Google lets candidates use Gemini in engineering interviews

Google is experimenting with the technical interview format: in a pilot program, software engineering candidates can open Gemini during the interview itself. The news, relayed by the Chinese outlet Jiqizhixin, spread quickly through developer communities. The reason is clear: for years, preparing to get hired at a large company meant hundreds of LeetCode problems, and now the employer itself says an assistant in the room is acceptable.

What exactly is changing in the interview

The classic scheme used to be the foundation: the interviewer gives an algorithmic problem, and the candidate writes a solution from scratch, often on a whiteboard or in a simple editor. According to the description of the new pilot, that skill does not go away. But other skills are added to it.

Google wants to test how a person reads code they did not write, how they find errors in it and how they judge the quality of a solution. In the available fragment of the source material, the list breaks off at the word "evaluate," so the full list of criteria is not yet known. The logic is readable, though: if a model generates working code in seconds, an engineer's value increasingly depends on the ability to understand that code, catch a hidden defect and decide whether it is fit for production.

Key parameters of the pilot, based on available information:

• format: a pilot program, not a blanket replacement of existing interviews;
• tool: Gemini, Google's own model, available to the candidate during the session;
• retained: writing an algorithmic solution from scratch;
• added: reading someone else's code, locating problems, assessing the quality of solutions;
• there are no public figures yet on the number of participants, the timeline or the effect on selection.

Why this is a logical step

The gap between interviews and real work has long been a running joke. In everyday practice, a developer uses autocomplete, a chat with a model and agents in the IDE. Banning all of this during an interview tested a skill that people barely use in that form. There was also a practical risk: candidates can quietly connect an AI, and interviewers end up playing detective instead of assessing ability.

Legalizing the assistant removes that problem. If the tool is allowed and everyone knows it, the winner is the person who can frame a task, check the answer and notice where the model is wrong. Other companies have already taken similar steps: according to 2025 publications, Meta and Canva tested interviews with an AI assistant. In this sense, Google is joining the trend, and doing so with its own model.

There is a flip side. Assessing "code reading" is harder than checking whether a solution passes tests. Interviewers need new rubrics, calibration and uniform criteria, otherwise subjectivity will grow. The question of fairness also remains open: some people work well with an assistant and some do not, and that does not always reflect engineering level.

How candidates should prepare

Learning algorithms is still necessary: fundamental thinking is still being tested. But a few practices are worth adding to the preparation list.

1. Go through other people's pull requests and look for bugs in them without running the code. 2. Practice working with an assistant: setting the task, defining constraints, checking the output against edge cases. 3. Learn to explain out loud why you accepted or rejected a model's suggestion. The reasoning, it seems, is what will be evaluated. 4. Do not rely on the assistant as a source of truth: model errors in subtle spots remain commonplace.

For hiring teams, the takeaway is different: if you plan to allow AI in interviews, rewrite your evaluation criteria first. Otherwise you will get the same interviews, only with prettier answers.