How to use AI in a coding interview without hurting your score

Published 30 October 2026 · 2 min read

In an AI-enabled round, the assistant is a tool you're being watched using. Strong candidates use it like a sharp colleague: small questions, checked answers, and their own reasoning in charge. Here's a playbook built from what companies and candidates have said about these rounds.

What's actually being scored

No company publishes a full rubric, but the public statements point the same way:

Google's pilot scores "prompt engineering, output validation and debugging skills" (via Business Insider).
Canva says it expects candidates to use AI tools and still own the result (Canva engineering blog).
Meta's candidate instructions reportedly tell you not to ask the AI for a complete solution (candidate report).

So the score comes from your judgment: how you break the problem down, what you ask, and whether you catch the AI's mistakes.

A loop that works

Orient first, without the AI. Spend two or three minutes reading: the entry point, the main data structures, and the failing test. Say out loud what you think the code does.
Form a hypothesis. "The off-by-one is probably in the stock check" is worth more than "AI, find the bug".
Ask a small, specific question. For example: "What does reserve() return when qty equals the stock?" Narrow questions get answers you can check.
Check the answer. Run the test, try an edge case, or trace a tiny input by hand. Don't move on until it holds.
Type the change yourself and explain it. One sentence on why it's right and what you'd test next.

Questions worth asking the assistant

"Summarise what this file does and which functions call it."
"What edge cases does this function not handle?"
"What's the time complexity of this loop, and which input makes it slow?"
"Write three test cases for this behaviour." Then read them, because generated tests can be wrong too.
"Explain this error message in the context of this code."

Mistakes that cost candidates

Pasting the whole problem in and hoping. It's slow, it reads as having no plan, and it's what the instructions warn against.
Accepting code you haven't read. If it fails later, you can't explain it, and that's the signal the round is designed to catch.
Going quiet. Interviewers can't score judgment they can't hear. Narrate what you're checking and why.
Over-trusting explanations. Assistants state wrong things confidently. Verify claims about the code against the code.
Forgetting the clock. These rounds have several phases. Don't spend 40 minutes perfecting phase one.

How to practice

Use real, unfamiliar code. Pick a small open-source project, break something on purpose, then find it again with an assistant that can't edit files.
Time yourself by phase: bug fix, then a new feature, then a performance pass.
Review your own transcript. After a session, reread what you asked the AI. Were the questions specific? Did you check every answer?

For company specifics, read Meta's AI-enabled coding interview, Google's code comprehension interview and which companies allow AI.