Two of the four largest engineering employers have now decided that an interview banning AI measures something increasingly unlike the job. Meta started first, piloting an AI-enabled coding round from October 2025. Google followed: from the second half of 2026 it is piloting a code comprehension round in which candidates use Gemini on an existing codebase, and Brian Ong, its VP of recruiting, has confirmed the pilot on the record. This is a narrower change than the headlines suggest, and it does not mean Google interviews are now open-book. The algorithm rounds are untouched. But the new round tests a skill most candidates have never rehearsed, and two other parts of the loop are shifting at the same time. This guide covers what is confirmed, what is not, and how to prepare.
How solid is this, actually?
Worth stating plainly before anything else. The details come from an internal Google document reviewed by Business Insider, not from a published Google policy page. What lifts it above rumour is that Google's VP of recruiting, Brian Ong, went on the record confirming the pilot and its rationale. So the existence of the pilot is well established; the operational specifics below are reported rather than officially documented, and pilots change. Treat this as a well-sourced picture of a moving target, and confirm your own loop with your recruiter.
The Three Changes
A new AI-assisted code comprehension round
The headline change. You are given an existing multi-file codebase and asked to read it, find and fix bugs, implement a feature, and optimise performance, with Gemini available throughout. The internal framing is reported as "human-led, AI-assisted": you drive, the model helps.
Googleyness and Leadership becomes partly technical
The behavioral round now reportedly includes a technical design discussion about your own past engineering work. It is no longer purely a culture and collaboration round, which means preparing stories about teamwork alone is no longer sufficient for it.
An open-ended engineering round for junior candidates
For junior candidates, one traditional technical round is reportedly replaced by an open-ended engineering problem. Less "implement this function", more "here is a messy situation, what would you do".
What has not changed matters as much. The conventional data structures and algorithms rounds remain, still delivered in a plain shared editor with no autocomplete, no syntax highlighting, and no way to run your code. Anyone reading this as permission to stop practising algorithms has misread it. Our full Google process guide covers those rounds, the four scored attributes, the hiring committee, and team matching.
The Code Comprehension Round in Detail
The premise is a reversal of the traditional format. Instead of an empty editor and a self-contained problem, you get a codebase you have never seen and a set of tasks that require understanding it first. That single difference invalidates most of the preparation candidates arrive with.
| Traditional Google round | Code comprehension round | |
|---|---|---|
| Starting point | An empty editor | An existing multi-file codebase |
| Task | Solve one self-contained problem | Read, debug, extend, optimise |
| Tools | No autocomplete, no compiler | Gemini available throughout |
| Main skill | Algorithmic insight, narrated | Comprehension and direction of a model |
| Scored on | Correctness, reasoning, communication | AI fluency: prompting, validation, debugging |
| Preparation that works | Timed problem drilling | Working in repositories you did not write |
| Status | Unchanged, still in the loop | Pilot, junior and mid-level US teams |
What "AI Fluency" Means in Practice
Reporting on the pilot names three competencies: prompt engineering, output validation, and debugging. Those are worth taking literally rather than treating as a vague gesture at using AI well, because they map to three distinct things an interviewer can observe.
Prompt engineering: can you direct the model usefully?
Not incantations. The observable skill is giving the model enough context to be useful: which file, which constraint, what you already ruled out. A candidate who pastes an error with no context and a candidate who says "this function is called from two places with different assumptions about ordering, here is the failing case" get very different answers, and the interviewer sees why.
Output validation: do you check what comes back?
The single clearest signal in the round. Reading a suggestion, spotting that it mishandles an edge case or misunderstands the codebase, and saying so out loud is exactly the behaviour being measured. Accepting plausible-looking code without reading it is the failure mode the round exists to detect.
Debugging: can you fix it when it is wrong?
Models produce confident, subtly wrong code. The round hands you a codebase with real bugs in it, and some of what Gemini gives you will be part of the problem rather than the solution. Composure and method when something does not work is the third of the three.
Notice that none of the three is about typing speed or knowing Gemini's interface. They are the ordinary judgment of a working engineer who happens to have a fast, unreliable colleague. That is the whole design of the round.
Google vs Meta: The Same Idea, Different Implementations
| Meta | ||
|---|---|---|
| Started | Second half of 2026 | Piloted from October 2025 |
| Who | Junior and mid-level, selected US teams | E6 and below, plus E7+ and M1 |
| Model | Gemini only | Reported choice of several providers |
| Round | Added code comprehension round | Replaces one of two coding rounds |
| Duration | Not reported | Reported at 60 minutes |
| Environment | Not reported | CoderPad with built-in assistant |
| Scored on | Prompting, validation, debugging | Judgment, review, communication, ownership |
| Other loop changes | Behavioral round now part technical | Rest of loop unchanged |
The implementations differ but the underlying bet is identical: that reading and directing code is now a larger part of the job than producing it from scratch, and that an interview should measure the thing the job actually requires. Our guide to Meta's AI-enabled coding interview covers that format in the same detail. If you are running both loops, prepare once for the shared skill and confirm the specifics separately.
