Linus Torvalds is no stranger to making headlines, but his latest foray into the Linux kernel community has caused a particular stir. In response to a mailing list discussion about Sashiko, an AI-powered system that reviews proposed kernel patches, Torvalds stated that Linux is not an “anti-AI project”. His language was characteristically blunt, but the underlying argument was pragmatic: AI tools can be useful, provided they help human caregivers instead of burdening them with extra work.
What Linus Torvalds actually said
The argument is more pragmatic
When Linus, the creator and host of Linux, writes something, the community usually stops what they’re doing to read it. Linux reportedly allows the kernel commit of AI-generated code, but only if the author is willing to take responsibility for that code. AI tools can reduce some of the barriers to software development, explaining unfamiliar code, and preparing an initial application for testing, but it’s up to the contributor to make sure everything is correct.
There are legitimate arguments about AI’s environmental value, training data, misinformation, and economic impact. Torvalds acknowledges that some of these broad questions remain unresolved, but they are separate from the narrower technical question under discussion: Can an AI system provide useful assistance during software development and code review? This is an ongoing discussion between the communities, but Torvalds for Linux is open to using these tools to improve the platform because noted in a mailing list comment:
I understand that some people don’t really like AI, but this is an area where I am fully prepared to put my foot down as a top-level servant.
AI is a tool like any other tool we use. And it’s clearly useful.
The solution is to make sure those LLM tools _help__ them instead of causing them pain. There is no question on that side.
We’re not forcing anyone to use it, but I very loudly ignore people who try to argue against other people’s use of it.
In the kernel community, we do open source because it results in better technology, not for religious reasons.
So we make decisions primarily on technical merit. Don’t be afraid of new tools.
If you read between the lines, Torvalds comments on how these coding tools can be useful, and the practical question is how they can be deployed without causing too much of a burden for human maintainers to handle. Torvalds is not explicitly suggesting that LLM be a kernel maintainer, but instead he advocates the developer’s freedom to use the tool as they see fit, providing code that can be checked, validated, and included in the wider Linux kernel.
It’s more than just using artificial intelligence to code
Linus has previously broached the idea of using AI to help with coding, but this particular mailing list conversation was about AI-powered research, not actual vibe coding. Sashiko is an agent system specifically designed to check for proposed Linux kernel patches. Architecture, execution flow, resource management, security, and hardware bugs are all at its crossroads. Sashiko doesn’t change at all. He simply interprets everything that human understanding requires.
AI is a tool like any other tool we use. And it’s clearly useful.
The project was developed by the Linux Foundation and is funded by Google (for LLM tokens and resources). From the outset, it’s understandable how someone could take what Torvalds has been saying in recent months and ask for a huge patch from a chatbot with a complete startup and redirect it without doing due diligence. That’s not really what’s going on here. Sashiko, in its testing, was able to detect more than 50 percent of bugs based on the last 1,000 top “fix” commits.
AI is here to stay regardless of the bubble
We have already let the genie out of the bag
Even if today’s investment boom in AI-powered programming finally closes, it’s unlikely to go away. Code completion, automated analysis, test generation, and natural language interfaces are now becoming part of development environments. Individual products and companies may fail, but the basic idea of using statistical systems to aid software development is unlikely to disappear with them. Everyone is getting on the AI train.
The AI crashes into everythingmostly indecent but not going anywhere. Even if the bubble bursts like it did in the early 2000s, we will continue to use AI in new and wonderful ways, like our online services like Amazon. This also applies to coding. Chatbots and agents designed specifically to help with code can make software development more accessible. It’s also great for learning because you can actively see how the code works.
The solution is to make sure these LLM tools help caregivers instead of causing them pain.
The same goes for the whole Linux kernel discussion. Torvalds has never stated that AI is right or should be used as the only commit stage, but if it can be used to help caregivers, that can only be positive for the platform as a whole. It’s not about delivering more code thanks to the power of AI, it’s about eliminating bugs earlier, not missing anything, creating sandboxed tests and even suggesting changes within a certain set of parameters.
As someone who used to laugh at all these chatbots and now actively uses LLMs in a local setting, I can definitely agree with Torvalds’ views on how AI can be incredibly useful when used correctly.
Software development can be scary
But AI can help make it more accessible and reliable
I am not a Linux kernel maintainer. I never committed a change to the repo. It’s not because I don’t want to, it’s just because it’s out of my comfort zone. Give me barebones Linux to play with and I’ll be happy for weeks, but actually helping shape the future of the platform that powers everything is a whole different animal. Also, I know no C, I have solid roots in web languages from school and college.
Using AI to complete or make changes to the code before analysis can be a great way to leverage the power of the system to get the code to fix any potential issues. But as Torvalds stated earlierOnce code is submitted from someone who has used AI assistance, that submission is their responsibility. Since the provided code is from the contributor, not the chatbot, there is no “Okay, it was the chatbot”. It doesn’t matter how the code is finalized.
So are my interactions when I get Qwen to do some PHP or JavaScript. I’ll still check every line of code it generates to make sure it matches what I want. I’ll run it in a sandbox to make sure it’s not a gimmick and is suitable for a wider application. Even though Gwen created the lines of code, I still feel it’s my responsibility to make sure it works as intended. The same goes for creating commits for the Linux kernel.
An AI might help me learn how the kernel subsystem works, but that wouldn’t make me qualified to submit a patch that I can’t independently explain and defend.
But the conversation has evolved to include using LLMs to code with Sashiko. In fact, it is completely different. How the tool works is to break down the investigation into phases, verify claims against the application, trace error paths, and look for leaks and bugs. It can even be seen as a form of Grammarly for the core. Like Sashiko, many experienced coders don’t ask LLM to complete the entire task. Very specific with smaller, pre-defined goals.
There is a clear path forward
The strongest objection is not that an LLM can never produce anything useful. Generation is cheap, while the viewing process remains expensive. A contributor can create ten reasonable-looking patches in minutes, while an experienced maintainer can spend hours trying to determine if they’re safe. If not used correctly, AI will not eliminate work and will transfer work from the person who submitted the code to the person responsible for maintaining the project.
The Linux project already has clear and concise guidelines for working with artificial intelligence. Human contributors must review code, supporting code must be labeled accordingly, and contributors must understand and defend each submission. You can’t ask ChatGPT to make heavy kernel changes and submit them without knowing what’s going on. This is the right way to manage AI tools by holding the contributor (not the model) accountable. It also forces tests and proofs with clear disclosure.
A tool like Sashiko is valuable in helping caregivers better manage their increased workload. Sashiko can help improve filtering and hold human contributors accountable, instead of masking every AI-powered commit as anything but AI. Torvalds discourse “Is AI useful?” addressing where it can be used, under what control and who is responsible. AI-generated code flood repositories are terrible, and I would be the first to complain about agents using them.
But using tools like Sashiko to spot potential bugs, handle duplicate lines, or even come up with new ideas before the code goes to the repo can have a positive impact. The choice is not between blindly trusting AI and banning it outright. It’s between careless automation and responsible assistance, and Torvalds is right on this distinction.
