Artificial Intelligence 5 min read 57m ago

Open Source Rebels Against AI Code

A review of 120 open source projects found that 37 have adopted a total ban on AI-assisted contributions, with GCC, QEMU, SDL, Gentoo, Zig, and Ghostty among those rejecting all
Debian's developer community is currently voting on whether to prohibit AI use across code, documentation, translations, and bug reports, with the outcome still undecided.
Greg Kroah-Hartman stated on the Linux developers mailing list that at least one third of outputs from even the best current AI tools are "flat out wrong or harmful", a figure

A significant and growing number of open source software projects are banning or tightly restricting contributions generated by large language models, reflecting a deepening scepticism among senior developers about the practical value of AI tools in professional engineering environments.

The scale of the backlash

A review of AI policies across 120 open source projects, conducted by researcher Rakshit Yadav, found that 37 had adopted a total prohibition on AI-assisted contributions. Projects including GCC, QEMU, SDL, Gentoo, Zig, and Ghostty have moved to reject all such submissions outright. Development platforms Codeberg and Sourcehut, along with the app store Flathub, have extended those restrictions further — prohibiting AI-generated software, documentation, bug reports, and review comments alike.

The Linux kernel takes a more measured position: AI-assisted contributions are permitted, but the LLM used must be attributed, in the interest of transparency. Projects that do permit AI involvement typically require a human to have read and filtered all generated output before it is submitted for review, in an attempt to curb what critics have labelled "AI slop".

Ad Unit · 728×90 / Responsive

Debian's unresolved vote

The debate is now playing out within one of the most influential Linux distributions. Debian's developer community is currently voting on whether to permit or prohibit AI use across contributions — a ballot that includes a proposal for a total ban covering code, documentation, translations, and bug reports. The outcome remains undecided.

The timing is pointed: the AI datacentres that power large language models themselves run predominantly on Debian and other Linux-based systems, and those models have been trained extensively on the very open source codebases whose maintainers are now considering excluding them.

Why senior engineers are pushing back

The case against liberal AI use in open source centres on the asymmetry between those generating contributions and those reviewing them. As AI tools lower the barrier to submission, experienced maintainers report being inundated with poorly structured, LLM-generated code that consumes significant human time to evaluate and reject — time that, in volunteer-driven projects, is in finite supply.

Greg Kroah-Hartman, a prominent figure in Linux kernel development, stated on the Linux developers mailing list that "even with the best of the current and next generation tools, at least 1/3 of the results they generate are flat out wrong or harmful." The dynamic has been compared to long-standing "script kiddie" behaviour, in which inexperienced users deploy tools they do not fully understand, generating work that costs reviewers far more effort to assess than it took to produce.

A secondary concern is pedagogical. Senior developers report a reluctance to mentor junior contributors who appear to be outsourcing comprehension to AI rather than developing their own understanding — a pattern that, if it persists, risks degrading the pipeline of expertise on which complex systems depend. As analysis of what genuinely drives productivity suggests, tools that bypass human judgement and the accumulation of real expertise tend to underdeliver on their promise.

What benchmarks show — and what they obscure

AI models are measurably improving on standardised evaluations. The best frontier models achieve a pass rate of approximately 50% on Humanity's Last Exam and resolve just under 77% of issues on SWE-bench, a benchmark focused on software engineering tasks. Yet those figures simultaneously confirm a substantial, persistent error rate. In engineering contexts — unlike consumer entertainment applications — a failure rate of roughly one in four carries material consequences.

This tension, between the commercially promoted progress visible in benchmark dashboards and the day-to-day frustrations reported by practising maintainers, lies at the heart of the growing divide between AI optimists and those closest to production codebases.

A tool whose limits are being underestimated

The argument advanced by those most familiar with how large language models function is not that the technology is without value. It is, rather, that LLM outputs are often fluent and confident in appearance even when technically incorrect, and that many users lack the domain expertise necessary to distinguish correct results from plausible-looking ones — a problem of information asymmetry that market distortions have historically proved slow to correct.

With Debian's vote unresolved and policy debates spreading across the open source ecosystem, the question of how to govern AI-generated contributions is fast becoming one of the defining institutional questions in software development — and the answer being reached, in project after project, is that without meaningful human accountability, the costs outweigh the benefits.

Topics

Artificial Intelligence Open Source SoftwareSoftware Engineering

People

Greg Kroah-HartmanRakshit Yadav

Organizations

DebianFlathubLinux

Ad Unit · 300×250 / Responsive