On December 10, 2025, Karat launched NextGen, a human-led, AI-enabled interview format that grades how well you use an AI assistant instead of whether you sneak one in. If you are still drilling LeetCode for a rubric that assumes an empty tab, you are prepping for a shrinking format while 71% of US tech postings now demand the opposite skill.
The gap between what interviews test and what jobs require has finally cracked open, and the resume most engineers are sending in is worse off than they think.
Why Karat blinked first
Karat's NextGen release is the first major interview vendor to concede that hiding AI in the loop is now the wrong optimization target. Karat runs over 600,000 interviews across customers like Atlassian, Duolingo, PayPal, and Citi, so when it changes the rubric, the market moves with it.
The 2025-2026 Karat AI Workforce Transformation report set up the pivot with three numbers that do not reconcile under the old format:
- 71% of engineering leaders say AI now makes technical skills harder to evaluate.
- 62% of candidates already use AI during interviews even when the format prohibits it.
- Almost two-thirds of companies still ban AI in interviews, and fewer than 30% have updated assessments or trained interviewers to spot AI-ready talent.
Karat co-founder Jeffrey Spector put the diagnosis plainly: "Most companies are still hiring based on a pre-LLM rubric." NextGen is the fix. It runs in a VS Code-based IDE against a production-grade codebase with a built-in AI assistant, and a human interview engineer scores your AI literacy alongside your code.
DocuSign CTO Sagnik Nandy endorsed it at launch: "AI is transforming engineering, but the real breakthroughs happen when human judgment and AI capabilities work together. What's been missing is a way to measure that combination reliably."
Amazon is the visible counter-example, cracking down on AI use in loops. Both camps cannot be right, and the postings data suggests Amazon is the one adjusting to Karat, not the other way around.
The 181% AI-fluency jump the resume market has not caught up with
Dice's 2026 data puts demand at 71% of US tech postings requiring some form of AI fluency, a 181% year-over-year jump, and its July 2026 Jobs Report pushes that number to 75%. Supply has not moved anywhere close.
Refolk's index of professional profiles shows the mismatch is wider than any hiring pundit is saying out loud.
| Segment | Count | Note |
|---|---|---|
| US Software, Senior, and Staff Engineers | 563,439 | Refolk's index baseline |
| Same group listing GenAI, LLM, or Prompt Engineering | 5,550 | Refolk's index |
| Same group listing GitHub Copilot | 466 | Refolk's index |
| Share signaling any AI-fluency skill | ~0.99% | Derived, 5,550 / 563,439 |
| Share signaling Copilot specifically | ~0.08% | Derived, 466 / 563,439 |
| Gap vs. 71% of postings requiring AI fluency | ~70x mismatch | Derived, demand over supply |
71% of US tech postings require AI fluency, but under 1% of US SWE profiles in Refolk's index list a GenAI, LLM, or Copilot skill.
The concentration is narrow too. The employers where AI-fluent SWEs cluster in the index are Google, LinkedIn, Figma, Airbyte, Medplum, PayPal, and BCG X. For Copilot specifically, Microsoft holds 10 of 25 in the sample, followed by Google and Intuit. If you do not work at one of those, your resume is probably invisible to the fluency filter Dice is describing, and you almost certainly have more evidence of AI use in your commit history than in your bullets.
Why this is a scarcity trade, not a saturation trade
Most resume advice about AI treats it like a keyword everyone is stuffing. The data says the opposite. Adding two credible AI-fluency signals with real evidence is currently one of the highest-ROI edits available, because you are competing against a pool where 99% of engineers have nothing to say about AI on paper. That will not last. It is a 2026 window, not a permanent one.
What NextGen actually scores, and why hiding AI fails faster
NextGen scores AI literacy explicitly: how well you understand the model's capabilities and limitations, and how effectively you use it in the flow of writing code. LeetCode prep does not touch any of that.
The mechanic worth internalizing, from AceRound's read of Karat's format: in classic Karat, injecting AI answers gets caught through session-behavior analysis and follow-up questions. In NextGen, the interview hinges on your explanation of what the AI produced. So faking it fails faster than failing honestly, because you cannot narrate a diff you did not think through.
The prep shift is concrete:
- Narrate while coding. Say out loud what you are about to ask the assistant, why, and what you expect back. This is the single most under-drilled skill in the transition.
- Reject suggestions on camera. Accepting every autocomplete is a red flag. Rejecting one with a one-sentence reason ("this pulls in an extra dependency for a five-line function") is a green one.
- Diff before you accept. Read the suggestion, name what it changed, then accept. Silent tab-tab-tab reads as passive.
- Know the model's failure modes. Hallucinated APIs, wrong version syntax, security anti-patterns. Naming one unprompted signals literacy.
- Ask the assistant a bad question on purpose. Then correct the prompt. Interviewers scoring AI literacy specifically want to see prompt iteration, not one-shot luck.
Under NextGen, weak candidates who inject AI answers fail faster than honest ones, because the rubric grades your explanation, not the code.
The 2025 State of AI-assisted Software Development reported 90% of respondents already use AI in their work, up 14.1% year over year. If you are in that 90% at your day job and none of it shows in your interview behavior, the gap reads as dishonesty even when it is just habit.
Rewriting the AI-fluency resume for 2026
The right resume rewrite is evidence-based, not title-based. Adding "AI" to your job title is visibly gamed and will not survive a five-minute recruiter scan; naming the tool, the workflow, and the outcome will.
