How to Detect AI Cheating in Video Interviews: Patent-Pending Multi-Modal Methods That Actually Work

A mid-sized SaaS company hired a senior backend engineer in Q3 2025. His interview was flawless — structured answers, precise technical vocabulary, clear system design thinking. He started on a Monday. By Thursday, his team lead pulled HR aside: "He can't write a basic database query without help. Something's wrong."

It took three weeks and $50,000+ in onboarding, salary, and team disruption to confirm what they suspected: he'd used an invisible AI overlay tool during his video interview. Every answer he gave had been generated by an LLM and displayed on his screen in a transparent layer that was completely invisible to the interviewer and their proctoring software. The company's tab-switch monitoring never flagged a thing — because he never switched a tab.

This isn't a fringe case anymore. It's the new normal.

Hoogway.ai detects AI-assisted cheating using a patent-pending multi-modal approach that analyzes three layers simultaneously: eye movement and gaze patterns (catching off-screen reading), voice modulation (distinguishing spontaneous speech from script-reading), and transcript analysis (identifying AI-generated versus authentic responses). If one detection layer misses something, the other two catch it.

The Numbers That Should Scare Every Hiring Team About AI Cheating in Video Interviews

Interview cheating crossed from "occasional concern" to "systemic problem" in late 2025:

• 38.5% of video interviews triggered cheating flags in Fabric's analysis of 19,368 sessions (July 2025 – January 2026). The rate jumped from 9% to 45% in just three months.
• 20% of U.S. workers admitted to secretly using AI during job interviews (Blind, 2025).
• 59% of hiring managers suspect candidates of AI-assisted misrepresentation (Gartner/Sherlock AI, 2026).
• 35% of candidates showed cheating signals by December 2025, more than double the 15% rate from six months earlier (Fabric, 2026).
• Gartner projects 1 in 4 candidate profiles will be entirely fake by 2028.

The question for hiring teams isn't whether cheating happens. It's whether your detection can keep up with the tools candidates are using.

Why Your Current Proctoring Doesn't Work

Browser-based proctoring — tab-switch detection, lockdown browsers, screen monitoring — was built for a simpler era. When "cheating" meant opening Google in another tab.

In 2026, the cheating tools have outrun those defenses entirely:

Invisible overlay tools (Cluely, Interview Coder, Leetcode Wizard) use low-level GPU hooks to render AI-generated answers directly on the candidate's screen. These overlays are invisible to screen sharing. The interviewer sees a clean screen. The candidate reads floating answers. Adoption of these tools doubled from 15% to 35% between June and December 2025 (Fabric).

Real-time audio transcription tools convert the interviewer's questions to text, feed them to an LLM, and display multiple answer options in real-time. The candidate picks the most natural-sounding one and reads it aloud.

Deepfake and proxy setups enable a qualified expert to answer questions while the candidate on screen lip-syncs using a face overlay. The FBI has issued warnings about state-sponsored actors using these techniques.

Tab-switch monitoring catches roughly 18% of cheating methods (Fabric, 2026). The remaining 82% requires behavioral signal analysis that browser-based proctoring simply cannot provide.

What a Missed Cheat Looks Like

A software engineering candidate is asked to explain how she'd design a rate-limiting system.

What the interviewer sees: She pauses for 3 seconds, then delivers a flawless answer covering token bucket algorithms, sliding window counters, and distributed rate limiting across microservices. Eye contact looks normal. No tabs switched. Impressive.

What actually happened: During that 3-second pause, an invisible overlay rendered the AI-generated answer on the lower-left of her screen. She read it while maintaining approximate camera direction. Her delivery was smooth because the tool offered three response options and she picked the one matching her natural vocabulary level.

What the old proctoring caught: Nothing.

The Cheating Tool Ecosystem

Understanding what you're defending against explains why multi-modal detection is necessary:

Tier 1: Free/Low-Cost — ChatGPT in a second browser or device, friend on earpiece, notes on a second monitor. Detection difficulty: low. Behavioral signals are obvious.

Tier 2: Dedicated Software ($20–50/month) — Invisible overlay tools, real-time audio-to-LLM transcription. Detection difficulty: medium. Overlays are invisible to screen monitoring, but gaze patterns and voice delivery still show reading behavior.

Tier 3: Advanced Fraud ($50–200+) — Deepfake face/voice cloning, professional proxy services, combined overlay + coaching systems. Detection difficulty: high. Requires multi-modal analysis.

The economics make the math clear: a $50/month cheating subscription versus a $150,000 engineering salary creates a risk-reward ratio that heavily favors cheating — unless detection makes it reliably risky.

The Candidate's Experience of Cheating (Why It's So Tempting)

Understanding why cheating is exploding helps explain why detection can't rely on candidate goodwill:

A software engineer preparing for interviews in 2026 sees TikTok videos showing Cluely in action. She watches someone get a perfect interview score using an invisible overlay. Her LinkedIn feed shows peers landing offers at companies she's applying to. She knows some of them are using these tools. She's been job searching for 3 months after a layoff. She has a mortgage.

The "prisoner's dilemma" kicks in: if she doesn't cheat but her competition does, she's at a disadvantage. If everyone cheats, the honest candidate is the only one penalized. Fabric's research confirms this pattern — the spike in late 2025 was driven not by individual bad actors but by a social cascade where candidates saw others succeeding with these tools and felt compelled to keep up.

This is why detection needs to change the equation, not rely on honor systems. When cheating is reliably detectable, the rational calculation flips: using a $50 tool that will likely get you flagged and cost you the opportunity becomes a bad bet. Multi-modal detection doesn't just catch cheaters — it discourages cheating by making it risky enough that honest candidates are no longer penalized for being honest.

