Who pays for the AI patch flood

Open-source maintainers are no longer arguing whether generative models will touch their trees. The fight is over who absorbs the cost when they do. In the space of a few weeks, LLVM floated a ban on AI-written pull request text, the kernel dropped its Assisted-by tag after a Maintainers Summit consensus, Linus Torvalds treated a swollen merge window as evidence that AI work is settling into ordinary development, and Emacs maintainers reopened whether LLM-touched packages belong on NonGNU ELPA at all. The shared complaint is not novelty. It is bandwidth: verbose, low-signal submissions that force scarce reviewers to massage other people's tooling output.

The LLVM discussion is the sharpest statement of that complaint. Nikic's RFC leaves code assistance alone under the existing "human in the loop" rule, but proposes to "ban the use of AI-generated or substantially AI-assisted textual communication," including pull request descriptions and RFC prose. The stated problem is extractive volume. New contributors arrive with long, heavily formatted writeups full of routine testing noise; reviewers spend time reading and rewriting them before the patch itself can be judged. Rengolin put the fairness claim cleanly: by farming the description out, a submitter is "offloading the understanding to the tool and pushing the onus of proof onto reviewers," which "is already provably pushing reviewers away from our project." Jrtc27 added a social edge: even a perfect harness that mimics a developer's voice still means "firing them off to another human to spend their precious time on when you have not done the same is just rude."

Pushback inside LLVM is less about loving slop than about enforcement and second-order effects. Chandlerc described a disciplined workflow (strict prompts, then ruthless human editing) and preferred clearer commit-message norms plus a higher burden of proof on newcomers over a categorical text ban. Void warned that the line between AI and human prose is already blurred: people will lightly edit model output and call it theirs, while hand-written text that uses a semicolon or an em dash can draw false accusations. Who arbitrates? Others answered that communities already punish plagiarism without perfect detectors; the rule sets expectations and enables action when patterns are clear. Shafik and rnk framed the deeper failure mode as pre-LLM review scarcity that models simply hit where projects can least afford it. Chandlerc's harder line was moderator logic: treat extractive contributions the way successful chat rooms treat spam, "delete and ban quickly to preserve the quality of the community," and own that this will feel less welcoming because "we can't welcome spam sustainably."

The kernel is living the same scarcity under different policy instruments. After summit consensus, Sasha Levin posted a patch to remove Assisted-by from the documentation and checkpatch, ending a short experiment in mandatory disclosure of LLM help. The tag was meant to track the evolving role of tools while still requiring a human Signed-off-by and full responsibility under the DCO. Developers judged it useless in practice; disclosure without capacity does not triage a flood. On the lists, Theodore Ts'o cast the maintainer dilemma without romance: faced with waves of newcomer patches, "the choices is either to (a) load-shed, or (b) burn out. And of the two, (a) is the better of two bad choices." Lyude asked why energy goes into automating non-deterministic review aids rather than the shortage of maintainers. Side threads about whether models can be made deterministic drew a blunt dismissal from Linus Torvalds: "the whole 'deterministic' part is a total red herring," a straw man for dislike of the tools, given that compilers, Monte Carlo methods, and humans are already non-deterministic in ways projects routinely manage.

Torvalds's 7.4-rc1 announcement then reframed volume as a possible maturation signal rather than pure crisis. The merge window was unusually large (on the order of 20k commits versus a more common 15k), partly AMD register noise, partly expectations that 7.4 could be the next LTS, and partly what he called the "by now obvious" factor. His preference was explicit: he is "much happier with the flood of AI triggered work going into the merge window than it showing up mainly as extra fixes during the -rc period," hoping assistance is becoming "more part of the normal development" instead of a stabilization-phase surprise. That is not a quality guarantee. It is a claim about when the cost appears, and an empirical bet that earlier, reviewed AI-shaped work beats late thrash. The -rc cycle will test whether the bet holds.

Emacs surfaces a third lever set: legal uncertainty and provenance rather than pure reviewer DoS. Lucius Chen asked whether Clutch, a database client whose development used Codex and Claude Code and disclosed that in Assisted-by headers, could enter NonGNU ELPA. Richard Stallman restated a precautionary GNU line aimed at copyleft enforceability when code may lack a human author: until the project knows whether that is a real problem, "the GNU Project policy is not to include LLM-generated code in GNU packages (including GNU Emacs)." Already-present code is not in a rush to be purged, but should be recorded. Eli Zaretskii and others noted the friction of asking case by case while CONTRIBUTE was silent; Stallman drafted explicit policy text for that file, and debate continues over where assistance (review hints, completions, heavy rewrite of a draft) crosses into installed LLM output. Transparency can look like a penalty when silent packages slip through, which is its own legitimacy problem.

Across these projects the technical substance is mundane and therefore hard. PR text is part of the review interface; a description that does not prove the author understood the diff shifts semantic load onto maintainers. Tags only help if someone reads and trusts them. Bans on communication modes invite cat-and-mouse editing. Accepting AI-shaped code at scale without matching reviewer supply makes load-shedding rational and mentorship optional, which Chandlerc explicitly endorsed as an acceptable loss of some good patches. Rejecting LLM-tinged packages on legal hypothesis protects a licensing theory still under investigation, at the cost of turning away disclosed work.

What remains unresolved is not whether models will keep producing patches and prose. It is which cost-allocation rules communities will treat as legitimate when the next wave arrives: forbid certain AI surfaces, demand tags that may not be read, silently load-shed, or normalize the flood and judge outcomes in the merge window. LLVM's text ban is still an RFC. The kernel has discarded its attribution tag and is watching whether a large, AI-tinged cycle stays healthy through -rc. Emacs has written a clearer freeze on new LLM-generated code while it studies the copyright question. None of those settles the underlying scarcity. Reviewer time is the scarce resource; generative tools made that impossible to ignore.