Nostal6ie via iStock / Getty Images Plus
Follow ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- Linux kernel developers are using AI to support project maintenance.
- Writing kernel code with AI remains an open question.
- Copyright and open-source licensing concerns still loom.
The Linux developer community has rapidly shifted from debating how to use AI to quietly embedding it deeply into the Linux kernel engineering workflows. Indeed, Linus Torvalds now describes himself as “a huge believer” in AI as a maintenance tool.
At the same time, kernel developers meeting in Tokyo for Open Source Summit (OSS) Japan, Linux Kernel Maintainer Summit, and Linux Plumbers were formalizing how large language models (LLMs) will fit into long-term processes such as stable backporting, Common Vulnerabilities and Exposures (CVE) triage, and tooling policy.
Also: Why AI agents failed to take over in 2025 – it’s ‘a story as old as time,’ says Deloitte
As Torvalds said in his OSS Japan keynote, AI has finally reached the point where it is genuinely helpful for Linux maintainers, rather than just hype around code generation. He emphasized that his interest lies in AI systems that pre-screen patches and merges, surfacing issues before they reach his inbox, rather than in tools that attempt to write complex kernel code outright. AI is not ready for that yet.
Before the conferences, Sasha Levin, an NVIDIA distinguished engineer and stable-kernel maintainer, had spelled out where the Linux kernel maintainers and developers already had agreement on AI’s use: Human accountability is non-negotiable; some form of disclosure is needed when AI is used; and maintainers will decide for themselves how to use AI tools.
The maintainers still haven’t reached consensus on some issues. For example, does AI make fundamentally different kinds of errors than humans? Should AI work be held to a different standard than human-produced code? And what are the legal ramifications of introducing AI code into the kernel?
Also: This new Linux platform will let you update your next car at home – as soon as 2027
This last is a hot-button issue. How copyright and AI mix hasn’t been settled yet. Major lawsuits are still making their way through the courts. Complicating matters, it’s well-known that AI coding tools were largely trained on open-source code under a variety of different open-source licenses.
Using LLMs on thankless jobs
Earlier this year, Levin revealed he’d already wired LLMs into two of the most thankless jobs in the project: identifying backports and security fixes. For AUTOSEL, the system that identifies kernel patches for backporting to stable releases, Levin rebuilt the selection engine around embeddings, treating each commit as a point in a semantic space and asking LLMs to find patches similar to past backports. The system distributes queries to multiple models, lets them “vote,” and then hands shortlisted candidates to human maintainers, effectively acting as an extra stable maintainer that never forgets historical decisions.
The same infrastructure now underpins Linux’s in-house CVE workflow, which took on responsibility for assigning its own CVE numbers in early 2024. The CVE team replaced brittle shell scripts with Rust tooling largely drafted with LLM help. Levin used AI both to generate code and to help classify which commits are likely to be security-relevant. Levin argues that these tools do not replace human judgment but make an otherwise unscalable, seven-days-a-week triage process doable as the kernel patch volume continues to grow. Linux maintainer burnout is a known, serious concern. It’s a thankless job.
Also: The fix for messy AI agent ecosystems might finally be here – and it’s open source
Levin also said a patch merged for Linux 6.15, credited to him but entirely generated by AI, with changelog and tests included. The patch, a small but non-trivial hash-table change, demonstrated, in his view, what AI does well: narrowly scoped, mechanical transformations in which the model can reason about bit-fields. He added that AI tests more patiently than a tired human. A separate AI-generated git-resolve script, added for 6.16 to resolve ambiguous commit IDs and shipped with unusually thorough self-tests and documentation, is now used daily by kernel developers.
But the AI-written hash patch also contained a subtle mistake: it dropped a __read_mostly attribute, a performance hint that should have been preserved, prompting a wave of criticism once the omission was dissected in the LWN’s comment threads. Commenters also argued that undisclosed AI authorship violates the Developer’s Certificate of Origin and community expectations of honesty and review.
