The AI-Assisted Coding Interview: What Senior Engineers Face in 2026

For roughly fifteen years, the unspoken premise of the technical interview was that the browser tab you were not allowed to open was the one with the answer in it. Then, in October 2025, Meta started opening that tab for candidates on purpose -- and grading them on what they did next.

The AI-assisted coding interview is not a proctoring story or a cheating story. Those stories have been covered to death. This is the opposite: an interview format where the company hands you a frontier model, tells you to use it, and scores how well you direct it. It is now live at Meta, at Canva, and in pilot at Google, and it changes what "preparing for a coding interview" means at the senior and staff level more than anything since the whiteboard went away.

Here is what the round actually looks like, what the rubrics measure, and how to prepare for a format where the hard part is no longer producing code.

The Round That Did Not Exist Two Years Ago

Meta rolled out its "AI-Enabled Coding Interview" in October 2025, initially as a pilot and, through 2026, as a standard part of the software engineering loop. Per Hello Interview's breakdown of the format, it replaces one of the two traditional coding rounds rather than being bolted on top of them. interviewing.io's guide, built from real candidate transcripts, describes a roughly 60-minute session in a CoderPad environment with an AI panel sitting beside the editor.

Canva got there first. In June 2025 the company published "Yes, you can use AI in our interviews", replacing its CS Fundamentals round with an AI-Assisted Coding round for backend, frontend and machine learning candidates. Copilot, Cursor and Claude are all permitted. By late 2025, Canva had moved from permitting AI to expecting it, and redesigned the problems themselves -- swapping Conway's Game of Life for a messier airport takeoff-and-landing control system that an LLM cannot one-shot from memory.

Google is the newest entrant and the most cautious. Per Exponent's tracking of the pilot, Google began testing a Gemini-enabled "code comprehension" round in May 2026 for junior and mid-level SWE candidates on select US teams, with a wider rollout signalled for the second half of 2026. Senior candidates are not yet in scope -- but the direction of travel is not subtle when the same company's CEO says that 75% of new code at Google is now AI-generated and approved by engineers, up from about 50% the previous year.

Who Allows AI, Who Bans It, and Why You Have to Ask

There is no industry standard, and the variance is not random -- it tracks how a company thinks about the job itself. The practical consequence for a candidate is that the single highest-value question you can ask a recruiter this year is which regime you are walking into.

Zoom out and the split is geographic as well as cultural. Karat's 2026 research with 400 engineering leaders across the US, India and China found that 71% say AI has made technical skills harder to assess, that 62% still prohibit AI in interviews, and that permissiveness varies enormously by market -- roughly 38% of US leaders allow AI in the interview versus about 68% in China. Most strikingly, leaders in that survey estimate that more than half of candidates use AI anyway, whether or not it is permitted.

So ask. In writing, before the loop: "Is AI permitted in the coding round? If so, which tools, and is it expected or merely allowed?" A recruiter who cannot answer will find out. The asymmetry is brutal -- guessing wrong toward "allowed" at Amazon ends your candidacy; guessing wrong toward "banned" at Meta means you hand-write a solution in a round explicitly designed to measure delegation.

Inside Meta's AI-Enabled Coding Interview

Meta's version is the most documented, and it is the one senior candidates are most likely to hit. The environment is a three-panel CoderPad session: problem statement, editor, and an AI chat panel. Candidates choose a model -- reported options have included GPT-5, Claude Sonnet 4.5, Gemini 2.5 Pro and Llama 4 Maverick -- and the interviewer watches the entire interaction, prompts included.

The problem is structured in phases rather than as a single puzzle: you are typically handed an existing implementation with defects to diagnose, asked to extend it with new functionality, and then pushed to optimize it under changing requirements. It is deliberately closer to a Monday morning at work than to a LeetCode set.

Meta's own framing, relayed through CoderPad, is blunt: candidates should use AI, but need to demonstrate that they understand the code that comes back. Explain the output. Test before you trust. Do not try to prompt your way out of a problem you have not understood.

What the Rubric Actually Scores

The rubric is where this format stops being a novelty and starts being a different exam. Meta's AI-enabled round evaluates four dimensions: problem solving, code quality, verification, and communication. Note what is missing -- there is no "produced a working solution unaided" line item, because unaided is no longer the measurement.

Google's pilot, per Exponent, converges on a similar shape from a different angle: it scores prompt engineering, validation of AI output, and the ability to debug what the model suggests. Canva's interviewers, writing in their inside guide to the round, describe the pattern that separates strong candidates from weak ones: the strong ones used AI strategically for well-defined subtasks while retaining control of the overall solution.

The industry-wide data says the same thing. In the CoderPad 2026 research, when AI is allowed in an assessment the top signal hiring teams look for is whether the candidate catches and fixes the AI's mistakes -- ranked above explaining trade-offs, improving output through iteration, and handling edge cases. The round is not measuring whether you can get to a green test suite. It is measuring whether you would have noticed if the tests were wrong.

Verification Is the Skill Being Tested

If that emphasis seems disproportionate, look at what the industry has learned about AI-generated code over the past two years.

The Stack Overflow 2025 Developer Survey, with more than 49,000 respondents, found 84% using or planning to use AI tools -- while positive sentiment fell and active distrust of accuracy rose to 46%, up from 31% the year before. The most-cited frustration, at 66%, was "AI solutions that are almost right, but not quite," and 45% reported that debugging AI-generated code takes longer than expected. Almost-right is precisely the failure mode a 60-minute interview is built to surface.

Nor is the productivity gain automatic. METR's randomized controlled trial of 16 experienced open-source developers across 246 real tasks found that developers were 19% slower when using early-2025 AI tools, while believing they had been 20% faster (full paper). The perception gap is the whole point: an interviewer watching you feel fast while going slow is watching the exact behaviour the rubric penalizes.

