I feel we need to have something written we can point to, but I do not have the time or patience for a debate on this subject, so I’ve put down a strict ban for now, on the principle that it’ll be much harder to tighten than to loosen the rules later.
zw --- README-hacking | 32 +++++++++++++++++++++++++++++++- 1 file changed, 31 insertions(+), 1 deletion(-) diff --git a/README-hacking b/README-hacking index db88c9a6..746b50dd 100644 --- a/README-hacking +++ b/README-hacking @@ -107,6 +107,8 @@ that, too: * Submitting patches +Thank you for helping improve Autoconf! + All patches should be submitted to <[email protected]> for review, in context or unified diff format against the latest sources. In most cases, a patch should include a test case, to ensure that @@ -133,7 +135,35 @@ you'd use doc/Copyright/request-assign.future: https://git.sv.gnu.org/cgit/gnulib.git/tree/doc/Copyright/request-assign.future -Enjoy! +** Usage of LLMs and other “generative AI” tools when working on Autoconf + +Autoconf is written in an unusual combination of shell, M4, and Perl, +and must generate configure scripts that work even on ancient systems +with buggy and/or non-standard shells. Therefore, application of +mechanical code-auditing tools (LLM-based or otherwise) to Autoconf, +or to generated configure scripts, is likely to produce a large volume +of false positive diagnostics. The Autoconf maintainers do not have +time to dig through that volume to find real bugs, if any. Please do +not send us any such reports. For similar reasons, please do not use +LLMs to “help” you write bug reports, even when you found the bug “by +hand”, as it were. + +As of August 2026, there is no official GNU-wide policy on the +acceptability of LLM-generated code in GNU projects. Autoconf has +only a few people working on it nowadays, and arguing about the +acceptability of LLM-generated patches is not a good use of our time. +Therefore, Autoconf’s interim policy is to decline *all* patches which +include, or are derived from, content generated by an LLM or similar +“generative AI” tool, *whether or not* they are large enough to +require copyright assignment. + +Use of an LLM as a personal assistive tool (e.g. for automatic +translation between natural languages) does not *necessarily* make +your patch “content generated by an LLM;” however, use of machine +translation when communicating with Autoconf’s maintainers is strongly +discouraged. If you are not comfortable communicating in English, we +would rather you use a language you *are* comfortable in. Trust us to +find a translator if we need one. ----- -- 2.54.0
