Well-known Linux Kernel maintainer Greg Kroah-Hartman said in an interview that artificial intelligence-driven code review can achieve a real leap in Linux systems. At present, because artificial intelligence generates so much garbage, some open source projects even prohibit the submission of reports generated by artificial intelligence. However, how to use artificial intelligence technology still depends on the maintainer's decision.
At least in the Linux Kernel, kernel maintainers are not averse to using artificial intelligence to review bugs, and even maintainers like Croa-Hartmann praise artificial intelligence for its prominent role in code review.
Croa-Hartmann mentioned that in one test he asked the AI to find problems, and then the AI reviewed the code and found 60 problems and provided solutions. After human review, Croa-Hartmann found that about 20 of the fixes were wrong, but the remaining 40 fixes were real and effective.
It's important to note that even if these fixes were real, they would still require manual cleanup, change log improvements, and integration work, all of which Croa-Hartmann would have to do manually again.
In the Linux Kernel project, Croa-Hartmann obviously will not allow AI to directly merge modification plans. Croa-Hartmann's approach is that patches will have the same development label, and AI is mainly responsible for code review.
In the end, the code will be merged only after the team has thoroughly confirmed that it is valid. This can prevent omissions in AI from causing problematic code to be merged into the kernel. Other members of the kernel team are also actively using AI for code review.
Kroa-Hartman credits longtime kernel developer Chris Mason (of Meta Corp.) with being a pioneer in AI code review workflows. Mason has long been involved in AI reviews of eBPF and networking code.
Although AI cannot be comprehensive and there are still errors in some places, AI can indeed find many obvious problems, so Croa-Hartmann always emphasizes that the role of AI is to supplement human defenders rather than authoritatively.