Interview integrity is the degree to which a hiring interview measures the candidate who will actually show up to work — their own knowledge, their own reasoning, their own communication — rather than a script generated in real time by someone or something else. In 2026, that integrity can no longer be assumed. AI copilots that whisper answers through invisible overlays have made the unproctored video interview one of the least reliable signals in the hiring funnel, and the data now shows it clearly.
The numbers: cheating is no longer an edge case
An analysis by interview-intelligence firm Fabric of 19,368 live interviews conducted between July 2025 and January 2026 flagged 38.5% of candidates for AI-assisted cheating. The rate wasn't static — it roughly tripled over the first few months of the study, from 9% in July 2025 to 45% by September. Software engineering interviews were hit hardest at 48%, against 12% in sales.
Two findings from that dataset should worry hiring teams more than the headline rate. First, 61% of flagged cheaters scored above the passing threshold and advanced in the process — the cheating worked. Second, detection by humans alone is close to a coin flip: a 2025 meta-analysis across 56 studies put average human deepfake-detection accuracy at just 55.5%.
Identity fraud is compounding the problem. In a Greenhouse survey of more than 4,000 respondents, 31% of hiring managers said they had interviewed someone they suspected was using deepfake technology. Gartner found 6% of job seekers admitting to some form of interview identity fraud, and projects that by 2028 as many as one in four candidate profiles could be entirely fake.
Why the live interview became the weakest link
The tooling has professionalised. For roughly the price of a streaming subscription, candidates can run real-time answer engines during behavioural interviews, or coding assistants that render suggestions as GPU-level overlays — invisible to the screen-sharing software recruiters rely on. Audio loopback pipelines transcribe the interviewer's question and return a polished answer within seconds.
The tell-tale signs — a three-to-five second "flatline" delay before every fluent answer, eyes tracking text that isn't the camera, answers that are structurally perfect but personally empty — are real, but they demand trained, attentive interviewers. Most organisations don't have them: surveys of HR teams found roughly half had received no training at all on interview fraud, and only about a third of employers had deployed any detection software.
The result is an uncomfortable asymmetry. The candidate's tooling has improved faster than the interviewer's. An unstructured conversational interview, run over video with no identity verification and no environmental monitoring, is now an open-book exam where only one side knows the book exists.
"Just go back to in-person" doesn't scale
The instinctive reaction is retreat: in-person interview requests have jumped dramatically — from about 5% to 30% of roles between 2024 and 2025 — and 72% of recruiting leaders say they now use in-person rounds specifically to counter AI fraud.
For final rounds, that's sensible. As a screening strategy, it collapses at volume. A BPO drive processing 3,000 applicants, or a campus season spanning forty colleges, cannot fly everyone in for a face-to-face conversation. In-person screening also reintroduces the problems structured assessment was invented to solve: scheduling bottlenecks, interviewer inconsistency, and the well-documented bias of unstructured conversation.
The better response is not to abandon remote evaluation, but to rebuild it so that integrity is measured rather than assumed.
Rebuilding the funnel: assess first, interview later
A defensible 2026 screening funnel inverts the traditional order. Instead of using interviews to filter and tests to confirm, it uses proctored, structured assessment as the primary filter — and reserves human interviews, in person where stakes justify it, for a small verified shortlist.
Proctoring that reports a band, not a verdict
Binary "cheated / didn't cheat" flags fail in both directions: they miss sophisticated fraud and they wrongly condemn candidates with poor lighting or a flaky webcam. Modern AI proctoring instead combines an identity baseline (does the face on camera match throughout?), continuous face and gaze detection, and screen and environment monitoring into a graded integrity signal. AssessAll, for example, issues a High/Medium/Low integrity band alongside every score, so a recruiter can fast-track high-band results, re-verify medium-band ones, and investigate low-band ones — proportionate responses instead of blanket accusations.
Questions that AI answers badly
Generic knowledge questions are exactly what answer engines are optimised for. Scenario-based items — "here is a specific, messy situation; what do you do and why?" — force candidates to commit to judgements, trade-offs, and first-person reasoning that canned AI output handles poorly and that AI-assisted grading can evaluate at scale. Situational judgement tests and AI-graded open responses also produce richer evidence for the interview that follows: instead of re-screening, interviewers can probe the candidate's actual answers.
Time-boxing and item randomisation
Tight, per-item timing compresses the window in which a candidate can relay a question to an external tool and read back the response. Randomised item pools mean leaked questions decay in value quickly. Neither is bulletproof alone; combined with proctoring, they raise the cost of cheating above what most opportunistic candidates will pay.
Verify identity once, carry it forward
If the person who passed the assessment is provably the person who shows up to the interview — same identity baseline, same verified record — deepfake substitution between rounds becomes much harder. This is where verified skill credentials earn their keep: a portable record tied to a proctored session gives later rounds something concrete to check against.
What to change this quarter
Move screening decisions off unproctored interviews and onto proctored assessments; treat early-round video interviews as unverified until proven otherwise. Train interviewers on the handful of reliable tells — response latency, gaze behaviour, perfect-but-impersonal answers — because trained humans plus machine signals outperform either alone. And audit your own funnel: if 38.5% is anywhere near your true cheating rate, your current pass lists already contain people your process didn't actually evaluate.
Cost is no longer a defensible excuse. Usage-priced platforms have pushed proctored assessment down to roughly ₹30 (about US$0.50) per assessment on pay-as-you-go models like AssessAll's — cheaper than the first minute of an interviewer's time, and orders of magnitude cheaper than the 30–150% of first-year salary that a fraudulent hire can ultimately cost.
The takeaway
AI hasn't broken hiring — it has broken the assumption that a live conversation is self-verifying. Teams that move integrity checking upstream, into proctored structured assessment with graded integrity signals, will keep the speed of remote hiring without inheriting its new fraud problem.