Torvalds himself later said he would have scrutinized the change more carefully had he known it was 100% tool-generated. This is why there will be an explicit tag or disclosure required when programmers turn in AI-derived code patches.
At OSS Japan, Torvalds said he believes LLMs should be treated as the next step in the evolution of compilers rather than as human replacements. He compared their adoption to the shift from assembly to higher-level languages: initially controversial, but eventually accepted as a way to free developers from drudge work such as writing boilerplate or meticulously drafting commit messages in a second language. In practice, developers are already using LLMs as powerful linters and summarizers for both code and email.
Also: Gemini’s command line tool is a productivity game changer, and it’s free – how I use it
Some maintainers also told me that AI-driven preselection can reduce a day’s worth of patch triage to minutes. For those of you who don’t know, the Linux kernel development pipeline runs on mailing lists, specifically the Linux Kernel Mailing List (LKML). This is an enormous saving of time and energy. Shuah Khan, a Linux Foundation fellow, agreed that AI has proven very useful for us in reviewing and improving productivity.
Beyond developer tooling, the broader Linux ecosystem is also moving toward AI workloads and AI-assisted operations. Hardware vendors are racing to expose accelerators and NPUs to Linux. At the same time, storage and filesystems are being tuned to feed GPU pipelines at scale. In parallel, projects outside the kernel, from academic experiments like Code-Survey to commercial offerings, are exploring how LLMs can map large codebases, track feature lifecycles, and uncover bug-dense subsystems by mining commits and mailing lists.
Not everyone is convinced
Despite glowing reports, not everyone is convinced about AI yet. One senior maintainer said, while most people’s “experience with AI felt falsely positive. We heard concerns and are working to bring AI into our processes responsibly. It’s something that can definitely help reduce the burden on maintainers and developers. On the other hand, we need to be very careful about bringing something this new and that’s currently proprietary into our workflow.”
Also: I’ve tested free vs. paid AI coding tools – here’s which one I’d actually use
Jonathan Corbet, senior kernel developer and editor-in-chief of LWN, added he was concerned about maintainers “depending on proprietary tools. If we become dependent on this for our review process, we really risk repeating the BitKeeper disaster of 20 years ago.” BitKeeper was the first Source Control Management (SCM) system used to track Linux’s code. Its creator changed its license so it could no longer be used for Linux. The result, while wonderful in the long run — it caused Torvalds to create Git — was extremely upsetting at the time.
Another concern shared by Shuah and Dan Williams, an Intel senior principal engineer and kernel maintainer, is how to ensure new developers actually understand what they’re doing. Williams said, “I do career talks at high schools, and I tell them the most important thing you can learn in school, and you will use it, is to ‘show your work.’ And I feel like AI is the ultimate, ‘I don’t have to show my work because the AI told me it is correct.'” That’s not helpful.
Also: Bad vibes: How an AI agent coded its way to disaster
This is a real problem both for people learning how to code and for people who want to move from being entry-level coders to mature programmers. As Stefania Druga, research scientist at Sakana AI and former AI research scientist at Google DeepMind, said in an OSS Japan presentation, AI tools are automating many of the junior tasks that once served as stepping stones for newcomers.
“We need experience to get a job, and a job to get that experience,” she said. “If AI is going to automate these entry-level tasks, how are real people supposed to learn these skills?” Good question, and one that students, employers, and Linux kernel developers must all address.
That’s a future problem.
Questions remain
For today, the trends are converging on the same conclusion: AI is becoming part of Linux’s plumbing rather than a bolt-on gimmick. For now, the focus is squarely on augmenting maintainers under crushing patch loads, automating the most tedious parts of stable and security work, and capturing the benefits of pattern-matching machines.
Also: Why people keep flocking to Linux in 2025 (and it’s not just to escape Windows)
Will AI eventually write substantial parts of Linux’s code? Stay tuned. That question remains unanswered, and its resolution may end up depending more on copyright law than on technical expertise.
