GCC AI Policy: The 15-Line Rule Is Not in the Text

GCC AI policy on AI-generated code copyright: what is declined

The GCC AI policy declines any legally significant contribution that includes LLM-generated content or derives from it. GCC is the compiler suite that turns C, C++ and other source code into programs a machine can run, and an LLM (large language model) is the kind of AI that generates code or text from a prompt. "Legally significant" is the policy's own phrase, and it comes with no number; the figure of 15 lines is from a separate GNU guide.

On July 29, 2026, David Edelsohn announced on the GCC mailing list that the GCC Steering Committee had accepted the policy recommended by the GCC AI Policy Working Group. Jonathan Wakely led the group, and its members were Carlos O'Donell, Sudakshina Das, Jason Merrill, Joel Sherrill, Sam James, Robin Dapp and Arthur Cohen. The committee said it expects the policy to keep evolving and will revisit it periodically.

So the policy sets a test of legal significance, not a ban on every LLM-assisted patch. It has one accepted case and two carve-outs:

• Accepted: maintainers may take a legally insignificant LLM-generated contribution if it meets the usual requirements for any contribution and is clearly marked.
• Exception: maintainers may take legally significant contributions to test cases, even when an LLM generated all or part of them.
• Excluded: code that does not primarily belong to GCC but is imported from other projects, such as libsanitizer, is outside the policy.

Where 15 lines comes from: GNU guidance on AI-generated code copyright

The GCC AI policy contains no line count. The number 15 comes from GNU's guide for maintainers of GNU software, in the section on legally significant changes. The policy page uses only the words "legally significant" and does not cite the guide.

The guide sets the threshold for copyright paperwork. A person who contributes more than around 15 lines of legally significant code and/or text needs to provide copyright papers for that contribution.

A change of just a few lines, fewer than 15 or so, is not legally significant. The guide adds two points that change how the number reads:

• A regular series of repeated changes is not legally significant. The guide's example is renaming a symbol, even when the symbol is renamed in many places.
• Small changes add up. A series of minor changes by the same person can become a significant contribution. What counts is the person's total, and it does not matter which part came when. When recording a small patch, the guide says to look for earlier changes by the same person first and check whether the past and present changes together are legally significant.

Read together, "around 15 lines" is a rough size for one person's total contribution, and it is not a per-patch allowance. The guide itself hedges the number ("around", "or so"), and the GCC policy sets no cutoff of its own.

Assisted-by tag on every LLM commit under the GCC AI policy

Every commit message for a contribution of LLM-generated content must include an "Assisted-by:" tag. A commit message is the note attached to each change in the project's history, and the tag is a labeled line in it that shows an LLM helped produce the change.

The wording is "any contribution", so the tag is not limited to large patches. A small bug fix drafted with an LLM can be accepted when it is legally insignificant, meets the usual requirements and is clearly marked, and its commit message still has to carry the Assisted-by tag. The same applies to the test-case exception above: a legally significant test-case contribution can be accepted, but the tag is still required.

The tag only records that a machine took part. It does not replace the human steps the policy requires.

Signed-off-by and the Developer Certificate of Origin: humans only

Only a human may add the "Signed-off-by:" tag, which certifies the Developer Certificate of Origin (DCO). The DCO is a short statement a contributor makes about where a change came from and their right to submit it, and Signed-off-by is the line in a commit message that attaches that statement to the change. The policy puts the rest of the responsibility on people as well:

• Every contribution must be submitted by a human who understands the changes and is prepared to answer questions about them.
• The decision to include a contribution in the project must also be made by a human.
• An LLM may not commit code to the project repository.

In practice, a contributor who uses an LLM still has to understand the change, be ready to answer questions about it and sign off on it personally. The policy bars the tool from signing off or committing.

Research, review and debugging fall outside the policy

The policy does not cover AI used for research, analysis, bug discovery and reporting, patch review or debugging, as long as the output is not included in the contributions to the project. The policy adds two notes. AI output used for bug discovery and reporting should not be sent verbatim without due consideration, and AI in patch review should support human review, not replace it.

Personal use is outside the policy too. Screen readers, text-to-speech, direct translation and spelling or grammar assistance are named, provided the contributor verifies the tool's output.

The boundary is the output. If an LLM helps someone track down a bug and the person then writes the fix by hand, the policy does not apply. If the LLM's output is pasted into the patch, the policy applies, including the legal-significance test and the Assisted-by tag.

Conclusion: the GCC AI policy follows copyright, not line count

Three points sum up the GCC AI policy. The test for declining a contribution is "legally significant", not a number of lines. The figure of around 15 lines belongs to GNU's guide for maintainers, where it is a rough size, and under that guide small changes from one person are counted together.

The policy is a "for the time being" rule. It is expected to evolve with the community or as the overall GNU Project position is updated, and it will be reviewed by the start of 2027 at the latest.

The policy document is published under CC0 1.0 Universal. To check which lines differ between two versions of a text, try the text comparison tool below.