A useful before-and-after pattern:
| Before (2023 phrasing) | After (2026 AI-fluency phrasing) |
|---|---|
| Built internal tooling in Python | Shipped internal CLI in Python, drafted with Cursor and Claude, ~40% of PRs assistant-authored, reviewed line by line |
| Led migration to Postgres | Led Postgres migration, used Copilot for schema stubs and test scaffolds, hand-wrote all migration scripts |
| Owned checkout service | Owned checkout service, added an LLM-based fraud triage step (GPT-4o, ~120ms p50 added, false-positive rate down) |
| Mentored two juniors | Mentored two juniors on AI-assisted workflows, wrote the team's Copilot usage guide (when to accept, when to reject) |
Three rules for the rewrite:
- Name the tool. Cursor, Copilot, Claude Code, Windsurf, Aider, GPT-4o, Gemini. Recruiters searching Boolean strings need the literal string.
- Show the workflow. "Drafted with X, reviewed line by line" is the phrase that separates users from taggers. Karat's rubric is asking exactly this in the interview; put it on the page first.
- Quantify the ratio. "~40% of PRs assistant-authored" beats "used AI daily." Ratios read as measurement, not marketing.
This is the exact work Refolk takes off you: paste your history and the posting, get your resume back with the AI-fluency evidence pulled forward from projects you actually shipped, phrased for the fluency keywords the JD is scanning for. It also drafts the cover letter and scores how well you actually fit, so you can see whether the fluency gap is real for a given role before you spend the afternoon on it.
Senior roles are where the fluency signal pays
Per Dice via CIO.com, 71% of the increase in software development job postings between May 2025 and May 2026 was at senior level, and 37% of those mention AI in the title. AI/ML title postings are up 173% year over year with median salaries running 22% above IT overall.
The takeaway: senior candidates who prep NextGen well and rewrite for fluency can jump the queue in a hiring market that is otherwise brutal. Juniors get a compressed on-ramp, because the postings favor people who can already narrate AI-assisted work.
The five-step prep plan for a NextGen-style loop
Assume your next loop is NextGen or a copycat, because Karat's largest customers will pilot the format and the rest will follow. Prep like this:
- Install the assistant your target uses. Copilot for Microsoft, Cursor for most startups, Claude Code for AI-native shops. Interview in the tool you practice in.
- Run three sessions on a real repo you did not write. Not a toy problem. NextGen uses a production-grade codebase; your prep should too. Pick an open-source project, open an issue, and fix it end to end with the assistant on.
- Record yourself. Watch it back once. If you cannot explain a chunk of accepted code, that is exactly where NextGen will catch you.
- Practice rejecting. Deliberately ask for a bad implementation, then reject it out loud with a reason. This is the single behavior that separates "AI literate" from "AI dependent" on the rubric.
- Drill the explanation loop. For every ten lines the model produces, be ready to say what it does, why you asked for it that way, and one thing you would change. This is the scoring surface.
Two of these (session recording, explanation drills) are things no LeetCode platform has because they are not code problems. They are communication problems dressed as code problems.
Where the interview and the resume have to match
Your resume and your interview behavior now need to tell the same story about AI, or the mismatch itself becomes the red flag. A resume that reads "shipped with Copilot on 40% of PRs" paired with a candidate who never touches the assistant on screen is worse than either signal alone.
The synchronization matters most in three places:
- Bullets that name a tool. If Cursor is on the page, use Cursor in the loop. If you list Claude, do not open Copilot.
- Ratios on the page. If you claim assistant-authored PRs, be ready to talk about a specific one: what you asked, what it returned, what you changed.
- Workflow language. "Reviewed line by line," "prompt iteration," "rejected suggestion because" should appear in both your resume and your interview narration. Consistency is a scored variable now, even if no rubric names it.
Refolk tailors each resume to the specific posting so the AI-fluency language on the page matches the stack the JD names, which keeps you from walking into a Copilot shop with a Cursor resume. Doing that by hand for every application is what kills people at 20 or 30 apps in; letting Refolk score fit first means you skip the roles where your fluency story does not line up and go deep on the ones where it does.
From Karat's 2025-2026 AI Workforce Transformation report. NextGen is the format catching up to behavior that already happened.
FAQ
What is the Karat NextGen interview format?
NextGen is Karat's human-led, AI-enabled interview, launched December 10, 2025. You work in a VS Code-based IDE against a production-grade codebase with a built-in AI assistant, while a human interview engineer scores both your code and your AI literacy: how you prompt, what you accept, what you reject, and how you explain the output. It is Karat's response to the 62% of candidates already using AI covertly and the 71% of engineering leaders who say AI has made classical technical assessment unreliable.
How do I prep for a pair-programming-with-AI interview?
Practice narrating your work in the assistant your target company uses (Copilot, Cursor, or Claude Code), on a real codebase you did not write. Drill three behaviors: explaining what you expect before you ask, rejecting a bad suggestion with a one-sentence reason, and reading the diff before accepting. Standard LeetCode prep will not cover any of these, because they are communication skills scored against the AI output, not algorithm speed.
What should an AI-fluency resume actually say in 2026?
Name the tool, show the workflow, and quantify the ratio. "Shipped internal CLI in Python, drafted with Cursor and Claude, roughly 40% of PRs assistant-authored, reviewed line by line" beats "used AI daily" because it reads as measurement, not marketing. In Refolk's index, under 1% of US software engineers list any GenAI, LLM, or Copilot skill, so two credible fluency signals with real evidence make you visibly scarce right now.
Is it safe to admit I used AI on past projects?
Yes, and increasingly it is riskier not to. Karat, IEEE-USA, and Dice all point to "pair with AI" as the replacement for classic live coding, and 71 to 75% of US tech postings now require some form of AI fluency. Saying you shipped with an assistant, reviewed the output, and can explain the tradeoffs is the exact story the new rubric rewards; staying silent about it reads as either dated or evasive.