Google is piloting a new coding interview. Candidates get Gemini in the room, sanctioned, on the record. Business Insider surfaced the internal doc: junior and mid-level roles, select US teams, wider rollout if it works.

The round to watch is code comprehension. Candidates open an existing codebase and read it, debug it, optimize it, with Gemini available the whole time. Google calls the format "human-led, AI-assisted."

The rubric is the interesting part. Interviewers grade "AI fluency, including prompt engineering, output validation, and debugging skills." Prompt engineering and output validation are judgment calls. Google turned judgment into a scored dimension of a coding interview.

Meta moved first, in October 2025, with a pick-your-model version. Google requires Gemini. Either way the two biggest hiring pipelines in the industry stopped pretending candidates work without AI.

They had reasons. A LeetCode round stops measuring anything once the candidate has a model open in a second tab, and most do. Google could keep escalating detection or start grading the collaboration. It picked the second.

If you run hiring anywhere smaller than Google, this pilot is worth copying and hard to copy. You need a real codebase, an AI the candidate can use, a log of every prompt and paste, and a grader who can tell direction from dictation. That is a lot of interview infrastructure for a five-person team.

We built that. In a Crackd screen, candidates work a real codebase for 90 minutes with AI on. We log every prompt, every paste, every backtrack, and grade the judgment: what they asked, what they trusted, what they threw away. Google built this round for Google. Crackd is the same round for everyone else.