The short answer is sometimes, and the only reliable way to know is to ask your recruiter which tool is allowed for which round. In 2026 that one question separates a candidate who looks prepared from one who gets flagged for the exact same behavior somewhere else.
The current picture looks like this. Meta started running an AI-assisted coding round in October 2025 for mid-level and senior software engineers, the E4 and E5 bands, and it replaces one of the old algorithm rounds rather than adding to them. Canva moved earlier, swapping its Computer Science Fundamentals interview for an AI-Assisted Coding round back in June 2025 across backend, frontend, and machine-learning roles. Google has been piloting an approved-assistant coding round for junior and mid-level candidates on select US teams. And a set of companies, Stripe and Shopify among the more vocal, tell candidates to bring whatever they would actually use on the job.
So “can I use AI” isn’t a yes-or-no question. It’s per company, per team, and per round, and the policies shift fast enough that a guide written last spring is already partly wrong. What follows is how to figure out where your specific interview lands, and how the scoring changes once AI is on the table.
The three stances you’ll actually run into
Almost every interview sorts into one of three buckets. Knowing which one you’re in is the difference between good prep and wasted prep.
| Stance | Where you see it | What you’re graded on |
|---|---|---|
| AI expected | Meta’s AI-assisted round, Canva, teams that hand you Cursor or Copilot | How you direct the tool, catch its mistakes, and defend the design, not whether you can write a two-pointer from memory |
| AI quietly allowed | Many startups and product teams that say “use your normal setup” | The same bar as always, plus you own every line the model wrote and have to explain it |
| AI off | The default live editor for most large-company algorithm rounds | Your own reasoning and code, with completions disabled and a hidden assistant treated as cheating |
That third row is worth dwelling on. CoderPad, HackerRank, and CodeSignal all ship an “AI-off” mode that kills Copilot-style autocomplete, and a lot of teams still turn it on for the live round even when they have no formal ban written down anywhere. The absence of a policy does not mean the editor will let you paste from ChatGPT.
One number to carry around, loosely: in a 2026 read of the New York market, a little over a quarter of employers said they allow AI in technical interviews, and recruiters expected that to climb toward half within a couple of years. Don’t quote the figure as gospel. The direction is the point. The default is drifting from banned toward allowed, but it hasn’t flipped, and a large middle group has no written rule at all.
Ask this, close to verbatim
Before the coding round, send your recruiter a line like: “For the coding interview, am I allowed to use an AI assistant, and if so, which one and in what mode?” It reads as prepared, not sketchy. Recruiters field this constantly now and would rather answer it than deal with a disputed result later.
The answer tells you more than the yes or no. “Cursor, and we’ll ask you to walk through what it generates” means the round is really about judgment and code review. “No, the editor has completions off” means you’re on the classic bar and should prep accordingly. Silence, or a vague “use your best judgment”, is the dangerous case, and there you should assume off and confirm again on the day.
What changes when the tool is allowed
When AI is on, interviewers stop caring whether you memorized Kadane’s algorithm. They watch how you work with a tool that is frequently, and confidently, wrong. The questions shift with it.
Expect to be handed a plausible-looking generation and asked to find the bug in it. Expect “why did you accept that suggestion?” and “what would you change about what it wrote?” A common setup: ask the model for a rate limiter, take the token-bucket implementation it produces, and the interviewer asks what breaks when two requests hit it at the same instant. If you can’t see that the counter isn’t thread-safe, the tool has just handed you a problem instead of solving one.
The real failure mode in these rounds is not being slow. It’s accepting a wrong answer smoothly. A candidate who reads each generation out loud, says what they’re checking, and rejects a suggestion with a reason looks far stronger than one who tab-completes a full function in silence and moves on. Slow down and narrate the review. That narration is most of what’s being scored.
Practice auditing the model’s first pass
If your round allows AI, prep with the tool on, but spend the reps on correction rather than generation. Take a medium problem, let the model write the first pass, then hunt for what’s wrong with it. You’ll start to recognize the recurring mistakes: off-by-one errors at boundaries, an O(n^2) solution dressed up to look linear, ignored concurrency, and calls to library methods that don’t exist. Spotting those quickly is the skill the AI-assisted round actually tests.
Take-homes are a different world
Everything above is about live rounds. On an async take-home, assume the reviewer expects you used AI, because nearly everyone does, and grading has already adjusted. A polished solution no longer impresses on its own. What earns points is the reasoning in your README, the tests you wrote, the tradeoff you chose and why, and the edge case you handled that the model wouldn’t have flagged. Plenty of teams now pair the take-home with a short live “explain your code” follow-up specifically to check that you understand what you submitted. If you leaned on AI for the take-home and can’t defend the design in that call, the gap shows within a minute.
What happens if you sneak it in when it’s off
The temptation is real when nobody appears to be watching. The detection is better than it looks. Codility’s CodeLive anti-cheat, HireVue, and some Karat configurations track tab switches, browser focus, paste events, and response timing. The strongest signal isn’t any single sensor. It’s the gap between polished code appearing instantly and a candidate who then can’t explain the tradeoffs inside it. Interviewers are trained to open that gap with follow-up questions, and it opens fast.
The downside is asymmetric. Getting caught doesn’t just cost you the offer. Several companies hand out a one-to-two-year ban on reapplying, and at firms with multiple brands under one parent, that ban can cover all of them. You’d be risking years of access to a company to skip a problem you could have prepped for in an afternoon. It isn’t a good trade.
Prepping for a round you can’t classify yet
When you genuinely don’t know which stance applies and can’t get a straight answer, prepare for AI-off and layer the AI-on skills on top. The classic bar still holds under completions-off, so the old grind on patterns and data structures isn’t wasted. Then add a few sessions of reading and correcting model output, because if the round turns out to be AI-assisted, that’s the part you can’t fake in the moment. Some interviewers who run an off round will still ask, as a discussion, how you’d approach the same problem with AI on the job. Having a real answer there costs you nothing and signals that you’ve thought about the tool as a working engineer, not a shortcut.
The strongest candidates in 2026 have one thing in common: they walk in already knowing which game they’re playing, because they asked before the day. That clarity is one short email away, and skipping it is the most avoidable mistake in the whole process.
Practice the behavioral round:
