If you have a Karat screen on the calendar in the next few weeks, the rules you rehearsed for probably no longer apply. On December 10, 2025, Karat launched NextGen, a format that puts an AI assistant inside the interview and bans the one on your second monitor. The candidates who pass now are the ones who can talk about what the AI produced, not the ones who typed fastest with it hidden.
What Karat NextGen actually is
Karat NextGen is a human-led, AI-enabled technical screen where you work on a multi-file project with an embedded AI assistant while a Karat Interview Engineer probes your reasoning, tradeoffs, and validation in real time. It launched December 10, 2025 out of Seattle and is already the standard format for clients including Atlassian, Duolingo, and PayPal.
The mechanics that matter:
- The AI assistant lives inside Karat's interview environment. Bringing your own (ChatGPT, Copilot in another window, a hidden phone) is a violation.
- The Interview Engineer, Karat's trained human evaluator, asks you why you queried what you queried, what you accepted from the AI's response, and what you changed.
- The session length hasn't changed from the classic 60-minute screen. The rubric has.
- Karat has run more than 600,000 interviews across its client base, so this rubric shift moves a lot of hiring decisions at once.
Karat co-founder Jeffrey Spector's line for the launch was blunt: "Most companies are still hiring based on a pre-LLM rubric." NextGen is the answer he built.
Why the flip happened now
Karat inverted the rules because the old ones stopped working: candidates were already using AI in banned interviews, and clients wanted to hire the engineers who could use it well. Karat's own 2026 Engineering Interview Trends report says 62% of candidates use AI during interviews even when the format doesn't allow it, which turns "no AI" into an unenforceable policy that quietly rewards better cheaters.
The demand side is louder. Karat's 2025-2026 AI Workforce Transformation Report surveyed 400 engineering leaders across the US, India, and China and found:
- Leaders estimate a 34% average productivity lift from AI on their teams.
- Nearly 70% of engineering leaders plan to strengthen AI capabilities through strategic hiring.
- Roughly two-thirds of companies still prohibit AI use in interviews.
- Less than 30% are updating assessments to identify AI-ready talent.
From Karat's 2026 Engineering Interview Trends report, the number that made the old rubric untenable.
The gap between "we want to hire AI-fluent engineers" and "our interview forbids AI" is the exact space NextGen fills. Sagnik Nandy, CTO at DocuSign, appeared in Karat's launch materials backing the human-plus-AI evaluation model. Expect the client list to grow well past Atlassian, Duolingo, and PayPal through 2026.
The scoring inversion, in plain English
In NextGen, a silent candidate who ships a correct answer scores worse than one who ships partial code with clear reasoning. That is not a rhetorical flourish. It is how the Interview Engineer's rubric weights the session.
The pattern Karat's own guidance describes:
- A candidate completes 2 of 3 parts cleanly, narrates the tradeoff for each prompt to the AI, spots a bug in the AI's suggestion, and correctly identifies the time complexity. Strong hire.
- A candidate rushes through all three parts with poor naming, no narration, and can't explain why they accepted the AI's second suggestion over the first. No hire, even with more code shipped.
The mechanism is simple. An embedded AI can produce a working function in seconds. What it cannot do is defend the choice out loud under a follow-up from someone who has run thousands of these sessions. That defense, the validation and the tradeoff, is what the rubric now scores.
Silence plus a correct answer is now worse than partial code plus articulate prompt rationale.
The corollary is that classic cheat tools are double-losing. Screen-injection extensions were built for the "hide it" era. In NextGen, if an injected answer lands on your screen and you did not prompt it yourself, you cannot defend it, and the Interview Engineer's follow-up is exactly the moment you fail. Even in the classic Karat format, the detection guide at karat.com/detect-ai-use-technical-interviews flags continuous screen switching, instant fully formed solutions, and inconsistent explanations as red flags. Both formats punish the same behavior now.
The talent-pool number nobody is publishing
The engineers who have publicly signaled AI-collaboration skill are a rounding error of the broader pool, which means the prep gap is enormous and the edge from closing it is outsized. In Refolk's index of professional profiles, only about 0.46% of current US software engineers list the skills the NextGen rubric is trying to detect.
Here is the shape of it:
| Segment | Count | Source |
|---|---|---|
| All US software / backend / full-stack engineers | 356,560 | Refolk's index, baseline |
| Same segment listing Prompt Engineering + LLM | 1,647 | Refolk's index |
| US Senior/Staff engineers listing Prompt Engineering | 1,012 | Refolk's index |
| Share of broad pool signaling AI-collaboration skill | ~0.46% | Derived (1,647 / 356,560) |
| Candidates using AI in banned interviews | 62% | Karat 2026 Trends report |
| Companies still prohibiting AI in interviews | ~66% | Karat 2025-2026 Workforce report |
Two things jump out. First, 62% of candidates use AI in interviews, but fewer than 1 in 200 have bothered to say so on their profile. The market cannot yet tell the difference between someone who prompts thoughtfully and someone who alt-tabs to ChatGPT. NextGen is Karat's attempt to make that difference legible.
