Torvalds: AI great for hobby, but in the kernel it's stressing maintainers

Linus Torvalds isn't joining the skeptics' chorus. Talking with Dirk Hohndel, Ericsson's head of software technology, at Open Source Summit Europe in Prague, the Linux creator said he "really likes using AI" — then immediately drew a line. Personal toy projects are one thing. Responsibility for the kernel is something else entirely. "You have to be very careful when you use AI for something real and important," he said.

Torvalds explained he doesn't write much kernel code these days; he acts as a collection point, a maintainer. He saves the actual programming for fun, and that's where AI enters as a positive "gateway drug." He described a personal project — a guitar pedal on a microcontroller with a C interface — where he wanted a modern UI but doesn't know Java. He fed the model a prompt, got a terrible result, then showed it his working C code and asked it to adapt. For him, that lets beginners clear the high bar of polished software and feel like they're building something relevant from day one.

The picture changes when it's the kernel. Hohndel asked about Sashiko, an agentic code-review system for the Linux kernel, and Torvalds confirmed its reviews are already appearing publicly on the Linux Kernel Mailing List (LKML). Some subsystem maintainers now require such review before accepting patches, and the expectation is that eventually all of them will. The tools are finding real security issues and surfacing bugs in old, unmaintained drivers that went unnoticed for years.

The benefit is clear — the code base improves — but there's a cost. Torvalds admitted the volume the tools generate is stressing maintainers "to the point where it's a problem." When patches and bug reports arrive faster than the human system can digest them, the technical improvement becomes an operational burden.

A counterpoint from the field: System76 has banned AI-generated content in its COSMIC desktop project. Where Torvalds sees a gateway to fun and learning, others see a risk to quality and maintainability — especially when the tool produces volume the community can't swallow.