Fifteen lines. That is the number the GCC steering committee landed on, and it is the reason this policy is more interesting than the headline suggests. On July 29, 2026 the committee accepted a policy declining any "legally significant contributions which include LLM-generated content or are derived from LLM-generated content," and it borrowed the GNU Project maintainer guidelines' existing definition of legally significant: roughly 15 lines of code and/or text. The decision arrived by email and a policy commit, 4d0793a6a14bf9bfe9e92ac1599840780355199d.
GCC is not a project that can shrug about provenance. It compiles most of the software running the internet, it lives under the Free Software Foundation's copyright assignment regime, and its legal posture has been the model other GNU projects copy for decades. When it says it cannot accept code whose copyright status is unclear, that is not squeamishness about AI. It is the same rule it has always had, applied to a new source of unclear ownership.
RelatedCan an AI Lawyer Replace a Real One?
What the policy actually permits
The ban is narrower than "no AI near GCC," and the exceptions tell you what the committee was actually worried about. Using an LLM for research and analysis is fine. Using one to find and report bugs is fine. Using one to review patches is fine. What you cannot do is put the output into the contribution itself. Maintainers may also accept legally significant test cases that an LLM generated, which is a carve-out worth noticing: a test case asserts what the compiler should do, it is not the compiler, and if a machine writes a good regression test the copyright exposure is a different shape.
Read together, those exceptions draw a clean line. LLM as a tool that informs a human is fine. LLM as a source of the artifact being contributed is not. The committee also said the policy "will evolve and will be revisited periodically," which is what a group writes when it knows the legal ground has not settled.
How does this compare to other projects?
| GCC | QEMU | Gentoo | NetBSD | Linux kernel | Fedora | LLVM | |
|---|---|---|---|---|---|---|---|
| Stance | Declines significant LLM content | Declines suspected AI content | Forbids AI-assisted contributions | Presumes AI code tainted | Permits | Permits | Permits with limits |
| Mechanism | GNU 15-line threshold | Blanket refusal | Unanimous Council vote | Core approval required to commit | Assisted-by: trailer | Assisted-by: recommended | Unreviewed AI treated as extractive |
| Carve-outs | LLM test cases accepted | None stated | None | Case by case | Agents may not add Signed-off-by | Disclose if used largely unchanged | Banned on good-first-issue tickets |
| Core concern | Copyright provenance | Copyright provenance | Copyright, quality, ethics | Copyright provenance | Human accountability | Transparency | Reviewer burden |
The split is not really optimists against pessimists. It tracks legal structure. GCC operates under FSF copyright assignment, where somebody must be able to say who owns a contribution and sign paperwork saying so. "A model produced it" is not an answer that regime can process. The Linux kernel runs on the Developer Certificate of Origin, where a human attests they have the right to submit, so the kernel could solve its version of the problem by naming the human and making them the liable party. Both projects made the only move their own licensing machinery allowed.
Where did the line-count idea come from?
- 2024Gentoo Council votes unanimously Contributions made with natural-language AI tools are forbidden, citing copyright, quality and ethics.
- 2024NetBSD treats AI code as tainted No commits without explicit core approval, using the project's existing tainted-code framework.
- 2026Linux kernel settles on
Assisted-by:After long debate, AI-assisted code is permitted and the human submitter carries the liability. - Jul 29, 2026GCC steering committee adopts its policy First to attach a measurable threshold by reusing the GNU 15-line definition of legal significance.
- OngoingPeriodic revisit The committee explicitly expects the policy to change as the copyright picture clarifies.
An alternative circulated during the debate. Guix proposed splitting contributions on "creative" versus "non-creative" rather than counting lines, which is arguably closer to how copyright law actually thinks. It did not pass. The line count won because it is checkable, and a rule a maintainer can apply at 11pm without consulting a lawyer beats a more accurate rule nobody can apply consistently.
Is a 15-line rule enforceable?
Not really, and the committee has to know that. There is no detector that reliably identifies LLM-generated C, and any classifier claiming to would be wrong often enough to be worse than useless when the penalty is rejecting a legitimate contributor. Enforcement rests on contributors telling the truth.
That is less naive than it sounds. Policies like this mostly do not work by catching people, they work by removing the excuse. Before July 29, a contributor pasting model output into a GCC patch could tell themselves nobody had said not to. Now the norm is explicit, so the same act requires deciding to break a stated rule, which is a meaningfully different thing psychologically and a much stronger position for the project if a dispute ever reaches a court. The rule is a liability instrument at least as much as a technical control.
Who does this actually affect?
Fewer people than the reaction will suggest. GCC's patch flow comes overwhelmingly from a small population of long-term contributors employed by Red Hat, SUSE, ARM, AMD and similar, and those engineers already work under employer policies about generated code. The casual drive-by contributor pasting Copilot output into a compiler backend is largely hypothetical, because compiler internals are one of the areas where current models are least useful anyway.
RelatedKimi K3's Weights Are Due Today. Can You Run Them?
The real effect is precedent. GCC is the reference implementation for how GNU projects handle copyright, and a numeric threshold is portable in a way a philosophical stance is not. Expect other projects to reuse the 15-line rule specifically, because it is far easier to adopt a number than to relitigate the whole question. Also expect friction as more contributors use agents by default and start having to reason about whether their own patch crossed a line they did not measure.
- Whether other GNU projects copy the number. The 15-line threshold is the exportable part. Watch glibc, binutils and GDB.
- The first real dispute. Policy gets tested when a valuable patch is rejected on suspicion alone. How the committee handles that will matter far more than the text.
- Whether the test-case carve-out widens. If LLM-generated tests prove genuinely useful to GCC, pressure to extend the same logic to other non-creative code will follow.
- Movement on AI output copyrightability. The entire policy is downstream of an unsettled legal question. A clear ruling in either direction forces a revisit, which the committee already said to expect.
Our take
The interesting decision here is not the ban, it is the refusal to invent anything. The committee could have written a bespoke AI policy with new definitions and new processes. Instead it reached for a definition of legal significance that has been sitting in the GNU maintainer guidelines for years and pointed it at a new problem. That is the correct instinct, and it is why this policy will probably outlive most of the AI policies written in the last two years.
The obvious objection, that 15 lines is arbitrary, is true and does not matter much. Every workable threshold is arbitrary. What a threshold buys you is that two maintainers looking at the same patch reach the same conclusion, and that is worth more than a philosophically superior standard that produces inconsistent outcomes. Guix's creative versus non-creative framing is a better description of copyright and a worse rule to hand a volunteer.
Worth being honest about the limits, though. This policy protects GCC's legal position, not its code quality, and those are different goals that happen to point the same direction today. If model output becomes reliably good at compiler work while its copyright status stays murky, the tension gets uncomfortable fast, and the periodic revisit clause is where that fight will happen.
- OfficialLWN: GCC steering committee announces AI policy — the policy text, commit hash and the Guix alternative that did not pass
- ReferenceGNU maintainer guidelines — the pre-existing definition of legally significant that supplies the 15-line figure
- GovernanceGCC steering committee — who makes this class of decision for the project
- BenchmarkGENZ TECH AI Coding Leaderboard — independently sourced scores for the models this policy is about
Original analysis by GenZTech, built from LWN's report on the GCC steering committee AI policy published July 30, 2026.