Second, the 1,012 Senior and Staff engineers who do list Prompt Engineering cluster at Google, Meta, LinkedIn, Airbyte, BCG X, and Spring Health, mostly in the SF Bay Area and NYC. If you are applying into a NextGen-standard client like Atlassian or Duolingo from outside that cluster, putting the signal on your resume is close to free alpha. This is the exact work Refolk takes off you: paste the posting, get your own resume back rewritten to surface the AI-collaboration work you actually did, in the language the Karat client is scoring against.
1,647 of 356,560 profiles in Refolk's index list Prompt Engineering plus LLM.
How to prep for a Karat NextGen interview
Prep for narration and validation, not raw LeetCode speed. The classic Karat screen is still a 60-minute live-coding session with communication scoring, but the weight on your voice, your prompt rationale, and your ability to reject an AI suggestion out loud is now the ballgame.
A concrete prep loop that maps to the NextGen rubric:
- Solve two multi-file problems a day with an AI assistant open and your mic on. Record yourself. Play back the audio. Count the seconds of silence.
- Narrate three things for every AI response: what you asked, what you kept, what you changed. If you can't articulate the third, you did not actually evaluate the output.
- Practice rejecting the AI on purpose. Ask a question you already know the answer to. When the AI hands back a subtly wrong version, name the bug out loud and fix it. This is the exact signal the Interview Engineer is listening for.
- State complexity unprompted. After you get anything working, say the time and space complexity before the interviewer asks.
- Prep for both formats. You will not always know in advance whether your screen is NextGen or classic. Same brand, opposite rules. Two 60-minute drills a week, one of each, is the safe posture.
- Use Karat's own resources. Brilliant Black Minds, Karat's mock-interview program, offers free technical mock interviews and is a legitimate way to get a live human reading of your narration.
The Interview Engineer is not looking for you to sound like a Loom voiceover. It is looking for four things: a stated goal, a prompt rationale, an evaluation of the response, and a decision. If every AI interaction in your session has those four beats, you are inside the rubric.
What to put on your resume before your next Karat screen
Put the AI-collaboration signal on the resume itself, in the language of the work, not as a buzzword line under Skills. The candidates who benefit most from NextGen are the ones whose profile already tells the Interview Engineer's client what kind of collaborator they will be.
Concrete lines that map to the rubric:
- Replace "used ChatGPT" with a specific outcome: "prompted Copilot to draft the pagination layer, rejected the first cursor design for an off-by-one on empty pages, shipped v2 in 3 days."
- Name the tool. Cursor, Copilot, Claude Code, and internal LLM gateways all read as more specific than "AI."
- Include one line where you overrode the AI. This is the single strongest signal that you can defend an output.
- List Prompt Engineering and LLM as skills only if you can back them up in a 30-second story. The 0.46% of the pool who list them credibly are the reason Karat built NextGen in the first place.
Refolk pulls that language out of your own history for you: you paste the posting, Refolk rewrites your resume against it, drafts the cover letter, and gives you a fit score before you submit. For a Karat funnel where the rubric now rewards specific narration of AI work, that pre-submission fit score is the closest thing to a preview of the Interview Engineer's read.
FAQ
How do I know if my Karat interview is NextGen or the classic format?
You often won't, until you are in the session. Karat rolls out NextGen client by client, and companies on the same day can be running opposite formats. The safe assumption for any Karat screen booked after December 10, 2025 at Atlassian, Duolingo, or PayPal is NextGen. For other clients, prep for both: one drill with an embedded AI assistant and narration, one classic live-coding drill with no AI. If you get NextGen, do not bring your own AI. If you get classic, follow the rules of the invitation exactly, because Karat's detection guidance flags screen switching and instant fully formed solutions as red flags.
Will using the embedded AI aggressively hurt my score?
No, but using it without narration will. NextGen is designed to reward AI collaboration fluency, not AI avoidance. The Interview Engineer is scoring whether you can state a goal, explain your prompt, evaluate the response, and make a decision. Candidates who prompt frequently and defend every acceptance tend to outscore candidates who try to prove they didn't need the AI. Silence plus correctness is the losing pattern.
Does raw LeetCode practice still matter for a Karat NextGen interview?
Less than it used to, but not zero. You still need fluency to read the AI's output fast enough to evaluate it, and pattern recognition on data structures still helps you spot bugs the AI introduces. What has changed is the marginal hour of prep. If you have ten hours before your screen, spend two on algorithm patterns and eight on narrated, AI-assisted multi-file problems with your mic recording.
What is the fastest way to signal AI-collaboration skill on my resume?
Rewrite one existing bullet to include the tool you used, the AI suggestion you rejected, and the outcome. One credible line beats a "Prompt Engineering" skill tag with no backing story. Only 1,647 US engineers in Refolk's index have paired Prompt Engineering with LLM on their profile, and the ones who convert those signals into interviews are the ones with a specific artifact behind the claim.