GCC bans AI-generated code contributions over 15 lines, drawing a line other open-source projects may follow

GCC draws the line at 15 lines of AI-generated code

The GCC steering committee has adopted an AI contributions policy that rejects code produced by or derived from large language models when it crosses the GNU Project's threshold for copyright significance: roughly 15 lines.

The policy states that the project will decline "legally significant contributions which include LLM-generated content or are derived from LLM-generated content" 1. The definition of "legally significant" comes from the GNU Project's maintainer guidelines, which set the bar at "around 15 lines of code and/or text" before copyright paperwork is required 2.

That threshold was not designed for AI. It has served for years as a practical benchmark in the GNU Project's copyright assignment process, marking the point where a contributor has added enough original work to trigger formal copyright paperwork. GCC is now repurposing the same line as the boundary for machine-written code.

The policy carves out exceptions that matter. GCC maintainers may still choose to accept AI-generated test cases that cross the 15-line threshold 1.

That distinction reveals the real fault line the policy is navigating. GCC is not questioning whether LLMs help developers work. It is drawing a boundary around what a contributor can submit under their own name. Using an LLM to find a bug, understand a codebase, or review a patch is permitted. Submitting the model's output as a contribution is not. The question the policy forces on every contributor is whether they can stand behind every line they submit.

The 15-line threshold raises two immediate practical questions: who counts the lines, and what counts as "derived from"?

The GNU guidelines already address the first. They warn that "a series of minor changes by the same person can add up to a significant contribution" and note that "it is irrelevant which parts of it were contributed when" 2. Splitting an LLM-generated patch into 14-line chunks would not clear the bar. The threshold applies to a contributor's cumulative body of work, not to individual commits.

The second question is harder. "Derived from LLM-generated content" is broad language. A developer who asks a model for a structural suggestion and then writes the code by hand has arguably produced work derived from LLM output. So has someone who pastes a function from a model and refactors it. GCC's policy does not yet distinguish between these scenarios, and the committee says it expects the policy to evolve and will revisit it periodically 1.

Whether the approach holds up depends on enforcement, and the LWN comment thread splits along familiar lines. One commenter argues that policies banning AI produce covert AI use, and that rules punishing honesty breed dishonesty 1. Another counters that contributors who use AI covertly will eventually be caught when they cannot answer questions about their own code, and that the resulting reputational damage is the real deterrent 1. A third commenter reports having identified AI-derived patches from non-regular contributors fairly consistently, because the output tends to contain fundamental errors in close proximity to code that is otherwise correct 1.

Both sides of the enforcement debate have a point, and the tension between them will shape whether other projects adopt similar rules. GCC sits inside the GNU Project, the same ecosystem whose licensing framework shaped how free software handles attribution and ownership. A policy adopted here carries weight beyond GCC's own codebase. Other maintainers facing the same questions will read this rule and ask: can a line-count threshold hold when AI coding tools produce output that increasingly resembles human-written code, and can a community enforce a policy that depends on contributors self-reporting their use of tools many developers now reach for by default?

GCC's answer is to start with a number borrowed from copyright law, exempt the use cases where AI output stays out of the codebase, and revisit the rest. The committee is explicit that the current wording is a starting position, not a final one. For maintainers and builders watching the open-source ecosystem, the policy is a concrete reference point for a debate that is only beginning. Whether the 15-line threshold holds or bends under pressure will tell other projects how far these boundaries can stretch.