In other words: the point of detection isn't punishment. It's restoring fairness.

How Hoogway's Three-Layer Detection Works

Hoogway's patent-pending malpractice agent runs during proctored async video interviews and analyzes three independent signal layers:

Layer 1: Eye Movement and Gaze Analysis

Normal behavior: Eyes move upward or sideways when accessing memory, blink at natural rates, gaze returns to camera when speaking.

Cheating behavior: Consistent horizontal scanning (reading lines of text), gaze fixation at screen edges where overlays render, reduced camera contact during answers.

What it catches: Side-screen reading, overlay text consumption, secondary monitor use.

Layer 2: Voice Modulation Analysis

Normal behavior: Irregular pacing, tonal variation, pauses at natural thought boundaries, energy that reflects engagement.

Cheating behavior: Flatter tonal range, uniform reading-speed pacing, pauses at line breaks rather than thought boundaries, disconnect between answer sophistication and delivery energy.

What it catches: Script-reading, AI response reading, proxy audio coaching.

Layer 3: Transcript AI-Generation Analysis

Normal behavior: False starts, self-corrections, tangential details, personal anecdotes, vocabulary consistent with demonstrated experience level.

Cheating behavior: Perfect structure even in casual speech, thorough coverage without tangents, vocabulary exceeding what resume and earlier answers suggest, consistent style regardless of question difficulty.

What it catches: ChatGPT-drafted responses, LLM-generated explanations, AI-polished answers exceeding genuine capability.

Why Three Layers Beat One

Any single method has blind spots. Beat gaze detection by memorizing AI answers. Beat voice analysis by practicing natural delivery. Beat transcript analysis by paraphrasing AI bullet points.

But beating all three simultaneously is extremely difficult. If you read from a screen (Layer 1 catches it), or read with flat delivery (Layer 2 catches it), or deliver suspiciously perfect content (Layer 3 catches it) — the combined signal flags the concern even when individual layers show ambiguous results.

What the Hiring Manager Sees

Integrity Score: 0–100 confidence that responses were authentic.

Layer Breakdown: Eye/gaze score + timestamped flags ("Sustained left-screen reading at 4:32–4:55 during Question 3"). Voice score + delivery analysis. Transcript score + AI-generation probability per answer.

Evidence Timeline: Visual timeline showing where flags occurred, linked to those moments in the recording.

This is evidence for a human decision — not an automatic rejection.

Hoogway vs Other Detection Approaches

No system catches 100%. A candidate who memorizes AI content and delivers it naturally has effectively learned the material — a different problem from real-time cheating. Hoogway catches the vast majority of AI-assisted cheating during the actual interview.

What About Neurodivergent Candidates?

This is an important question that most cheating detection articles dodge.

Candidates with ADHD, autism spectrum conditions, or other neurological differences may naturally exhibit atypical gaze patterns or speech rhythms. Someone with ADHD might look away frequently not because they're reading an overlay, but because that's how their attention works. A candidate on the autism spectrum might deliver answers with less tonal variation not because they're reading a script, but because that's their natural communication style.

Hoogway addresses this with intentional design: integrity scores are confidence levels (0–100), not binary labels. A score of 65 doesn't mean "cheater" — it means "review this candidate's flagged moments and apply your judgment." The system surfaces evidence for a human decision, specifically because behavioral diversity exists and edge cases require human context that algorithms don't have.

Teams should calibrate their threshold with awareness of this. Some organizations set a lower bar for initial screening (flag only below 40) and a stricter bar for regulated roles (flag below 60). The important thing is that no candidate is auto-rejected based on an integrity score alone.

Implementing Multi-Modal Detection: Step by Step

Step 1: Audit your current detection. Have someone on your team attempt an interview using a second device with ChatGPT open. If your system doesn't flag it, you have a gap. Most teams are surprised by how easy it is to beat their existing proctoring.

Step 2: Define your integrity policy. Before deploying detection, decide: What score triggers review? What score blocks advancement? Do flagged candidates get a supplementary live round? Document these decisions for compliance.

Step 3: Integrate into your video stage. Hoogway's detection runs natively during the proctored async interview. No separate proctoring vendor, no candidate-side software, no friction.

Step 4: Train hiring managers on reading integrity data. An integrity score of 45 with a timestamped flag at minute 4:32 is useful information — but only if the manager knows what to do with it. Build 15 minutes of integrity-data training into your hiring manager onboarding.

Step 5: Document everything for compliance. NYC Local Law 144, the EU AI Act, and expanding state regulations expect documented integrity processes. Timestamped evidence with defined criteria provides the audit trail regulators want to see.

For US Regulated Industries and India High-Volume Hiring

United States (Regulated Industries)

In financial services, healthcare, and defense contracting, hiring a fraudulent candidate isn't just an HR problem — it's a security and compliance risk. The FBI has specifically warned about AI-enhanced identity fraud in remote hiring targeting defense and government contractors. Hoogway's integrity scoring provides documented evidence for compliance files, and the multi-modal approach satisfies the "reasonable measures" standard that regulators increasingly expect.

For healthcare organizations subject to credentialing requirements, a candidate who fakes technical competency in an interview but can't perform once hired creates patient safety risks that go far beyond the cost of a bad hire.

India and High-Volume Markets

When processing 1,000+ candidates for BPO, IT staffing, or campus roles, cheating rates compound fast. If 35% of candidates use AI assistance (Fabric, 2026), that's 350 potentially misrepresented interviews in a single batch. Manual detection — having a human watch each recording for signs of cheating — is impossible at this volume. Automated multi-modal detection makes integrity verification feasible exactly where it matters most: high-volume hiring where you can't manually review every interview.