The AI-assisted coding round gives you a small codebase and a terminal. You also get an AI assistant you’re allowed to use. It tests whether you can judge the assistant’s output, not whether you can type code from memory. Candidates fail it in two ways. Some accept suggestions they can’t explain. Others refuse to use the assistant and run out of time. Expectations also differ between companies. Sometimes they differ between interviewers. So ask at the start how you’re expected to use the AI. Then narrate every accept or reject decision out loud. Say why a suggestion is right or what you don’t trust about it. Solve the core logic yourself first. Use the assistant for the parts you already understand. Below is one short exchange with the assistant, told the weak way and the strong way. It’s followed by how the best-documented version of this round runs.
Meta’s version of the AI-assisted coding round is the clearest picture so far
Evan King of Hello Interview, an interview-prep company, has written up Meta’s AI-enabled coding interview from candidate reports. He writes that Meta began rolling it out in October 2025. He quotes an internal Meta message describing the goal. It’s meant to reflect how future employees will work. It should also make LLM-based cheating less effective. He has talked with candidates from E5 to E7 and M2.
He describes three phases built around one extended problem, starting with fixing a bug in the provided code. Two details matter for preparation. First, the rules on AI use during the bug-fixing phase vary by interviewer. Some ban it for that part. Others leave it open. Second, candidates report that the assistant seems tuned to be less helpful. It tends not to point out bugs or hand over complete solutions. King’s advice is to solve hard parts yourself. Then use the AI to speed up what you already understand. His company sells prep for this round, so weigh his framing with that in mind.
This round isn’t Android-specific. Meta’s version is simply the best documented. Expect a similar shape wherever an Android loop adopts it.
The example: accepting an AI fix you can’t explain
Say the codebase has a Kotlin function that pages through results. A test fails with an index error. You ask the assistant. It suggests changing a loop bound from size to size - 1. The test passes. Here’s the weak version of what happens next.
Okay, that fixed it. Moving on to the next part.
If the interviewer asks why the bound was wrong, the candidate has nothing to say. They accepted a change because it compiled and passed one test. The round is built to catch exactly that.
The same moment, narrated
The assistant suggests size minus one. Before I accept it, let me check why. The loop reads index i plus one inside, so at the last element it reads past the end. So the bound fix is right for this loop. But the real issue is that the loop reads two elements at a time. The last page gets skipped when the count is odd. I’ll accept the bound change and add a test for an odd count. Then I’ll check the last page separately.
Now the interviewer has heard your reasoning. You checked the suggestion against the code and accepted it for a stated reason. You also spotted a second bug it missed. That’s the judgment the round is scoring. The code details are illustrative.
Refusing the assistant is the other trap
Some engineers avoid the assistant to prove they can do it “the real way.” In a round built around AI use, that can read as unfamiliarity with how teams work now. Writing boilerplate by hand in a format built for tool use also eats time. That leaves less for the parts that are scored.
The balance is simple. Do the reasoning yourself. Let the assistant write what you could have written yourself. A data class, a test scaffold or a known pattern are good examples. Then read what it wrote before you run it.
How to prepare for the AI-assisted coding round
- Practice in an editor with an AI assistant, on code you didn’t write. Get comfortable with where suggestions appear and how to prompt mid-problem.
- For every suggestion you accept in practice, say out loud why it’s correct. For every one you reject, name what you didn’t trust.
- Practice the bug-fixing phase without AI, since some interviewers ban it there.
- At the start of the real round, ask how the interviewer wants you to use the assistant. It takes ten seconds and removes the biggest source of misjudged pacing.
The same narration habit helps in the Android pairing round. There the interviewer also scores how you think out loud.