The code itself trends in a measurable direction too. GitClear's analysis of 211 million changed lines found duplicated code blocks rising sharply while refactored and moved code collapsed, a pattern their follow-up on the maintainability gap ties to elevated defect rates. On security, the canonical study remains Pearce et al.'s "Asleep at the Keyboard", which found roughly 40% of Copilot-generated programs in security-relevant scenarios contained vulnerabilities. And DORA's 2025 State of AI-assisted Software Development -- with around 90% of surveyed developers using AI at work -- frames the tool as an amplifier of whatever practices already exist, not a corrective for missing ones.

Put those together and the rubric's fourth dimension stops looking like a formality. Verification is the scarce skill. The interview is priced accordingly.

How to Run the 60 Minutes

A workable time budget for a 60-minute AI-assisted round, drawn from the phase structure Meta and Canva both use:

Minutes 0–5: clarify before you type. Constraints, scale, failure semantics, what "done" means. Do not open the AI panel yet. Interviewers consistently report that the first prompt tells them most of what they need to know about your seniority.

Minutes 5–10: state your approach out loud. Name the data structures, the invariants, the complexity target, the boundary between components. This is the artifact the AI cannot produce for you, and it is the one the interviewer is grading under "problem solving."

Minutes 10–35: delegate bounded subtasks. Ask for the parser, the retry wrapper, the boilerplate -- pieces whose correctness you can check by inspection. Keep the core logic and the architecture in your own hands. Read every line that comes back before it enters the file.

Minutes 35–50: verify aggressively and visibly. Write the test the AI did not write. Run the empty input, the duplicate key, the concurrent write, the off-by-one at the boundary. Say what you are testing and why. This is the highest-scoring stretch of the interview.

Minutes 50–60: optimize and narrate. When the interviewer changes the requirement -- and they will -- reason about the change before prompting about it. Explain what you would do differently with a week and a production dataset.

Prompting Like a Staff Engineer

The difference between a junior prompt and a staff prompt is not politeness or length. It is specification. A junior prompt outsources the thinking; a staff prompt has already done the thinking and outsources the typing.

The second prompt is scoreable. It shows the interviewer your data structure choice, your complexity target, your invariant, and your awareness of the tie-break edge case -- before a single line of generated code exists. If the model returns something wrong, you already know what "wrong" looks like, which is the entire skill.

Three habits reinforce this. First, prompt for a specific artifact, never for "the solution." Second, ask the model to justify a choice you already have an opinion about -- disagreeing with it out loud is a strong senior signal. Third, when you paste, narrate what you are checking as you read: "I want to see whether it handled the empty-runway case here." Silence while reading generated code is indistinguishable, from the other side of the screen, from not reading it.

The Failure Modes That Get You Down-Leveled

The reported rejection patterns are remarkably consistent across companies running this format:

• Prompting before understanding. Opening the AI panel in the first two minutes signals that you reach for the tool instead of the problem.
• Copy-paste without inspection. Code entering the editor faster than any human could have read it is the single most visible tell.
• Letting the AI own the architecture. Delegating the boilerplate is expected. Delegating the design is the thing being screened out.
• Vague prompts producing vague code. "Make it better" tells the interviewer you cannot articulate what better means.
• Failing to explain your own submission. The follow-up question -- "why does this line exist?" -- is unanswerable if you never read it.
• Skipping edge-case testing. The round's highest-weighted dimension, most commonly abandoned when the clock gets tight.

When AI Is Banned: Read the Regime Correctly

The permissive regime is not universal, and the penalty for misreading it is severe. Amazon instructs candidates not to use AI tools during interviews unless explicitly permitted, with disqualification as the stated consequence. Anthropic -- an AI company -- now invites candidates to use Claude on application materials but still bars it during live interviews. Goldman Sachs has taken a similar line.

Two things follow. First, "I did not know" is not a defence anyone accepts, so confirm the policy in writing and keep the reply. Second, in a banned regime, the compensating behaviours still matter: interviewers watching for AI use report cueing on delayed responses, typing during questions, and answers that sound read rather than reasoned. If you have practiced narrating your thinking continuously, you are simultaneously performing well and looking nothing like someone reading off a second screen.

It is also worth being realistic about where this is heading. When 62% of engineering leaders prohibit AI while believing most candidates use it anyway, the prohibition is not a stable equilibrium. Karat's own research arm has been rebuilding its assessment methodology around that reality, and the companies furthest ahead have concluded it is easier to grade AI use than to police it.

How to Practice for a Round You Cannot Memorize

You cannot grind this format the way you grind algorithm problems, because the thing being scored is a working style rather than a pattern library. Four drills map directly onto the rubric:

The broken-code drill. Ask a model to implement something moderately complex, then deliberately do not read the prompt output for correctness -- read it for defects, and time how long it takes you to find one. Do this daily for two weeks and you will develop the reflex the rubric is measuring. Our guide to the modern take-home assignment covers the same review muscle in written form.

The specification drill. Before every prompt you write this month, in real work, force yourself to state the function signature, the invariant, and the complexity target first. This is the single highest-leverage change, and it transfers straight into the interview.

The narration drill. Record yourself working with AI for 30 minutes with continuous commentary. Play it back. The gaps of silence are exactly where an interviewer loses the thread of your reasoning.

The tool-agnostic drill. Practice with at least two models. Candidates who have only ever used one tool are visibly slower when handed a different chat panel in a timed environment.

The broader loop has not gone away, either. The AI-assisted round sits alongside system design and behavioural interviews -- see our guides to the senior virtual onsite loop and what algorithm interviews still test at staff level -- and increasingly alongside questions designed to probe AI fluency without announcing themselves.