AI-assisted coding interviews: what Meta, Google and Canva test
Updated
An AI-assisted coding interview lets you use an AI assistant while you read, debug and extend real code, and it grades how you direct the assistant and check its output, not only whether the code works. Meta runs a 60-minute AI-enabled round on a multi-file codebase, Google is piloting a Gemini-assisted code comprehension round, and Canva asks candidates to bring their own AI tools. Most companies still ban AI: in Karat's January 2026 survey of 400 engineering leaders, 62% of organizations prohibit it in technical interviews.
Formats change quickly and vary by role, level and team. The details below come from company statements and published reporting as of October 2026. Confirm your own loop with your recruiter before you choose how to practice.
Which companies allow AI in coding interviews?
A growing minority. Karat found that 38% of US companies allow AI in live interviews, against 68% in China. The companies below have described their formats publicly or had them reported in detail.
| Company | Round | AI assistant | Status |
|---|---|---|---|
| Meta | 60 minutes in CoderPad on a multi-file codebase: fix or review, extend, then handle edge cases | Built in, with selectable models | Piloted from October 2025; rolling out to software engineering roles in 2026 |
| Code comprehension: read, debug and optimize existing code | Gemini | Pilot for junior and mid-level roles on select US teams from the second half of 2026 | |
| Canva | Realistic, ambiguous product challenges that one prompt cannot solve | Your choice, such as Copilot, Cursor or Claude | Piloted in 2025 as a replacement for the computer science fundamentals screen |
| Most others | Conventional coding rounds | Not allowed | 62% of organizations prohibit AI (Karat, 2026) |
What happens in Meta's AI-enabled coding round?
According to interviewing.io's guide, you get 60 minutes in a CoderPad environment with a directory tree, a terminal, a button to run unit tests and an AI assistant. The codebase is typically a few hundred to a few thousand lines across several files. One themed problem rises in difficulty: a bug fix or code review, then a new feature, then edge cases and scaling, with questions about runtime, tradeoffs and design choices along the way.
Candidates at E6 and below reportedly take it alongside a conventional coding round, while candidates at E7 and above, and M1 managers, take only the AI-enabled round. The common failure modes are accepting suggestions without reading them and following a confident but wrong answer. Read each proposed diff before you accept, reject or change it.
Meta gives candidates with a scheduled interview a practice CoderPad session with the same AI panel; Hello Interview's guide advises asking your recruiter for the link if it did not arrive. You are not expected to finish every part, and orientation in an unfamiliar multi-file project eats into the hour, so practice reading code you did not write.
What is Google's code comprehension round?
Business Insider reported in May 2026, from an internal document, that Google is piloting a round where candidates read, debug and optimize existing code with Gemini available as an assistant. A Google spokesperson confirmed the assistant is Gemini. Interviewers assess "AI fluency, including prompt engineering, output validation and debugging skills." The pilot covers junior and mid-level roles on select US teams, with wider rollout if it works.
The same report says the Googleyness and Leadership round gains a technical design conversation about your past engineering work. If you are interviewing at Google, ask which rounds in your loop are part of the pilot.
What do interviewers evaluate when AI is allowed?
The published criteria converge on judgment rather than typing. Canva lists knowing when and how to use AI, breaking down unclear requirements, and identifying and fixing issues in AI-generated code. Google names prompt engineering, output validation and debugging. Meta's format is built to expose candidates who rely on the assistant without understanding the result.
- Scoping: understanding the codebase and the requirement before asking for changes.
- Directing: giving the assistant bounded, specific requests instead of pasting the whole problem.
- Verifying: running tests, reading diffs and catching wrong or incomplete output.
- Owning the result: explaining the code you submit, its tradeoffs and what you did not check.
How should you practice for an AI-assisted round?
Practice the format, not just the algorithms. Work in an unfamiliar multi-file codebase under a 60-minute timer, with an assistant available. Fix a bug, then extend the behavior, and narrate your reasoning out loud or in notes. Afterward, review your prompts and test runs and pick one habit to change.
PraxisAI's timed bug-fix challenges come from real open-source repositories, which is close to the read, debug and extend work these rounds describe, and its take-home projects suit open-ended builds like Canva's. PraxisAI is not affiliated with these companies, and its environment differs from CoderPad and Google's tooling, so use it to build the habits rather than to rehearse an exact replica.