Screening integrity
Candidates using AI in interviews
Every vendor on this question has a detection feature to sell. Here is the version built on human review of the evidence.
Can you detect whether a candidate used AI in a video interview?
Not reliably, by any vendor. Truffle is built on human review instead: a one-way video interview gives you the recording and the transcript, so a human can judge the answer, plus questions that ask for specifics only that candidate would know, which a generated answer fails on its own merits.
- How answers are judged
- By a person, with the recording and transcript
- How answers are scored
- Against your criteria, with the evidence shown
- Identity verification
- Works alongside your existing provider
- What you get
- The recording, the transcript, the evidence
A confident detection score would be a false one
Telling whether a spoken answer was drafted with help is not a solved problem. The tells people reach for — a pause before answering, polished phrasing, eyes tracking to the side — are also what a nervous candidate reading their own notes looks like, and what a non-native speaker composing carefully looks like. A product that turned those into a percentage would be producing a number with no defensible basis, and in high-volume hiring that number would end up rejecting people.
So Truffle puts the recording, the transcript, and the evidence in front of a person instead. The useful move is not catching the tool, it is making the tool unhelpful.
Design the questions so a generated answer scores badly
Ask for their specifics
Not "how do you handle an angry customer" but "what was your AHT target on your last account, and what happened the month you missed it". A model can produce the first convincingly. The second needs a history the candidate either has or does not.
Score against criteria, not polish
AI Match scores each answer against the criteria you set for the seat and shows the line of evidence that earned the rating. A fluent answer that never lands on a criterion scores low, which is the behaviour you want regardless of how it was produced.
Verify later, not here
Treat the first round as a filter. The things that would be expensive to get wrong — identity, credentials, right to work — belong at a stage where you are on a live call or in person, with a verification provider you already use.
It raises the effort. Be precise about that claim.
A spoken answer recorded on camera inside a response window is a harder place to consult something mid-thought than a take-home form is. Follow-up questions probing a detail the candidate just mentioned are harder to handle with a prepared answer than a standalone competency question. Both of those are real, and both are differences of degree, not of kind.
What they are not is a guarantee. Anyone telling you their format makes AI assistance impossible is describing a marketing position rather than a threat model. The reasonable expectation is that a well-designed asynchronous first round costs a candidate more effort to game than a written screen, and that the answers which survive it are more informative — not that nothing gets through.
- Every answer keeps its recording and its transcript, so a human can check any score
- Candidate Shorts surface the thirty seconds a reviewer should watch, tagged by competency
- Question-level guidance states what the question is for and what a strong answer looks like
- The AI never auto-rejects anyone. A low score is a prompt to look, not a decision
“She'd been cut off twice already, so I stopped reading the script and just told her what I could actually do that day.”
Play 2:14 – 2:41The position
We would rather lose the query than sell you a detection score we cannot stand behind.
Candidate sentiment data on the format is in The AI Interview Myth report →
AI in interviews: common questions
Can you detect if a candidate used AI in a video interview?
Reliable detection of AI assistance is not a solved problem for any vendor, and a confident score would be a false one. Truffle gives you what holds up instead, the recording itself: a human can watch a candidate answer, see the pauses and the eye movement, and read the transcript alongside it.
Does a one-way video interview make AI assistance harder than a written screen?
Generally yes, because of timing rather than technology. A spoken answer given on camera within a response window leaves less room to consult something mid-answer than a take-home form does, and follow-up questions that ask for a specific example from the candidate's own experience are harder to generate convincingly than a generic competency answer. It raises the effort; it does not eliminate the possibility.
So how should a BPO handle it?
Treat the first round as a filter, not as proof. Ask for specifics only that candidate would have — the account they worked, the escalation they handled, what their AHT target actually was — and verify the things that matter at a later stage where you are face to face or on a live call. Design the questions so a generic answer scores badly, which is cheaper and more reliable than trying to catch a tool.
Does using AI to screen mean you are fine with candidates using AI to answer?
They are different problems. Truffle uses AI to read every applicant against criteria a human set and to show the evidence behind each score, with a person making every decision. A candidate using AI to generate an answer is substituting for the thing you are trying to measure. Truffle's job is the first: reading every applicant the same way, showing the evidence, and leaving the decision to a person.
Do candidates actually mind being screened this way?
Less than the discourse suggests. Across hundreds of voluntary reviews submitted after a Truffle video interview, 93% rated the experience positively — the full method and the negative reviews are in The AI Interview Myth report. The complaints that do come in are about clarity and length, not about the format itself.
More on hiring at BPO scale
Ask better questions, score against real criteria
One-way interviews with per-question evaluation guidance, evidence behind every score, and the recording kept beside it. No detection theatre.
7 days · 30 credits · no card required