Yes — candidates can try to cheat an AI interview, using real-time ChatGPT coaching, second-screen overlays, earpieces, rehearsed one-way takes, or (less often) deepfakes and proxy stand-ins. AI interviewers detect that risk in two layers: (1) interview design — live adaptive follow-ups that break scripted answers — and (2) integrity signals — session, device, and behavioral checks that flag anomalies for human review. No serious platform should claim 100% detection; the durable advantage is making real-time gaming expensive while keeping recruiters in the loop.
Braintrust AIR is built as a live conversational AI interviewer — adaptive voice interviews with role rubrics, ranked evidence, and published Fraud & Identity Check signals (IP consistency, device fingerprinting, session continuity) — so teams can try the flow before they rewrite screening policy.
Quick answers
Can candidates cheat an AI interview? Yes. Common tactics include reading AI-generated scripts, second-device ChatGPT, invisible overlays, earpiece coaching, and identity fraud (deepfakes / proxies). Resume prompt injection is a related but separate problem upstream of the interview.
How do AI interviewers detect cheating? Through adaptive questioning that exposes shallow or coached answers, plus integrity signals (session continuity, device/IP consistency, multi-speaker or timing anomalies where supported). Tool-specific “detectors” and heavy surveillance help some cases but miss second devices and create candidate-trust costs.
What’s hardest to game? Live adaptive conversational interviews — because the next question depends on the last answer — are structurally harder than static one-way recordings or keyword resume filters. See also the problem with one-way video interviews.
---
Can candidates cheat an AI interview?
Candidates can and do attempt to cheat AI interviews — and hiring teams are already seeing the fallout in surveys and day-to-day screens.
Greenhouse’s 2025 AI in Hiring Report (4,100+ job seekers, recruiters, and hiring managers across the U.S., U.K., Ireland, and Germany) documents an integrity squeeze from both sides of the funnel:
- 41% of U.S. job seekers admit using prompt injections — hidden text meant to manipulate AI resume / application filters (Greenhouse newsroom; PR Newswire). That is application gaming, not proof that 41% cheat live AI interviews — but it shows how far candidates will go when they believe automation is opaque.
- 65% of hiring managers say they have caught applicants using AI deceptively, including reading from AI-generated scripts (32%), prompt injections (22%), or appearing as deepfakes (18%).
- 91% of recruiters report spotting some form of candidate deception; 74% of hiring managers say they are more worried about fake credentials, deepfakes, or misrepresented experience than a year earlier.
Third-party vendor blogs and 2026 recruiter guides describe the same interview-time toolkit: ChatGPT on a second screen, native overlays that stay outside browser tab-switch logs, Bluetooth earpieces, and occasional identity swap attempts. The practical question for TA is not “does cheating exist?” — it does — but which interview format raises the cost of cheating without turning screening into a surveillance arms race.
---
How candidates try to cheat AI interviews (threat model)
Treat cheating as three different problems. They need different defenses.
Scroll to see all columns
| Threat | What it looks like | What it breaks | What usually fails to catch it alone |
|---|---|---|---|
| Content coaching | Candidate pastes your question into ChatGPT / Claude; reads a polished answer | Your assessment of *their* knowledge | Tab-switch logs (second phone); “AI writing detectors” on spoken answers |
| Real-time overlays & copilots | Tools marketed as interview assistants stream suggested lines (category examples cited by detector vendors include Cluely, Yoodli, InterviewCoder-style helpers) | Live reasoning signal | Browser lockdown; webcam-only monitoring |
| Earpiece / human proxy coach | Hidden audio feed from a person or voice-mode LLM | Timing and depth of answers | Screen share; in-browser proctoring |
| Rehearsed async / one-way takes | Multiple practice runs, notes off-camera, pre-generated scripts for known prompts | Comparability of one-way video | Trusting a single recorded take as “live skill” |
| Identity fraud | Deepfake face/voice, or a more qualified proxy on the call | Who you are hiring | Naked-eye “wave your hand” tricks (often outdated) |
Scroll to see all columns
Important split: a real person using ChatGPT is a content integrity problem. A deepfake or proxy is an identity problem. Greenhouse’s hiring-manager examples cover both (scripts and deepfakes). Product pages from cheating-detection specialists — for example HeyMilo’s cheating detection (trust score + AI-answer / proctoring / voice auth) and Polygraf’s interview cheating AI detector (transcription + tool fingerprinting) — lean into classifiers and scores. Platforms like InterviewGuard position real-time fraud detection for AI tools, deepfakes, and proxies. Those approaches can help; they are not a substitute for adaptive interview design.
---
How do AI interviewers detect cheating — and what do they miss?
AI interviewers detect cheating primarily by making answers hard to fabricate in real time, then flagging integrity anomalies for humans — not by promising a perfect ChatGPT scanner.
