Review bandwidth, not model quality, drives the AI contribution backlash

Across three major free-software projects in the same window, maintainers have stopped arguing primarily about whether large language models can write correct code. They are arguing about whether AI-shaped submissions are consuming the scarce resource that keeps projects alive: human review time. The shared pressure is not a sudden discovery that models hallucinate. It is the daily experience of verbose, low-signal patches and package proposals arriving faster than experienced people can absorb them, with burnout already on the table.

In LLVM, Nikic opened an RFC that leaves the project's relatively liberal stance on AI-written code largely intact, while proposing a hard line elsewhere: ban "AI-generated or substantially AI-assisted textual communication," including pull request descriptions and RFC text. The stated motive is concrete. Reviewers are drowning in "extremely verbose" descriptions packed with routine testing boilerplate and "gratuitous formatting," material that takes longer to read and still needs heavy editing before a change is mergeable. Beyond bandwidth, Nikic frames self-authored prose as a sanity check in the project's "human in the loop" model. If the submitter never has to say what the change does in their own words, that check never happens.

That framing drew sharp agreement from people who live in the review queue. Rengolin put the fairness point bluntly: asking a tool to write the description "offload[s] the understanding to the tool and push[es] the onus of proof onto reviewers," behavior he called unfair and already "provably pushing reviewers away." Shafik located the pain where projects can least afford it: "The core issue is code review bandwidth, it has been an issue before LLMs came on the scene but LLMs have caused pain exactly where we can least afford to deal with it." Chandler Carruth went further on process culture, arguing that open source increasingly needs the moderator's instinct: "delete and ban quickly to preserve the quality of the community," because "we can't welcome spam sustainably," even if that means owning a "less welcoming" posture and accepting that some good patches will be missed.

Opposition inside the same thread is less a defense of slop than a warning about unenforceable bright lines. Chandler and others who do use models described ruthless human editing after strict prompting, and asked for clearer expectations about commit-message substance rather than origin theater. Void and others pressed the detection problem: if a hand-written description is accused of being model output because of punctuation habits, who arbitrates, and how does a contributor prove a negative? Jrtc27 answered the etiquette half of that objection without resolving the forensics: even a perfect style clone does not erase the rudeness of "firing them off to another human to spend their precious time on when you have not done the same." The LLVM fight is therefore not really code-versus-no-code. It is whether the project will forbid a class of text that is cheap for submitters to generate and expensive for reviewers to process, while still allowing AI in the diff itself under human responsibility.

The Linux kernel hit a related wall from the opposite documentation direction. After a Maintainers Summit consensus, Sasha Levin posted a patch to deprecate and remove the Assisted-by tag that coding-assistant guidance had required for LLM involvement. Attribution, once sold as a way to track the evolving role of AI, was judged not worth the process weight. The surrounding thread makes clear why a label felt inadequate. Theodore Ts'o, discussing newcomer floods and prioritization, cast the maintainer choice in stark terms: load-shed or burn out, with load-shedding the better of two bad options. Lyude asked why automation of review keeps dominating the conversation instead of the underlying shortage of maintainers. Linus Torvalds, for his part, dismissed one popular objection as a "total red herring": non-determinism is not a special reason to reject LLM-assisted work, he argued, and treating it as decisive is often "a disguise for 'I don't like AI'" rather than an engineering criterion, given that toolchains and even classical algorithms have long produced varying or stochastic results. The kernel conversation thus splits the difference in practice. Tools are not anathematized on philosophical grounds, but ceremonial disclosure tags are being stripped, and the live anxiety is volume, triage, and whether humans remain accountable for what lands.

GNU Emacs and NonGNU ELPA add a third axis that is legal and philosophical as much as operational. Lucius Chen sought guidance before offering Clutch, a database client developed with disclosed help from OpenAI Codex and Claude Code. Richard Stallman restated a cautious GNU line: concern that copyleft "may not be enforcible on code that has no human author," and until that is resolved, "the GNU Project policy is not to include LLM-generated code in GNU packages." He distinguished packages already present (record known LLM-originated material; no rush to delete) from new submissions (do not add them for now), and pushed back on marketing language, preferring "Machine Learning" to "AI." Eli Zaretskii and others noted the awkward incentive surface: transparent authors can be slowed or refused while silent ones may pass, and contributors often learn the expectation only after substantial work. Stallman also pressed for finer distinctions, asking how completely a rewritten draft must change before it is no longer a derivative of model output, while treating pure error-checking automation as harmless. Even where NonGNU ELPA is discussed as more flexible than core GNU packages, the submission inquiry shows provenance policy becoming a gate at the packaging boundary, not merely a commit-message convention.

Read together, the through-line is load management under asymmetric cost. LLVM wants to outlaw AI-authored review text because that text multiplies reviewer minutes without proving submitter understanding. The kernel is discarding Assisted-by because tagging did not solve floods or burnout and may have added ceremony without leverage. GNU is freezing LLM-generated code at the intake door while it worries about authorship, license enforceability, and long-term human maintainability. In each venue, defenders of AI assistance stress residual human responsibility, careful editing, and the futility of pretending probabilistic tools are categorically unlike prior non-deterministic tooling. Critics answer that responsibility theater fails when the scarce side of the market is attention, and that "human in the loop" collapses if the human never had to compress the change into their own words.

What remains unresolved is enforcement and equilibrium. Projects still lack a reliable way to distinguish polished human prose from lightly laundered model output, so rules risk both false accusations and trivial evasion. They also lack agreement on whether the durable fix is stricter intake norms, more maintainers, automated triage that does not itself become nondeterministic noise, or simply accepting a colder contribution culture. The models are not leaving. The open question is which parts of the contribution surface communities are willing to declare off-limits so that the humans who still sign off on quality do not leave first.