Cheating On Technical Assessments Doubled In A Year (16% → 35%). Google & McKinsey Are Bringing Back In-Person Interviews. Here's The Middle Path.

CodeSignal’s February 2026 data made the rounds quietly, but the number deserves more attention: technical assessment cheating rose from 16% to 35% in 12 months. That’s not a gradual trend. It’s a doubling in a single hiring cycle, driven almost entirely by the accessibility of AI-generated responses.

Separately, CBS News reported that 50% of businesses have now encountered AI-driven deepfake fraud in hiring — candidates presenting AI-generated video, voice, or written responses as their own work.

The responses from the market have been predictable. Google and McKinsey reinstated mandatory in-person interview rounds (WSJ, mid-2025). Assessment vendors like InCruiter launched deepfake detection layers, flagging fraud in 25–30% of suspicious sessions. Sherlock, Metaview, and others are building detection into their platforms.

None of these solutions is wrong. All of them are incomplete. And for high-volume campus hiring in India — where you’re screening 3,000 candidates across 14 days — most of them simply don’t scale.

Here’s the middle path.

The Scale of the Problem: Three Data Points

Before examining solutions, it’s worth understanding the full shape of the problem.

CodeSignal (February 2026): Cheating on technical assessments doubled from 16% to 35% in 12 months. The primary driver: large language models that can generate accurate, well-formatted technical answers in seconds. The secondary driver: asynchronous assessment formats that give candidates unlimited time to consult these tools.

CBS News survey: 50% of businesses have encountered AI-driven deepfake fraud in hiring. This includes candidates submitting AI-generated video interviews, AI-cloned voice responses, and AI-written work samples as their own.

InCruiter deepfake detection (2026): Their detection system flagged fraud in 25–30% of suspicious sessions — approximately twice the rate that human interviewers caught the same fraud. The detection works. The problem is it’s reactive: you detect fraud after the candidate has completed the assessment, which means you’ve already invested the screening infrastructure.

The underlying challenge: the tools that candidates use to cheat are improving faster than the tools used to detect cheating. Detection is a permanent arms race.

How Companies Are Responding — And Why Each Falls Short

Response 1: Reinstate In-Person Interviews (Google, McKinsey)

The logic is sound. If a candidate is physically present in the room, AI-generated responses and deepfake video are not available to them. The integrity problem is largely solved.

The scale problem is not. Google and McKinsey can enforce in-person requirements because they have the infrastructure, the brand pull to demand it of candidates, and the candidate volume that makes in-person logistics manageable.

A company running a 3,000-candidate campus drive cannot bring 3,000 candidates to a physical location for first-round screening. The cost is prohibitive. The logistics are impossible during a 14-day offer window. And in India’s campus hiring context — where drives span multiple cities and university campuses simultaneously — physical in-person screening at first-round scale was never a practical option to begin with.

In-person works for final rounds. It doesn’t solve the first-round problem.

Response 2: Add Detection Layers (InCruiter, Sherlock, Metaview)

Detection platforms represent real innovation. Flagging deepfake video, identifying AI-generated text patterns, and monitoring browser behaviour during assessments are technically sophisticated capabilities.

The fundamental limitation: detection is after-the-fact. You learn that a candidate cheated after they’ve completed your assessment, consumed your screening capacity, and potentially passed through to the next round before the flag surfaces.

Detection also creates false-positive risk. Aggressively flagging suspicious sessions catches fraud but also catches legitimate candidates who happen to type quickly, have unusual response patterns, or use accessibility tools. In high-volume campus hiring, false positives have real costs — both in candidates you wrongly exclude and in the recruiter time required to review flagged sessions.

Detection is a useful layer in a comprehensive fraud strategy. It’s not a substitute for a fraud-resistant architecture.

The Third Path: Fraud-Resistant by Construction

The alternative to detection is prevention. Not physical presence — which doesn’t scale — but a screening format where AI-generated scripted responses don’t work.

This is where real-time conversational AI screening changes the equation.

In a text-based or pre-recorded video assessment, a candidate can pause, consult ChatGPT, copy a generated response, and submit it. The format gives them the time to do this, and the asynchronous nature means there’s no one to notice.

In a live voice conversation — where the AI interviewer is generating follow-up questions based on what the candidate just said — the scripted response breaks almost immediately.

Here’s why: ChatGPT-scripted answers are optimised for the question that was asked, not for the conversation that follows. When a voice screening agent responds to a candidate’s answer with a follow-up that branches off their specific phrasing, their specific example, or a specific claim they made — the scripted answer has no branch for that.

The candidate who read a ChatGPT-generated response about “a time they demonstrated leadership” will deliver it fluently. The AI interviewer who then asks “you mentioned the team disagreed with your approach — what was the specific objection and how did you address it in the moment?” is asking about something the script didn’t anticipate. The follow-up reveals whether the candidate was there or whether they borrowed the story.

This isn’t detection. It’s prevention by conversational architecture.

How Real-Time Conversational Probing Breaks Scripted Answers

The mechanism is worth understanding in detail, because it’s different from anything the detection approach offers.

A static assessment — written or pre-recorded video — presents a fixed set of questions. The candidate knows every question before they begin (or can find them online). AI tools are trained to answer exactly these question formats.

A dynamic voice conversation has no fixed script after the opening question. Each exchange generates the next question based on the previous answer. The candidate cannot know what they will be asked because the questions are generated in response to them specifically.

The adversarial test: can a candidate use AI assistance to navigate a dynamic conversational interview in real time? In theory, yes — they could type their answer into an AI tool and read the suggested response. In practice, the latency, the cognitive load of simultaneously listening, prompting an AI, and reading a response aloud while maintaining a natural conversational cadence is prohibitive. The tells are detectable within the first three exchanges.

For campus hiring at scale — where you’re screening for communication skills, structured thinking, and genuine engagement — the conversational format does double duty: it screens for the competencies you care about and it’s structurally resistant to the fraud vector you’re trying to prevent.

What This Means for High-Volume Campus Hiring

The cheating problem is worst at the exact moment it causes the most damage: first-round screening at scale, where volume is high, individual scrutiny is low, and fraudulent candidates who pass through consume interview slots and offer capacity before being identified.

For a company running 3,000-candidate drives, a 35% cheating rate on technical assessments means 1,050 fraudulent responses in the first-round pool. Detection-after-the-fact means those 1,050 candidates have already been through your process before you know.

Prevention-by-architecture means the conversational format itself filters them — not through a flag, but through the natural dynamics of a conversation they can’t script.

HireQwik’s live voice screening handles 200+ simultaneous conversations, classifies candidates into Strong Go / Go / On Hold / No Go, and produces a full transcript for every session — so your team can review the conversation that produced each classification.

If you want to hear what a fraud-resistant voice screen sounds like and see how the conversational probing works in practice, connect with the HireQwik team.