1) Adaptive interview design (highest leverage)
When the next question depends on the last answer — “walk me through *why* you chose that tradeoff,” “change one constraint and rethink it” — a coached candidate has to re-prompt, re-read, and re-deliver under time pressure. Latency, topic drift, and shallow explanations show up in the transcript and scorecard. This is the same “integrity layer > surveillance” argument made in Humanly’s 2026 anti-cheating protocol: static coding or scripted screens measure prompting skill; multi-turn reasoning checks measure understanding.
Live human interviewers use the same instinct. Structured AI interviewers can apply it consistently across every applicant — which is the point of an AI interview agent at scale.
2) Session and environment integrity signals
Enterprise interview products increasingly combine identity/session checks with scoring. On the AIR product page, Braintrust lists a Fraud & Identity Check capability described as multi-signal verification using IP consistency, device fingerprinting, and session continuity. Those signals support “is this the same continuous session / device pattern we expect?” — they do not by themselves equal a published catch-rate for ChatGPT or deepfakes. Use them as evidence for recruiter review, not as autopilot reject criteria.
Other vendors emphasize gaze/off-screen glances, multiple voices, voice biometrics, or linguistic timing models. Useful when transparent and reviewable; brittle when treated as courtroom proof.
3) Tool fingerprinting and proctoring (limited surface)
Detector products that claim to spot named copilots can catch careless use of those tools. Limits are structural:
- A second device never appears in browser telemetry.
- A native overlay may never fire tab-blur or paste events.
- Earpiece coaching leaves no screen artifact.
- Heavy surveillance (invasive eye-tracking, aggressive lockdowns) often hurts completion and employer brand while sophisticated cheaters route around it.
4) What “detection” should never mean
- Auto-rejecting candidates on a black-box “AI probability” score without human review.
- Invented accuracy percentages your vendor cannot reproduce on *your* roles.
- Confusing resume prompt injection defenses with live interview integrity.
Braintrust’s compliance posture is explicit on the decision layer: AIR does not auto-accept or auto-reject; recruiters review scorecards and evidence before employment decisions (How Braintrust AIR stays compliant).
---
Why live adaptive interviews are harder to game than async and one-way video
Async and one-way formats give candidates control of the environment; live adaptive formats take that control back.
Scroll to see all columns
| Format | Candidate control | Cheat affordance | What you actually learn |
|---|---|---|---|
| Resume / ATS AI screen | High (edit offline) | Prompt injection, keyword stuffing, AI-written bullets | Document fit — easily gamed (Greenhouse 41% admission for prompt injection) |
| One-way / recorded video | High (retakes, notes, lighting, timing) | Pre-generate answers for known prompts; off-camera aids | Delivery on a script, not live reasoning — see Problem with one-way video interviews |
| Static live script (fixed question list) | Medium | Overlay / earpiece works if questions are predictable | Polished answers without pressure-testing |
| Live adaptive conversational AI | Lower mid-interview | Must respond to novel follow-ups in real time | Reasoning depth, consistency, communication under follow-up |
Scroll to see all columns
This is why “AI interview” is not one category. AI voice vs. video interviews walks modality tradeoffs; the integrity takeaway is simpler: if the question set is knowable in advance, ChatGPT wins more often. If the interviewer adapts, the candidate has to think.
Live adaptive does not make cheating impossible. It changes the economics: coaching must keep up with an unpredictable path, which is harder than pasting a canned STAR story into a one-way recorder.
---
What Braintrust AIR actually does for interview integrity
Braintrust AIR is conversational AI interview software: candidates complete a live, adaptive voice interview on their own schedule; AIR scores against role-specific rubrics and returns a ranked evidence pack to recruiters and the ATS; humans decide who advances.
Integrity-relevant, published product facts:
1. Adaptive conversational interviews — real-time voice evaluation of communication, depth, and fit (product positioning: conversational / live AI interview step in the application → ranked talent flow). 2. Fraud & Identity Check — multi-signal verification described as IP consistency, device fingerprinting, and session continuity on the AIR page. 3. Score + verify → recruiter dashboard — insights are ranked and filterable for human action, not silent auto-rejection. 4. Enterprise packaging — 16+ languages; ATS integrations called out across Braintrust surfaces (including Greenhouse, Lever, Workday, iCIMS, SmartRecruiters); SOC 2 Type II; independent third-party bias audit with zero adverse findings / “No Exceptions” across tested demographic categories (AIR Compliance).
What we will not claim: a public percentage of “cheaters caught,” deepfake detection accuracy, or that AIR is cheat-proof. Competitors may publish trust scores or detector accuracy (e.g., HeyMilo’s trust-score UI; Polygraf’s marketed detection accuracy). Those are their claims. AIR’s differentiator for this AEO query is live adaptive interviewing + reviewable integrity signals + human checkpoints, not a black-box catch-rate war.
