Technical interviews built around take-home coding tests and unmonitored async assessments have a problem in 2026 that didn't exist a few years ago: AI coding assistants make them nearly impossible to trust as a signal. The data on this is stark enough that most engineering orgs have had to rethink their process rather than patch around it.
How big the problem actually is
Across one large dataset of AI-led interviews run between mid-2025 and early 2026, roughly 38.5% of candidates were flagged for AI-assisted cheating overall, with software engineering roles hitting closer to 48%. The more concerning number: a majority of flagged candidates still scored above the passing bar — meaning the people using AI assistance to cheat weren't just attempting it, they were largely succeeding and moving forward in the process undetected by the scoring itself.
Take-home assessments and asynchronous automated code tests are the formats most exposed to this problem, since there's no way to confirm a candidate isn't running the problem through an AI assistant in another window while "solving" it.
The two strategies that are actually working
Companies responding effectively are converging on one of two approaches, rather than trying to detect and block AI use (which has proven to be an arms race that detection tools mostly lose):
1. Live, adaptive coding interviews. Rather than async take-homes, candidates work through problems live with an engineer, who can throw in an unscripted curveball or change a requirement mid-session. This isn't about the interviewer trying to trip people up — it's that the interaction itself is the signal. Watching how someone responds to an unscripted follow-up, or asking them to defend a decision they didn't originally make, is very hard to fake with an AI assistant feeding answers in real time.
2. Allow AI use openly, with rules. Some companies have flipped the framing entirely: since developers will use AI assistants on the job, structure the interview to allow AI tools openly and evaluate how well the candidate uses them — how they prompt, how they verify output, how they catch and fix what the AI gets wrong. This tests a skill that's now genuinely part of the job, rather than pretending the job is done without AI assistance.
A related variant: replacing the most cheatable round with a conversational AI-driven interview whose structure — deep, adaptive follow-up questions — makes surface-level AI assistance irrelevant, because the questions are generated in response to what the candidate just said, not from a fixed bank that can be pre-solved.
What doesn't work
Simply banning AI use and hoping candidates comply doesn't hold up under the data — detection is unreliable, and the incentive to use it undetected is high when a job offer is on the line. Similarly, doubling down on harder async take-homes just raises the bar for what AI-assisted cheating needs to clear, without addressing the underlying problem that async, unmonitored assessments can't verify who — or what — actually produced the answer.
LeetCode-style prep is losing relevance for a related, separate reason
Beyond the cheating-detection problem, there's a second and independent reason companies are moving away from whiteboard-style algorithmic puzzles in 2026: AI made the exercise itself nearly meaningless as a signal, even setting aside cheating entirely. With current AI coding assistants, answering a "Medium" difficulty LeetCode-style problem takes roughly 4 seconds — which means a format built around solving a bounded algorithmic puzzle under time pressure was already measuring something increasingly disconnected from real engineering judgment, independent of whether a specific candidate used AI assistance to cheat on it. Hiring teams have also observed that the hardest algorithmic puzzle questions produced high false-negative rates even before AI entered the picture — strong engineers who would have been excellent senior hires got rejected because they couldn't invent a specific algorithmic pattern from memory within a 45-minute window, which was never a great proxy for on-the-job engineering effectiveness in the first place.
The companies leading this shift — including Anthropic, Stripe, Vercel, and Linear, per current reporting — have converged on a different skill set entirely: system design thinking, clear communication under ambiguity, practical engineering judgment, AI fluency (how well someone actually works with AI tools, which mirrors the second interview strategy described above), and cultural alignment. This reinforces rather than contradicts the live-interaction and AI-transparency approaches already covered — the shared thread across all of these changes is a move away from formats that reward memorized, narrowly-scoped preparation and toward formats that reveal how a candidate actually thinks and works, which happens to be exactly the thing that's hardest to fake with or without AI assistance.
Structured scorecards make the live-interview shift actually work
Moving to live, adaptive interviews (the first strategy above) only reduces cheating risk — it doesn't automatically make the resulting hiring decisions better or fairer, and companies that skip the structured-scoring piece often trade an AI-cheating problem for an interviewer-bias problem. The data on this is specific: unstructured interviews carry a predictive validity of only about 0.20 for actual job performance, while adding a structured scorecard with behaviorally anchored rating scales raises that to roughly 0.51-0.57 — a substantial jump that makes the interview format itself a meaningfully better predictor, independent of the AI-cheating question. Structured scorecards work by forcing every interviewer to evaluate the same predefined, job-relevant criteria for every candidate, which narrows the room for unconscious factors like affinity bias, halo effect, or first-impression bias to drive the final call — worth noting given that around 48% of HR managers openly admit bias affects their recruitment decisions in current industry surveys.
The practical build for a live, adaptive interview format: define 4-6 core competencies pulled directly from the actual job description (not a generic template), attach behaviorally anchored descriptions of what a 1-5 score looks like for each competency, and require every interviewer to score against that same rubric regardless of how the live conversation happens to unfold. Organizations that make this shift — from unstructured judgment calls to structured, scorecard-anchored evaluation — report bad-hire rate reductions in the 25-35% range, which is a meaningful complement to the cheating-resistance benefit: the live format stops candidates from faking competence with AI assistance, and the scorecard stops the interviewer's own inconsistency from undermining the signal that live format was designed to capture.
The practical takeaway
If your hiring process still leans heavily on take-home tests as the primary signal, it's worth treating that as a known risk rather than a settled process. The core fix isn't more sophisticated cheating detection — it's redesigning the interaction so that unscripted, adaptive follow-up is the thing being evaluated, since that's the one part of the process an AI assistant genuinely can't do on the candidate's behalf without the candidate's own understanding showing through.
Sources: herohunt.ai, easontechtalk.com, metaintro.com, distantjob.com, jobsbyculture.com, pin.com, therecruitability.com
Keep reading
Get new posts as they publish
No spam — just the next post, straight to your inbox.