivers a perfectly structured answer on a hard question, that's a flag.
Answer structure analysis
GPT-generated answers have a recognisable structure: they're comprehensive, well-organised, and almost always complete. Human answers tend to trail off, miss sub-points, and have natural backtracking.
Follow-up coherence
When the AI asks a follow-up based on the candidate's previous answer, can they build on it? Candidates reading from a script or using AI in real-time often struggle to connect their answers.

What it reliably catches

✓ Clear AI-assisted responses

When a candidate is pasting questions into ChatGPT in real-time, the pattern is usually obvious. Structured lists where none were asked for. Perfect definitions of niche concepts they got wrong moments earlier. Latency spikes on harder questions. The score catches this well.

✓ Pre-scripted answers read aloud

Candidates who memorise answers to common questions verbatim often show a specific pattern: near-instant responses, perfect recall, but poor follow-up coherence when asked to expand.

✓ Third-party assistance (someone else answering)

When another person is answering for the candidate, you usually see a consistent disconnect between answer quality and follow-up coherence. The answers are too good for someone who can't build on them.

What it doesn't catch (and sometimes gets wrong)

✗ Nervous candidates who pause a lot

Anxiety looks similar to external lookup in some scoring models. A candidate who thinks carefully before answering can score higher on suspicion than they deserve. Always read the transcript, not just the number.

✗ Non-native speakers who write more formally

Some candidates communicate more formally in spoken English than native speakers, because that's how they learned the language. This can superficially resemble AI-generated text. The score can be unfairly elevated.

✗ Well-prepared candidates with great recall

There's no universal rule that great answers = cheating. A candidate who has genuinely prepared, knows their domain deeply, and communicates clearly can produce answers that score as "suspiciously good." That's not a problem — that's what you want in a hire.

✗ Sophisticated AI use

A candidate who uses AI to prepare and then truly internalises the content, answering in their own voice, will not be flagged. This is actually the right outcome — if someone prepared that well and can answer naturally, they've done the work.

How to use the score correctly

The score is not a pass/fail gate. It's a flag that prompts a closer look. The right workflow:

0–40
Low suspicion
Take the assessment at face value. Candidate likely answered authentically.
41–65
Moderate suspicion
Look at the transcript. Were their follow-up answers as coherent as their initial responses? Flag for a deeper round 2 probe if advancing.
66–100
High suspicion
Read the full transcript carefully. Don't disqualify automatically, but do verify: probe specific answers in round 2. If they can't explain what they said, that's your answer.

The honest bottom line

The integrity score is a useful signal for catching obvious cheating. It is not a reliable proxy for candidate quality. Some of the most impressive candidates in a batch will score 60+ and deserve to advance. Some candidates who score 20 are coaching themselves through bland answers.

Use it the same way you'd use a plagiarism checker in a university setting: it narrows your investigation, it doesn't replace judgment.

The takeaway

An integrity score is a shortlist tool, not a verdict. Always read the transcript. Always probe flagged candidates in round 2. Never disqualify on the number alone.

See Round1's integrity scoring in a live interview.

Try a Demo Interview →← More articles