The rule, not the model: how Debian, LLVM, Emacs, and PostgreSQL are governing AI in the patch stream

Open source projects are no longer arguing about whether machine-generated text and patches will arrive. They are arguing about which rule is allowed to meet them, because reviewer time and the legitimacy of project decisions are already failing under the volume.

Debian's fight is over a general resolution that both sides claim. Johannes Schauer Marin Rodrigues told debian-devel he will stop uploading to NEW after learning, without an announcement, that queue processing throws uploads at a model. The review itself disclaims authorship: it was not written by a member of the DFSG team and has not been checked by one. His refusal is pointed. "I'm not going to read the machine-generated output." He reads the resolution's line, "Debian neither endorses nor prohibits the use of generative AI", as incompatible with a central team applying a model to every NEW upload. Ian Jackson, quoted in the same thread, called the practice "in egregious violation of even the pro-LLM decision in that vote". Emmanuel Arias, speaking for the people doing the work, answered that the queue is being kept low and that "Nothing is accepted without a proper human review."

The split is not only taste. Jonas Smedegaard argued that even lintian-style mail would be a problem if uploaders were spammed with "content noone could be held accountable for", while granting that probabilistic tools add a tension deterministic ones do not. Richard Laager's counter is institutional: "we voted and we collectively decided LLMs are okay for this project." He holds the DFSG team accountable for automated output in aggregate, whether the tool is deterministic or not, and treats a good-faith experiment as legitimate until results force an adjustment. Smedegaard's reply is that "LLMs are okay" is only one of several summaries, and not the one that uses the resolution's own words. Debian has a ballot and still no shared reading of what responsible use forbids when the person receiving the output did not ask for it.

LLVM is trying to draw the line in a different place. Its standing policy is liberal: essentially all uses are allowed if a human in the loop takes responsibility. Nikita Popov now wants a narrower ban, to "ban the use of AI-generated or substantially AI-assisted textual communication", including pull request descriptions and RFC texts, without changing what is allowed in code. The motive is concrete. Verbose, over-formatted descriptions from new contributors are "a drain on our limited reviewer bandwidth", and they also skip the sanity check of making the contributor explain the change. Jessica Clarke wants the harder line adopted soon so the project can spend less time "fending off the slop". Matt Arsenault is unconvinced that a disclosure-tag regime survives contact with volume. "Enforcement of any kind of policy is additional reviewer burden." He would rather not review model output that the author could have had a model review first, and he sees a spectrum: a short mechanical commit message is not an RFC. Reid Kleckner frames the etiquette as separating human conversation from instructions passed between agents, not as a hunt for perfect compliance. Aaron Ballman noted that the project council reads the tea leaves after discussion dies down; a decision looked more likely in December than the next day. Renato Golin, meanwhile, pointed at fuzzer and auto-fix traffic as the operational wound that has dominated the year.

Emacs is where a temporary copyright hold is being asked to become a permanent social rule. A NonGNU ELPA inquiry for Clutch, an interactive database client, disclosed development with Codex and Claude and asked whether Assisted-by headers made the package eligible. Richard Stallman drew a bright line for new packages and a slower one for old ones. His text, which Eli Zaretskii added to the contributor guide, states the concern that copyleft may not be enforceable on code with no human author, and concludes: "Until we know the answer, the GNU Project policy is not to include LLM-generated code in GNU packages (including GNU Emacs)." Packages already in the archive are to be recorded, not ripped out in a hurry. Stallman also wants the vocabulary changed, machine learning rather than the marketing term, and treats a rewritten draft as still possibly a derivative of model output, while tools that only point at mistakes do not, in his view, require the same worry.

Phil Hagelberg, writing for a group calling for Human Emacs, says legal defensibility is not the crux. "It would be a mistake to say that because something is legally allowable it must be acceptable". The harm they name is to trust and solidarity once generated code is allowed in, a corrosion they say they have watched in other communities. Stallman has said he will forward the incomprehensibility concern to people at the Free Software Foundation thinking about permanent criteria, and he does not treat Emacs's age as inherent safety: a few people could still rewrite some part with nonfree tools. Zaretskii's defense of the status quo is institutional, not moral. Radical changes are rare; when they arrive, many eyes review them because no one person knows the subtle parts. Separately he insisted that "the decision to accept or reject a contribution should not be based on the moral and ethical views and opinions of the contributor." A side argument about slave labor and training data showed how quickly the copyright question slides into whether the product of a practice can be refused without judging the person who submitted it. That question is not settled on the list.

PostgreSQL shows the third rule, triage, arriving before any policy text. Michael Malis, sitting on more than a thousand previously unreported findings from a bug-finding suite, from crashes to MVCC violations, asked for the least disruptive path and offered a catalog rather than a dump. "What's the least disruptive way for me to submit these issues?" Andres Freund's conditions are a de facto policy: deduplicate root causes, minimize reproducers, prefer regression and isolation tests over scripts, skip superuser stunts and convoluted inconsistencies, and separate security issues. "I think we have all grown extremely tired of overcomplicated LLM generated repros." Shihao Zhong, who has been sending model-found bugs, added the practitioner's filter: search the archives first, reproduce by hand, ask which workload would hit it, and review the proposed fix, because "the LLM often fixes the symptom, not the root cause." He will follow whatever policy the community ends up with. There is not one yet.

The through-line is that each project is spending legitimacy to protect attention, and choosing a different instrument. Debian is testing whether a resolution that refuses both endorsement and prohibition can survive an unannounced central workflow. LLVM is testing whether banning generated prose, while leaving generated code to a responsible human, reduces the reading load without creating a second load of enforcement. Emacs is testing whether a hold for copyright uncertainty should harden into a ban on the social grounds that legally allowable is not the same as acceptable. PostgreSQL is testing whether a flood can be shaped by repro quality and security sorting without a written rule at all.

None of those instruments is closed. Debian still has two honest readings of the same ballot. LLVM's council has not adopted the communication ban. Emacs has written the interim copyright rule and has not adopted the permanent one the Human Emacs letter wants. PostgreSQL has a queue of findings and a list of conditions, not a decision about who may send the next thousand. Reviewer time will not wait for the prose to converge.