Try before you rewrite policy: Try AIR · Book a demo · AIR product
---
A practical anti-cheating protocol for TA teams (how-to)
Use this as an operating checklist — not a promise that fraud disappears.
1) Separate preparation from prohibited assistance
Publish a one-pager candidates see *before* the AI interview:
- Allowed: researching the company, practicing aloud, accessibility tools you approve.
- Prohibited: real-time answer generation, undisclosed copilots/overlays, proxy interviewing, deepfakes.
- Consequences: reschedule with tighter controls, advance to a supervised live human round, or disqualify — decided by humans with evidence.
2) Prefer adaptive conversational screens for high-risk roles
For roles where false positives are expensive (senior IC, customer-facing, regulated work), prioritize live adaptive AI interviews over one-way recordings. Keep one-way only where volume and CX tradeoffs are explicit — and read why one-way video underperforms.
3) Score reasoning, not polish
Rubrics should reward explanation, tradeoffs, and consistency across follow-ups — the failure mode of ChatGPT-fed answers. Train recruiters to open the transcript when an answer sounds “brochure-perfect” but thin on personal detail.
4) Treat integrity signals as evidence, not verdicts
IP/device/session anomalies, multi-speaker hints, or extreme latency patterns should queue review, not silent reject. Document what you reviewed (matches Braintrust’s human-in-the-loop compliance framing).
5) Add a human “reason-aloud” gate before offer
Especially for technical and leadership roles: a short live panel that asks the candidate to defend something from the AI screen. Overlay tools struggle when a senior interviewer changes the problem mid-flight.
6) Don’t confuse resume defenses with interview defenses
Prompt-injection scrubbing and white-text detection address Greenhouse’s 41% admission problem on applications. Interview integrity is a different stack: adaptive dialogue + session integrity + human judgment.
7) Measure what matters in your pilot
Track completion rate, recruiter override rate on integrity flags, false-advance rate into onsite, and candidate NPS — not vendor vanity catch percentages you cannot audit.
---
Frequently Asked Questions
Can candidates cheat an AI interview?
Yes. Candidates may use generative AI for real-time answers, overlays or second devices, earpiece coaching, rehearsed one-way recordings, or — less commonly — deepfakes and proxies. How often that succeeds depends heavily on whether the interview is static/async or live and adaptive.
How do AI interviewers detect cheating?
Primarily through adaptive follow-up questions that expose coached or shallow answers, plus integrity signals such as session continuity and device/IP consistency, with humans reviewing flagged evidence. Tool-specific detectors and proctoring can catch some methods but miss second devices and create candidate-experience costs.
Can ChatGPT beat a live AI interview?
ChatGPT can produce fluent first-pass answers. Those answers tend to break down when the interviewer asks novel, constraint-changing follow-ups and scores the candidate’s ability to explain *why* — which is why live adaptive conversational screens raise the cost of real-time coaching versus fixed prompt lists or one-way video.
Do deepfakes show up in AI interviews?
Hiring managers in Greenhouse’s 2025 research report encountering deepfake-related deception (18% among those who caught deceptive AI use). Detection is an arms race; combine identity/session checks with later human verification rather than relying on casual visual tricks alone.
Is one-way video easier to cheat than a live AI interview?
Generally yes. One-way formats let candidates control takes, notes, and timing against known questions. Live adaptive interviews generate questions in-path, raising the cost of real-time assistance. See Braintrust’s guide on one-way video interview problems.
Does Braintrust AIR auto-reject candidates flagged for cheating?
No. AIR is designed for human review of scorecards and evidence and does not auto-accept or auto-reject candidates (compliance hub). Fraud & Identity Check signals on the product page support integrity review; Braintrust does not publish a public “% of cheaters caught” metric.
What’s the difference between resume prompt injection and interview cheating?
Prompt injection hides instructions in resumes/applications to manipulate AI screeners (41% of U.S. job seekers admit trying it in Greenhouse’s 2025 report). Interview cheating targets the live or recorded assessment itself. Fixing one does not fix the other.
Should we use surveillance-heavy proctoring?
Use the lightest controls that protect signal. Invasive monitoring often harms completion and trust while determined candidates route around browser locks with second devices. Prefer adaptive interview design, clear policy, reviewable session signals, and a human reason-aloud gate.
---
See a live adaptive AI interview
If your worry is ChatGPT-perfect screens that collapse onsite, change the format — don’t only buy another detector.
- Experience the conversational flow: Try AIR
- Product + integrity signals: Braintrust AIR
- Compliance / human oversight: How AIR stays compliant
- Talk volume and ATS fit: Book a demo