Greg Kroah-Hartman Will Reject LLM-Generated Patches for Linux’s Staging Drivers

Greg Kroah-Hartman says he will automatically reject patches generated with large language models for the Linux kernel’s drivers/staging/ subsystem, with one narrow exception for real security fixes. He set out the policy Monday in a post to the linux-staging and linux-kernel mailing lists, pointing to a recent “onslaught” of LLM-generated patches.

His reasoning rests on what staging is for. “drivers/staging/ exists PRIMARILY as a place for new kernel developers to learn how to get involved in kernel development,” he wrote. The tree is full of “low hanging fruit” in code cleanups and API changes, and nobody should rely on the code there; loading it sets the kernel’s TAINT_CRAP flag. Maintainers already refuse tool-generated “fix all the coding style issues in this file!” patches because clearing that work out would remove the on-ramp for new contributors.

“So this mean that anyone attempting to use a LLM to cleanup or ‘fix’ any code in drivers/staging/ is EXPLICITLY defeating the whole purpose of it existing in the first place,” Kroah-Hartman wrote. He also warned that hiding AI use won’t work: “it is VERY obvious when people submit LLM-generated patches.”

The exception covers security bugs. LLMs “are very good at finding suspect ‘security issues’ in kernel code these days,” he said, but even the best current and next-generation tools produce results that are “flat out wrong or harmful” at least a third of the time. A contributor who believes an LLM-found fix is valid can still send it, but only after testing it on the actual hardware for the driver, describing that testing, and being willing to defend the change.

Kroah-Hartman, who uses LLMs in his own kernel work, according to Phoronix, described staging as a “gym” for building skills. He wrote that an LLM “is turning into a reasonable tool to do ‘heavy lifting’ for some things,” but only for developers who have already trained enough to know what kind of lifting it can do, and pointed readers to an essay by Bruce Schneier on deciding when to use AI for a task.