A CNN headline this week put it bluntly: AI is changing software engineering so fast that the interview process cannot keep up. The piece described something that was a rumor a year ago and is now policy at the biggest engineering orgs in the world — the AI is in the interview room with you, and the interviewer is watching how you use it.
Google is rolling out a "code comprehension" round in its 2026 loop where you read, debug, and optimize a real codebase with Gemini available at your side. Meta started replacing one of its two coding rounds with an AI-enabled session back in October 2025 — you pick your model (GPT, Claude, Gemini, Llama), and you are scored on four axes: problem solving, code quality, verification, and communication.
Read that list of four again. Notice what is missing. "Can you produce the correct algorithm" is not on it.
The test inverted, and most candidates have not noticed
For a decade, the coding interview was a memory-and-recall exam wearing a problem-solving costume. You ground LeetCode until the patterns were reflexes, walked in, and reproduced the optimal solution fast enough to look smart. The AI assistant breaks that game completely. The model reproduces the optimal solution faster than you ever could. So reproducing it stopped being the signal.
When given an AI in the interview, weak candidates do one of two things: they ignore it and grind the problem by hand to "prove" they don't need it, or they paste the prompt, accept the first output, and move on. Both fail. The first signals you cannot work the way the team works. The second signals you cannot be trusted to catch what the model gets wrong.
What replaced raw recall is harder to fake and harder to cram: judgment under collaboration. Can you frame the problem so the model gives you something useful? Can you read its output critically instead of trusting it? Can you catch the failure mode it skipped? Can you narrate your reasoning so the interviewer knows the judgment is yours and not the model's?
This is the same shift the broader market is signaling. The Pragmatic Engineer's 2026 job-market analysis describes fewer but higher-skilled roles and AI skills appearing in 42% of job descriptions, up from 8% in 2022. Hiring data from this spring shows aptitude assessments surging — companies are filtering for thinking, because syntax is now a commodity the AI supplies.
The interview stopped asking "do you know the answer." It started asking "can I trust your judgment when the answer is one prompt away."
What the four scoring axes actually mean
Meta's rubric is the clearest public window into the new bar. Here is what each axis is really probing.
| Scoring axis | What it actually tests |
|---|---|
| Problem solving | Do you decompose the problem before prompting, or do you outsource the thinking to the model? |
| Code quality | Can you tell good output from plausible output, and improve what the model hands you? |
| Verification | Do you test, probe edge cases, and distrust the happy path the AI defaults to? |
| Communication | Can the interviewer follow your reasoning and see that the judgment is yours? |
Three of those four are not about code at all. They are about whether you can think out loud while directing a tool. That is a spoken skill, and it is almost impossible to develop by practicing silently against a screen.
How to prepare for an interview where the AI is allowed
The instinct is to get better at prompting. That is necessary but not sufficient. The differentiator is everything that surrounds the prompt — the framing before it and the verification after it.
The uncomfortable part: you cannot rehearse this against a static problem set. The AI-assisted interview is a live conversation where the interviewer introduces new constraints mid-session and watches how you adapt. The only practice that transfers is practice that talks back.
You are not the coder in the room anymore. You are the senior engineer reviewing a fast, confident, occasionally wrong junior. The interview is testing whether you would be safe to put in that seat. Act like it from the first sentence.
The companies did not lower the bar by letting you use AI. They raised it. The floor used to be "can you produce working code." The floor now is "can you produce working code and prove you knew why it was right." The candidates clearing that bar are the ones who have practiced explaining, steering, and verifying out loud until it is reflex.
Rubduck drops you into a live, spoken AI-assisted coding interview — you steer an AI in the editor while an interviewer probes your judgment, exactly like Google's and Meta's new rounds. You get scored on the same axes that matter: problem solving, verification, and how clearly you communicate. Start your free sessions →