How to Prepare
Keep drilling algorithms
This is first on the list deliberately. The conventional rounds are unchanged and still the larger share of the loop. Nothing about the pilot reduces the value of being fast and clean in a plain editor with no compiler.
Work in codebases you did not write, on a clock
Clone a mid-sized open-source project, pick a real open issue, and give yourself an hour to land a fix with an assistant open. The first ten minutes will feel bad. That feeling is the round.
Read before you prompt
Impose a rule: no prompt until you can say what the project does, where the relevant code lives, and what your change needs to touch. Prompting into a codebase you have not read produces confident nonsense, and the interviewer will see it happen.
Practise catching wrong answers
Ask a model for code you already know how to write and find the flaw in it. Validation is one of the three named competencies and it is the one that improves fastest with deliberate practice.
Narrate your prompting
Say why you are asking, what you expect, and whether the answer earned your trust. Silent prompting scores like silent coding always did at Google: the interviewer cannot credit reasoning they cannot hear.
Prepare a past project to defend technically
The Googleyness and Leadership round now reportedly includes a design discussion about your own work. Pick one project and be ready to justify its architecture, the alternatives you rejected, and what you would change.
Ask your recruiter which loop you are in
The pilot is narrow. If you are senior, outside the US, or on a team not included, none of this applies to you yet. Thirty seconds of asking beats a fortnight of preparing for the wrong round.
The Direction of Travel
Two companies is a pattern rather than a coincidence, and the reasoning both give is the same: interviews should look like the work. Google's own stated position is that a large majority of its new code is now AI-generated, which makes an AI-free interview an increasingly strange instrument for predicting performance. Expect more companies to follow, expect the emphasis to keep shifting from recall toward judgment, and expect the durable skill to be the unglamorous one — telling correct output from plausible output, and owning whatever you ship.
It is also worth being precise about what this is not. Neither company has made interviews open-book, and using an unapproved tool in a round that does not permit it remains a straightforward integrity problem. Our piece on where AI use is and is not permitted covers that boundary, which now varies not just by company but by round within the same loop.
Rehearse working out loud, under a clock
Amigo runs unlimited timed practice from your resume and target level, so you can train narrating your reasoning before the real loop.
Try Amigo free →Frequently Asked Questions
Is Google letting candidates use AI in interviews?
In one specific new round, yes. Google is piloting an AI-assisted code comprehension round in which candidates use Gemini while working through an existing codebase. Brian Ong, Google's VP of recruiting, confirmed the pilot on the record, describing it as making interviews 'more reflective of how our teams are operating in the AI era'. The conventional algorithm rounds are unchanged and remain AI-free.
Who is affected by Google's AI-assisted interview pilot?
Reporting indicates junior and mid-level software engineering roles on selected US teams, beginning in the second half of 2026, with expansion planned if the pilot succeeds. If you are interviewing at senior level, outside the US, or for a non-engineering role, it very likely does not apply to you yet. Confirm with your recruiter rather than assuming.
Which AI tool do candidates use in Google's interview?
Gemini, Google's own model. Unlike Meta's AI-enabled round, which reportedly offers a choice of several models from different providers, Google's pilot standardises on its own assistant. You do not bring your own tool.
What is the code comprehension round?
A round built around an existing multi-file codebase rather than a blank editor. Reported tasks include reading and understanding unfamiliar code, finding and fixing bugs, implementing a feature, and optimising performance, with Gemini available throughout. It tests whether you can work productively in code you did not write.
What does Google mean by AI fluency?
Reporting on the internal pilot document cites three named competencies: prompt engineering, output validation, and debugging. In plain terms, whether you can direct the model usefully, whether you check what it gives you instead of pasting it, and whether you can fix it when it is wrong. Validation is the one candidates most often skip.
Are Google's traditional coding interviews going away?
No. The classic data structures and algorithms rounds remain, still delivered in a plain shared editor with no autocomplete, no syntax highlighting, and no way to run your code. The AI-assisted round is an addition to the loop for candidates in the pilot, not a replacement for algorithmic screening.
What else is changing in Google's interview loop?
Two further changes are reported. The Googleyness and Leadership round now includes a technical design discussion about a candidate's own past engineering work, so it is no longer purely behavioral. And for junior candidates, one traditional technical round is replaced with an open-ended engineering problem.
How is this different from Meta's AI-enabled coding interview?
Both hand you an existing codebase with an assistant available, but the details differ. Meta's round is a reported 60 minutes in CoderPad with a choice of several models, and it replaces one of the two coding rounds. Google's is a Gemini-only code comprehension round added for junior and mid-level candidates on selected US teams. Meta began piloting in October 2025; Google's rolls out from the second half of 2026.
Found this useful?
Share it with someone preparing for an interview.
