38% of Your Tech Candidates Are Using AI to Cheat — and Your Live Coding Doesn't Catch It

38.5% of tech candidates showed signs of AI cheating in technical interviews between July 2025 and January 2026. This data comes from Fabric, based on a sample of 19,368 interviews. In purely technical roles, the number rises to 48%. Most companies still conduct live coding interviews over Zoom as if it were 2019.

How big is the problem really?

Cheating in technical interviews isn’t marginal — it’s the baseline scenario. Between June and December 2025, the percentage of candidates using unauthorized AI rose from 15% to 35%. In Fabric’s study of 19,368 interviews (Jul 2025 – Jan 2026), the picture is starker: 48% cheating in technical roles, compared to just 12% in sales. Juniors (0–5 years of experience) cheat at twice the rate of seniors.

The most alarming number isn’t the prevalence. It’s this: 61% of candidates who cheated passed the approval threshold (score ≥7.0). Without detection, these individuals advanced to the next stage. Meanwhile, Karat (the coding interview company) estimates that 80% of candidates use LLMs during code tests even when explicitly prohibited.

In the processes we observe at WeRecruitIT, the pattern aligns with the data: flawless screenshots, code that works on the first try, but the candidate can’t explain why they chose that structure when asked a follow-up question.

Why doesn’t traditional proctoring catch it?

Because the 2025 generation of cheating tools isn’t designed to be detected by Zoom or Google Meet. They’re designed to operate beneath the surface.

The most used tools (Cluely, Interview Coder, Leetcode Wizard) employ low-level graphics hooks: DirectX on Windows, Metal framework on macOS. They render responses as an invisible overlay drawn over the GPU output, beneath the layer that captures screen-sharing. The interviewer sees the candidate’s screen “clean.” The candidate sees the LLM’s response overlaid on their IDE.

How candidates cheat today (Fabric, 19,368 interviews):

Method	% of Cases	Detectable by Zoom/Meet
Dedicated assistants (Cluely, Interview Coder)	45%	No
Voice mode LLMs (ChatGPT in the background)	34%	Partially
Tab switching / second screen	18%	Yes
Live human help	3%	Sometimes

Specific proctoring vendors claim 85–95% detection. Plausible for standard flows, but the arms race has already begun: in 2025, six new tools like Cluely appeared, along with at least three open-source clones. Detection will always lag one or two steps behind.

What if the solution isn’t catching the cheater?

This is the uncomfortable part. The dominant narrative is “we need better proctoring.” The correct narrative might be different: redesign the assessment so cheating becomes irrelevant.

Look at the data from another angle. In 2026, according to Dice, 71% of tech job postings in the U.S. require some form of AI fluency — a 181% year-over-year increase. In other words, companies no longer want developers who code without AI; they want developers who code with AI. If you’re hiring someone to use Copilot all day, why conduct an interview that prohibits using Copilot?

Enter the “pair programming with AI” format, which IEEE-USA describes as the emerging replacement for classic live coding. The typical structure has three phases:

1. Problem decomposition. The candidate breaks down an ambiguous requirement. The interviewer evaluates analytical criteria — something that can’t be faked with an LLM because the question is discursive.

2. AI-assisted implementation. The candidate implements using AI explicitly allowed. What the interviewer observes is how they prompt, how they iterate, what they discard. Not the output.

3. Code review. The candidate reviews AI-generated code with planted bugs and identifies them.

These three stages measure things an overlay LLM can’t solve for you: criteria, communication, judgment. They remove the incentive to cheat because AI is already on the table.

Is the take-home test viable again?

Yes — but with a different version of the 2019 take-home test. According to CoderPad’s State of Tech Hiring 2025, 68% of companies now incorporate take-home tests (+12% year-over-year), and 41% run a hybrid model: take-home followed by a live review session where the candidate defends their submission.

Let’s define quickly: a structured take-home is an asynchronous task with a limited scope (1–4 hours), with a real product objective (not an algorithm puzzle), followed by a 30–45 minute conversation where the candidate walks the interviewer through their solution, justifies decisions, and answers follow-ups.

The structured take-home works even in an AI world because the question shifts from “did they write this?” to “can they explain each decision and what they would have done differently?” If Cursor wrote the code and the candidate doesn’t understand it, it falls apart in 5 minutes of defense.

McKinsey, in its Technology Talent Report 2025, found that companies using structured take-homes have 41% fewer early departures than those using only live interviews. The signal isn’t perfect, but it’s better.

What should you change in the next 90 days?

Three concrete moves, in order of impact:

First, accept that AI is part of the interview. Don’t ban it in a phase of the process. Design a phase where it’s explicitly allowed and where the observation is on how the candidate uses it.

Second, replace the whiteboard algorithm. Shift the first technical filter to a short work sample (1–2 hours) that looks at realistic work, not LeetCode.

Third, add the live defense. Any take-home should conclude with 30 minutes of “explain this to me,” where an overlay LLM doesn’t help and where a real candidate demonstrates criteria.

The 38% isn’t a statistic about candidates. It’s a statistic about the interview format you’re running. As long as the format remains “solve this algorithm live in a shared IDE,” the number will keep rising. And half of those who pass won’t know why their code works.

The most costly part of a bad hire isn’t the salary. It’s what that person breaks in the first six months before you realize it. If your evaluation process doesn’t separate signal from noise in 2026, you’ll pay that difference more times than you’d like.