On October 20, 2026, internal Google documents reviewed by Business Insider showed the company is rolling out AI-assisted software-engineer interviews: candidates use Gemini during a "code comprehension" round, reading, debugging, and improving an existing codebase, while interviewers grade AI fluency, prompt quality, output validation, and debugging as a baseline expectation for software engineering roles.
What Business Insider’s documents actually show
Business Insider’s reporting, based on documents it reviewed, describes Google testing several AI changes to its hiring process for software engineers, rolled out in phases across 2026 (reporting via dnyuz, October 20, 2026). The changes include letting candidates use AI in coding interviews and letting interviewers use Gemini to draft role-tailored feedback. Many engineers are now expected to show some degree of AI fluency as a bare minimum, and a Google spokesperson told BI that AI fluency is a requirement for all software engineering jobs.
The "code comprehension" round is the centerpiece. Business Insider first reported the pilot in May 2026: candidates for junior and mid-level software engineer roles on select US teams were allowed to use Gemini as an assistant while reviewing, debugging, and improving code, with interviewers scoring how effectively they work with the tool. The October documents show the formats have been rolled out more broadly across the company, though a spokesperson said they are still considered pilot-phase.
Google’s VP of Recruiting, Brian Ong, gave Business Insider the official line: "We’re piloting updates to our interview process for software engineers, including evaluating for AI fluency, to be more reflective of how our teams are operating." Google is also piloting Gemini for interview transcription and note-taking on Google Meet, and one document warns employees against building their own AI interviewing tools: "Do not use GenAI or LLMs to create your own standalone solutions or agents for hiring or interviewing use cases."
The context for the shift: in April 2026, Google said roughly three-quarters of new code at the company is AI-generated and reviewed by humans. When the draft is written by a model, the human skill that matters is reading the draft, verifying it, and fixing it. The old interview never tested that directly. The new one does.
What "AI fluency" actually means, stripped of jargon
AI fluency is not a new language to memorize. It is the ability to do professional-grade work with an AI assistant at your side: knowing when to use it, how to direct it, how to check what it gives you, and how to fix what it breaks. The durable skill is not writing code fast. It is reading, debugging, and validating AI-written code, because the writing is now the cheap part.
Zapier, which made AI fluency a hiring requirement for every new hire in May 2025, grades candidates on a four-level rubric, from unacceptable to adoptive to transformative (Zapier’s own account), built function by function around how its top performers actually use AI. The minimum bar is not "uses a chatbot sometimes." It is demonstrable, describable impact on quality, efficiency, or output. That framing is the one spreading across the industry: Google’s version is "AI fluency" as a bare minimum for software engineers; Canva replaced its computer science fundamentals screen with an AI-Assisted Coding round entirely.
The 18-month shift, tracked
The timeline above is ByteHerald’s original compilation, and the pattern it shows is worth stating plainly. In April 2025, Shopify CEO Tobi Lutke’s leaked memo made effective AI use a baseline expectation for all employees, with teams required to prove AI could not do the job before requesting headcount (TechCrunch, April 7, 2025). The interview industry moved first: CodeSignal launched AI-assisted coding assessments on May 28, 2025, powered by its Cosmo assistant, so companies could evaluate how candidates collaborate with AI in the IDE.
Zapier made fluency a universal hiring criterion that same month. In June 2025, Canva went further and made AI mandatory in its coding interviews, replacing the computer science fundamentals round with an AI-Assisted Coding round using tools like Copilot, Cursor, and Claude (The Register, June 11, 2025). Reports of Meta piloting an AI-enabled round surfaced in July 2025; candidate reports compiled by prep firms place the live pilot, replacing one of two onsite coding rounds with a 60-minute CoderPad session in a multi-file codebase, from October 2025 (TechScreen, 2026).
On December 10, 2025, Karat launched NextGen interviews, a human-led, AI-enabled format for employers (Karat press release). Karat’s own 2025-2026 report adds a sobering counterpoint: almost two-thirds of companies still prohibit AI use in interviews, and fewer than 30% are updating assessments or training interviewers for AI-ready talent. The leaders moved; most of the market has not caught up.
In May 2026, Datadog updated its candidate AI guidelines: no AI during live coding interviews or technical assessments, except in interviews explicitly designed for AI-assisted coding, with candidates told in advance. And in May 2026, Google’s internal pilot put Gemini in the "code comprehension" round for junior and mid-level roles. The October 2026 documents show that pilot expanding across the company, with AI fluency now a stated requirement for all software engineering jobs.
What interviewers actually score
Across the formats we verified, the graded triad is consistent, and it is not "how fast the AI writes." Google’s pilot scores prompting, output validation, and debugging. Canva’s AI-Assisted Coding competency evaluates whether candidates can direct the AI, critically review its output, and identify bugs it introduced. Meta’s AI-enabled round counts verification among its four scored areas: whether you catch the model when it is wrong.
ByteHerald’s analysis: the documents publish no numeric weights for these areas, so we assign none. What the evidence does show is a uniform direction. The scarce skill, the one candidates are filtered on, is judgment under assistance: stating constraints before prompting, testing what comes back, and owning the fix when the model is wrong. Speed of generation is free. Supervision is the skill.
How to prepare: the durable playbook
Interview formats change by team and by quarter, so prepare for the underlying skill, not one company’s rubric.
- Practice reading unfamiliar codebases under a timer. Open a repo you have never seen, give yourself 60 minutes, and explain what it does, where the bug is, and how you would fix it. That is the code comprehension round.
- Use one AI assistant exclusively, the way you would a colleague. Our original scorecard of five AI coding assistants compares published benchmarks, October 2026 pricing, and task fit, which tells you which tools the strongest candidates will have on hand; the point is to build deep fluency in one of them, not shallow familiarity with five.
- Deliberately practice the failure case. Feed the assistant a vague prompt, watch it produce subtly wrong code, and narrate exactly how you found and fixed the fault. Canva’s pilot found candidates with minimal AI experience struggled not because they could not code, but because they lacked the judgment to guide AI or spot suboptimal suggestions.
- Narrate everything. In the AI-assisted format, your thinking is the answer. The model writes; you prove you understand every line.
- Ask your recruiter, in writing, which loop you are getting and which tools are allowed. Formats are pilots; they vary by team and change often. Never assume.
The judgment call for hiring managers
For hiring teams, the 18-month record answers the question Canva’s engineering leadership asked in June 2025: fighting AI in interviews had become a policing problem, and policing was losing. Detection is impractical, covert use undermines fairness, and the interview no longer mirrored the job. The companies that moved first, Canva most explicitly, redesigned questions to be messier and more realistic, problems that cannot be solved with a single prompt, and graded the supervision instead of the generation.
The open question is calibration. Google is still in pilot phase, Meta’s format rests on candidate reports rather than an official guide, and Karat’s data says most companies have not touched their rubrics at all. The bar moved at the top of the market first. Everyone else is hiring against a pre-LLM rubric for a human-plus-AI job.
By ByteHerald Staff. Core reporting from Business Insider’s review of Google’s internal documents (dnyuz, October 20, 2026). Timeline compiled by ByteHerald from company announcements and press reports dated April 2025 through October 2026; scoring details reflect reported pilot formats, not official Google documentation, and formats vary by team. Nothing in this article is legal or employment advice.
Comments
Post a Comment