GCC, the compiler behind most of Linux, will reject LLM-written contributions over its 15-line copyright threshold

The GNU Compiler Collection, the toolchain that turns C, C++, and Fortran into the binaries most Linux distributions ship, has published its first written rule on AI-assisted contributions. The threshold it uses is a copyright-meaning test, not a code-quality one.

The steering committee has accepted an AI contributions policy recommended by the GCC AI policy working group. The rule, reported by LWN, declines any "legally significant contributions which include LLM-generated content or are derived from LLM-generated content." What counts as legally significant is not GCC's invention. The threshold inherits the GNU Project maintainer guidelines, where "around 15 lines of code and/or text" qualifies as significant for copyright purposes.

The threshold answers the question the policy is asking: who counts as the author under copyright law. A project that ships under a copyleft license has a real interest in being able to answer that. The same bar applies to a human contributor pasting a block from a vendor's docs or a Stack Overflow answer without attribution, so the AI clause is grafted onto an existing authorship test rather than a new rule.

The policy is narrower than the wire's "GCC bans AI" framing. LLM tools are still allowed for research, analysis, bug discovery, bug reporting, and patch review, provided the output is not included in contributions. Phoronix's coverage flags the same carve-out. A maintainer can use an LLM to read a 2,000-line patch, find a defect, and write a review comment. The same LLM cannot, under the policy, generate a 20-line fix that goes into the tree, because the fix crosses the 15-line threshold and counts as legally significant under the inherited bar.

Two carve-outs in the policy are the parts to watch. Maintainers may still choose to accept legally significant test cases that are LLM-generated, because tests sit in a separate category of contribution under copyright practice and are easier to attribute cleanly. The steering committee also explicitly frames the policy as one that is expected to evolve and be revisited periodically. The committee is publishing a starting point it expects to revise.

The open question is whether the template travels. GCC is the first major compiler toolchain to publish a written contribution rule that uses an existing copyright threshold as the boundary between AI assistance and AI authorship. Other large projects that have been debating similar policies, including the Linux kernel and the LLVM/Clang project, will be asked whether their own rules, formal or informal, line up with this one or diverge from it. Hacker News commenters, some of them self-described maintainers, are already weighing in on whether 15 lines is the right cutoff or a defensible one at all.

Inheriting the GNU threshold also changes the politics of changing it. The 15-line bar is older than the AI policy, and it is the bar a maintainer already applies when a contributor pastes a block of code from a third-party source. Lowering the bar for AI-generated contributions, or raising it for everyone else, would mean rewriting the underlying maintainer guideline, not just the AI clause. The committee's "evolving" language acknowledges that without committing to a specific revision.

For now, the rule is in place, the policy text is on the GCC site, and the committee has reserved the right to revisit it. The next live test is whether an LLM-authored patch that just barely crosses 15 lines counts as a contribution or as a tool-assisted review note. That distinction is the policy.