When programmers interviewed, they were finally able to turn on the bright AI.
Recently, a software engineer interviewed by Google for a reform that sparked a heated debate online.
According to the new pilot schedule, candidates will be able to use Gemini in interviews, and the contents of the exam will begin to change.
In addition to writing an algorithm problem from scratch, Google wants candidates to be able to read other people's code, locate problems, and judge the reliability of AI solutions.
Some netizens excitedly said: "AI has finally killed LeetCode interviews".
In addition to writing code, Google will also test how you read code.
For many programmers preparing to enter a large factory, LeetCode should be more familiar than a name.
Algorithmic puzzles, time-consuming research complexity, and writing the right code in a limited amount of time have almost become a must-have for technical interviews.
Google is experimenting with a different way of thinking.
According to a Business Insider report on May 7, candidates will be able to read, modify, and optimize an existing code in Google's pilot new link, Code Comprehension.
AI can also be involved in the process. A Google spokesperson confirmed that the tool provided to candidates in the pilot is Gemini.
Since AI can be asked directly, how should an interviewer judge the level of a candidate?
Google's answer is "AI proficiency" (AI fluency).
The interviewer will focus on how candidates design prompts, verify the output of AI, and debug code. Candidates need to be able to judge whether AI's suggestions are valid and responsible for the final results.
This means that even if Gemini quickly finds a bug, the candidate will need to verify that the fix is working and that the modification will introduce other issues.
Of course, this is only a limited reform at the moment.
The relevant arrangements are aimed at the second half of this year and mainly involve the positions of primary and secondary software engineers in some teams in the United States. Organizations such as Google Cloud, the platform and the equipment sector have participated in the pilot arrangements, and whether future deployments will depend on the actual results.
Brian Ong, Google's vice president of recruitment, said the adjustment to the interview process was to more accurately reflect how engineers actually work in the AI era.
In addition to understanding the code link, Google plans to include technical design discussions of candidates' past projects in some interviews, allowing junior candidates to address more open engineering issues.
In addition to being able to write code from scratch, companies have begun to examine whether candidates can process an existing code or even code generated by AI.
Karpathy: Handwriting skills are starting to deteriorate.
But reading code is another thing.
It's easy to see what's going on with Google's own engineering team.
In April, Google CEO Sundar Pichai revealed that 75% of the company's new code has been generated by AI and reviewed by engineers.
In just a year and a half or so, AI's involvement in software development has increased significantly.
For individual developers, similar changes occur faster.
In January this year, Andrej Karpathy shared his experience with Claude Code on X.
He found that his programming habits changed almost inverted in just a few weeks.
As of November 2025, about 80% of his programming work is still done manually, or with the help of code autocompletion; by December, about 80% of the work has been handed over to Agent, who is solely responsible for the remaining modifications and improvements.
Karpathy was shocked that this was one of the biggest changes to the way he worked in his nearly 20-year programming career.
Today, much of his programming work has turned to using natural language to describe the need for AI.
This change even had an unexpected impact: Karpathi found that his ability to write code manually began to gradually deteriorate.
However, he specifically distinguished between two abilities:
"Creating (writing) code and distinguishing (reading) code is a distinct ability in the brain".
Writing code often requires a programmer to remember many specific syntax and implementation details, while reviewing code relies more on understanding the logic and judging whether implementation is reasonable. Thus, a person may still be able to read the code even if he or she is not good at writing code from scratch.
But AI's code is also becoming increasingly difficult to check.
Karpathy mentioned in the post that the mistakes made by programming models today are no longer mainly simple grammar issues.
Instead, they may make the wrong assumptions when understanding the need, and then continue to complete the entire set of implementations according to this assumption. Sometimes the model writes thousands of lines of complex code, and after engineers point out a more concise implementation method, it can quickly reduce the code to about a hundred lines.
These mistakes look more like omissions made by inexperienced developers in a hurry.
So while Karpathy has found it difficult to imagine returning to a development that relies primarily on handwritten code, he still believes there is a need to pay close attention to what AI writes.
This is a new way of thinking about Google interviews.
If AI can do a lot of coding in a matter of minutes, engineers' tasks will increasingly shift to understanding code, verifying implementation, and intervening when AI makes mistakes.
As code generation becomes easier, the ability to review code becomes even more important.
Google wasn't the first, and some people even intentionally misrepresented AI.
In fact, Google is no longer one of the first companies to try this approach.
In June last year, design platform Canva announced in its official engineering blog that candidates for rear-end, machine learning and front-end engineers would be required to use AI programming tools in technical interviews.
For front-end and back-end positions, Canva even replaced the basic interview link with AI-assisted programming, which was previously based on algorithms and data structures.
The reason is straightforward: the old question is no longer as effective today.
Canva used AI to test its own traditional interview questions and found that the model often takes only a few seconds to generate the correct and explanatory answers, even without further input.
So Canva decided to make the topic more complicated.
For example, in the past, candidates may have been required to write "The Conway Game of Life", but now they may have been asked to design takeoff and landing control systems for busy airports.
This open-ended problem requires candidates to figure out their own needs, identify schemes, and coordinate the relationships between different functions. AI can complete the coding tasks, but it's difficult to handle all the design and engineering details with just one simple instruction.
Canva also shared observations from the pilot.
Successful candidates tend to be proactive in identifying needs, breaking down complex tasks into more precise parts, and then targeting AI. They examine the results generated, and after discovering problems, they adjust the scheme.
Some candidates who have programming skills but lack experience using AI may find it difficult to guide AI in deciding whether recommendations are appropriate.
Meta is starting to adopt a similar approach.
In July 2025, the media disclosed that Meta was developing coding interviews that allowed the use of AI. Subsequently, Meta officially introduced the pilot publicly and encouraged the candidates involved to use approved AI tools.
Meta Special emphasizes that using AI in such interviews is no longer cheating, as it is the tool engineers use in their actual work.
In October, the recruitment technology platform Codility released a new interview feature that allows interviewers to add hidden instructions to the AI programming tools used by candidates.
For example, allowing AI to intentionally write code that is bugged, or using unreasonable programming methods.
Candidates will not see these instructions. The interviewer will be able to observe how the candidates are handled, whether they will carefully examine the results generated, or directly copy the stickers, as received in full.
The platform also supports the use of AI controls in different scrutiny tasks, allowing enterprises to both evaluate the performance of candidates in collaboration with AI and check their ability to solve problems independently.
In the past, in order to prevent candidates from getting answers through AI, interviewees may have had to figure out how to identify traces of AI use.
In the past, a programming question may have started with "Please implement this algorithm".
In the future, the interviewer might give you an AI code directly and ask, "What do you think is wrong with this code, it's already working?"
And the AI tools at your fingertips won't help you answer that question.
As some netizens comment: "To ask candidates to use Gemini to complete tasks is a good interview in itself".
The link is as follows: