Linus Torvalds spent weeks chasing a stubborn graphics glitch. The display on his Battlemage G21 system kept going black. He applied 24 debug patches. He rebooted the kernel 18 times. And every step of the way, the AI assistant he consulted insisted the problem was impossible. Unsolvable.
But Torvalds kept going. The fix turned out to be a one-word change: round_up() became round_down(). The bug lived in the Intel graphics driver. MakeUseOf first reported the details on September 13, 2026, highlighting how the episode reveals far more about responsible AI use than most coverage captured.
The stubbornness that mattered
The fix landed in commit 818bebe. Torvalds let the AI write the final commit message and credited it by name, MakeUseOf noted. The AI didn’t just fail to find the answer. It actively discouraged further pursuit. Multiple times it declared the issue “impossible and unsolvable.” Torvalds pushed anyway. He reviewed every suggestion. He tested each hypothesis himself. The machine held the flashlight. The human decided where to point it.
This wasn’t vibe coding. No one handed the problem to an LLM and accepted the first output. Torvalds treated the AI as one voice among many. A fast but fallible one. He overruled it repeatedly. The result stands as working kernel code. Short. Direct. Correct.
Contrast that approach with the flood now hitting Linux maintainers. AI systems scan the kernel and surface thousands of potential issues. Greg Kroah-Hartman showed a graph at Kernel Recipes 2026 in Paris. For several years the project fixed roughly 500 CVEs per release. From 7.0 onward the number jumped to 1,000. Linux 7.2 exceeded 1,500. The count may top 2,000 in 7.3. TechSpot covered the surge on September 2, 2026. Most reports come from AI code review platforms and large language models. Many target old or deprecated code. Yet the volume overwhelms human triage.
Torvalds flagged the problem months earlier. In his Linux 7.1-rc4 post he described the private security mailing list as “almost entirely unmanageable.” Reports climbed from two or three per week two years ago to five to 10 every day in 2026. Maintainers spend hours filtering obvious noise. They hesitate to dismiss patches that might contain real fixes.
So the kernel project built rules. A formal AI Coding Assistants policy sits in the documentation. Contributors must add an “Assisted-by: LLM” trailer to commit messages when machine help played a role. AI agents cannot sign off with the legally binding Signed-off-by line. Only humans certify the Developer Certificate of Origin. The tag lives in git history. It can be counted. A companion document explains the rationale. The Next Web explained the policy on September 7, 2026.
Torvalds reinforced the stance in July. On a linux-media mailing list thread about an AI code review system called Sashiko, critics argued for restrictions. Torvalds responded bluntly. “Linux is not one of those anti-AI projects, and if somebody has issues with that, they can do the open-source thing and fork it. Or just walk away.” He added that he would “very loudly ignore” attempts to ban others from using the tools. AI counts as “a tool, just like other tools we use. And it’s clearly a useful one.” Ars Technica reported the exchange on July 16, 2026.
His rc2 announcement for Linux 7.3 carried the same wry tone. The release candidate arrived heavier than expected. Filesystem fixes, DRM patches, networking and driver updates piled up. Torvalds could not pin the size on one cause. “It might be just random,” he wrote. “But we’ll obviously all blame it on AI, because whether that’s really the cause or not, it’s an easy thing to blame.” He added a winking emoticon. The joke landed amid real shifts. AI tools scan continuously. They surface bugs humans once missed or deprioritized. The volume has become the new normal.
Yet the Intel graphics story shows the difference. Torvalds did not accept the AI’s conclusion. He audited its suggestions. He iterated. He owned the final change. That pattern matches the project’s explicit guidance. Humans review. Humans test. Humans sign off. The Assisted-by tag records assistance without surrendering responsibility.
The Linux kernel now carries more than 40 million lines of code. It runs on everything from watches to supercomputers. Changes ripple across millions of systems. A single flawed patch can cause widespread outages or security holes. In that environment, overreliance on any automated system carries risk. Torvalds’ debugging session offers a case study in mitigation. Question the machine. Test its assumptions. Keep pushing when it says stop.
The commit that fixed the graphics bug contains no grand claims. No marketing language. Just the technical record. The AI said impossible. Torvalds disagreed. One character change later, the black screens stopped. The episode deserves attention not because it proves AI works. It shows how skilled developers make AI work. They stay in control. They treat outputs as hypotheses, not truths. They document assistance transparently.
That standard travels beyond the kernel. Companies racing to embed LLMs in their engineering workflows face the same choice. Accept every suggestion and drown in technical debt. Or build processes that record assistance, demand human review, and reward engineers who know when to overrule the model. The second path scales. The first collapses under its own weight.
Torvalds has never required developers to adopt AI. He simply refuses to let opponents block those who find it useful. The policy remains technical. Based on merit. Measured by results. The Intel driver fix passed that test. The flood of AI-generated reports tests it daily. So far the project holds the line. Humans decide. Machines assist. The git log tells the story.
And the story continues. Linux 7.3-rc3 arrived shortly after the heavy rc2. More fixes landed. More AI scans ran. The maintainers keep triaging. Torvalds keeps reviewing. The process evolves, but the core principle does not. Good code comes from good judgment. AI can sharpen that judgment. It cannot replace it.
WebProNews is an iEntry Publication