The Policy
System76's COSMIC desktop environment now asks every contributor to confirm that their pull request contains no LLM-generated content - and the scope is wider than code. It covers source code, comments, and pull request descriptions.
Announced 2 October 2026 by Jeremy Soller of System76. It makes COSMIC the first major Linux desktop project to issue a blanket ban.
There is one exception: cosmic-flatpak, where upstream projects manage their own manifests.
The Reason Is Not What People Assume
This is not a claim that AI writes bad code. Soller's stated reason is "the extra review work for the COSMIC team".
What the maintainers observed: a rise in submissions from first-time contributors using LLMs, and those submissions were rarely accepted. The cost was not merging bad code. The cost was reading it.
That distinction matters because it tells you the policy would not be reversed by better models. A more capable model raises the acceptance rate; it does not reduce the number of people who can now produce a plausible-looking patch in thirty seconds for a project they have never read.
The Same Break, Two Days Apart
This is the second institution in a week to cap inbound volume rather than improve its filter.
On 1 October, arXiv limited preprint submissions to two a month per submitter after volume doubled in two years. On 2 October, COSMIC banned AI-generated pull requests outright.
Both are the same structural problem: producing a plausible artefact got cheap and assessing one did not. Peer review and code review were both designed on the assumption that writing a credible submission was expensive enough to act as a filter by itself. That assumption is gone, and the institutions built on it are now rationing intake because they have no other lever. Our piece on the arXiv limit covers the same dynamic in research.
Expect more of these. Every maintained open-source project with a public PR queue has the identical arithmetic.
Read the Rest of the Checklist
The LLM clause got the headlines, but it arrived inside a list of confirmations that is worth reading as a whole. Contributors must also confirm they:
- understand their changes
- can respond to feedback
- have tested the work
- certify it under the Developer Certificate of Origin
That was always the bar. Nobody wrote it down before because it went without saying - you did not open a pull request against a project you had not read, because writing the patch required reading the code.
The useful way to read this policy is not "COSMIC banned AI". It is that COSMIC had to write down something that used to be implied, and the thing it wrote down is about comprehension and accountability rather than about tooling.
What It Means If You Contribute
- Check the policy before you open a PR anywhere. This is now a live question per project, not a universal default, and it changes faster than contributor guides get read.
- A ban on generated content is not a ban on using tools to understand a codebase. COSMIC's policy is about what lands in the diff and the description.
- The confirmations are the real test. If you cannot explain your change or respond to review, the policy is not your obstacle - the review is.
No workload figures were published, so the scale of the problem COSMIC is responding to is not public. The policy is the only evidence of it.
FAQ
What exactly did COSMIC ban?
LLM-generated content in pull requests - source code, comments and pull request descriptions. Contributors must confirm their submission contains none of it. The one exception is cosmic-flatpak.
Why did they do it?
Review workload. Jeremy Soller of System76 cited the extra review work on the COSMIC team, after a rise in LLM-assisted submissions from first-time contributors that were rarely accepted.
Is this about AI code being bad?
The stated reason is review cost, not code quality. The distinction matters - better models would not solve a problem caused by how cheap it has become to produce a submission.
Are other projects doing this?
COSMIC is the first major Linux desktop with a blanket ban. arXiv imposed a two-submission monthly limit on preprints the day before, for the same underlying reason.