Developer Tools

System76 COSMIC Bans LLM-Generated Code, Comments, and PR Descriptions

System76's COSMIC PR template now bans LLM-generated code, comments, and descriptions. Jeremy Soller says review load from first-time AI PRs drove the change; cosmic-flatpak stays exempt.

By Abhinav Ramaswamy October 4, 2026 7 min read

System76’s COSMIC desktop project has drawn a hard line on AI-written pull requests. Contributors must now certify that a PR contains no LLM-generated content, covering source code, comments, and the PR description itself. Incomplete checkboxes can get the PR closed.

The rule is live in the project’s GitHub pull request template, including on cosmic-epoch, cosmic-comp, and cosmic-edit. It is not a soft disclosure policy. It is a ban on LLM-authored material in submissions.

That stance landed on Hacker News and Linux coverage in early October 2026, as more open-source maintainers wrestle with the same flood of AI-assisted first-time PRs. For Pop!_OS and COSMIC users watching how AI coding tools reshape open source, this is one of the clearest desktop-project responses so far.

What the New PR Checklist Requires

The template replaces the usual “describe your changes” note with a checklist. Contributors must read and acknowledge every item. PRs without a completed checkbox will be closed.

The first item is the one that matters most for AI tools:

I have not included any LLM (also known as AI) generated content in this PR, including code, comments, and descriptions.

The rest of the list is familiar open-source hygiene, restated with more force:

The contributor understands the changes in full and can answer review comments.
The commit message accurately describes the change.
The contribution is tested and works as described.
The contributor has read the Developer Certificate of Origin and certifies the contribution under its conditions.
A final checkbox confirms they followed all of the above, with a link to the Pop!_OS contributor guide.

That combination treats LLM output like unreviewed third-party code. If you cannot defend it in review, it does not belong in the PR.

Why System76 Says It Did This

On 30 September 2026, System76 engineer Jeremy Soller posted the policy rationale on FLOSS.social. He said COSMIC projects will no longer accept LLM-generated content in PRs, and that the primary driver was review load rather than a blanket cultural crusade against AI.

“The primary reason is simply our team’s load for reviews,” Soller wrote. “In particular, we would like to prioritize working on contributions from our own team and regular contributors, while we have been receiving far more changes from first-time contributors using LLMs. These changes are often unplanned and have had a low acceptance rate.”

In a follow-up post the same day, he pointed to the revised PR template on cosmic-epoch and noted one exception: cosmic-flatpak. There, upstream projects manage their own Flatpak manifests that point at their own code. COSMIC’s team mainly reviews those manifests for correct sandboxing before accepting changes.

That framing is narrower than “AI is banned forever.” It is a maintainer bandwidth decision: AI made it cheap to open PRs, and expensive to review them.

What Is Confirmed vs. What Coverage Added

Claim	Status
COSMIC PR template bans LLM-generated code, comments, and descriptions	Confirmed in GitHub templates across multiple pop-os COSMIC repos
PRs without completed checklist can be closed	Confirmed by the template wording
Primary reason is review load from first-time LLM contributors	Confirmed by Jeremy Soller on FLOSS.social (30 Sep 2026)
cosmic-flatpak is an exception for upstream Flatpak manifests	Confirmed by Soller’s follow-up
All AI tool use is banned, including bug finding	Not stated in the template or Soller’s posts; secondary writeups note non-generative uses are not clearly barred
Pop!_OS as a whole banned AI from every codebase	Overstated in some headlines; the published rule is the COSMIC PR template policy Soller described

How This Differs From Disclosure-Only Policies

Many projects still allow AI-assisted patches if contributors disclose them and take responsibility. COSMIC used to sit closer to that world. Coverage from Linuxiac and XDA Developers notes that contributors had already been asked to disclose AI-generated code and understand their submissions, and that large LLM-written PRs were often rejected in practice.

The new template is stricter. Disclosure is not enough. The PR must contain no LLM-generated content at all. That is a different governance model from projects that accept high-quality AI patches with human review.

It also sits against a wider 2026 backdrop. Linux kernel discussions and Ubuntu process changes have, according to secondary reporting, been strained by AI-generated bug reports and patches. COSMIC’s answer is closer to a hard reject than a “label it and ship it if good” rule.

What It Means for Developers Using AI Coding Tools

If you use Cursor, Claude Code, Copilot, or similar tools, the practical takeaway is simple: do not paste model output into a COSMIC PR and call it done. The checklist asks you to certify the opposite.

That does not automatically ban every local use of AI. Reading docs, exploring a stacktrace, or drafting notes outside the PR are not what the checkbox names. The named ban is content inside the PR: code, comments, and descriptions. Secondary coverage has argued that non-generative uses such as finding bugs are not clearly forbidden. System76 has not published a longer FAQ drawing those lines in official docs beyond Soller’s posts and the template.

For regular contributors, the message is also about prioritization. Soller said the team wants more time for its own work and for people who contribute often. First-time, LLM-heavy drive-by PRs are the load being cut.

This lands next to other stories we have tracked about AI coding agents changing open-source risk. Anthropic’s Claude Code mods can rewrite prompts and tool calls without a sandbox, which raises a different kind of trust question for agent plugins. Separately, Glow Labs reported AI coding agents leaking more than 13,000 internal screenshots to public GitHub repos. COSMIC’s policy is about review capacity, not those incidents, but all three point to the same pressure: AI makes it easy to generate volume, and humans still own the merge button.

What It Means for Indian Developers

India has a large population of developers using AI IDEs on Linux and contributing to open source from laptops that often run Pop!_OS or other Ubuntu-family systems. If you contribute to COSMIC or plan to, treat the checklist as a hard requirement, not etiquette.

Practical habits that still fit the rule: write the patch yourself, use AI only as a reference outside the PR if at all, test on your machine, and be ready to answer review questions line by line. A low-acceptance first PR that burns maintainer time is exactly what Soller said the team is trying to reduce.

For Indian teams building products on COSMIC or packaging Flatpaks, the flatpak exception matters. Upstream manifests can still move through cosmic-flatpak when sandboxing is correct. Application code in COSMIC itself cannot ride in on LLM prose.

What to Watch Next

Three signals will show whether this hardens or softens:

Whether more pop-os repositories copy the same template language.
Whether System76 publishes a longer contributor FAQ that clarifies allowed AI uses outside PR text.
Whether other desktop environments adopt similar bans, or stick with disclosure-only rules.

Until then, the verifiable facts are narrow and public: the GitHub checklist, Soller’s two FLOSS.social posts, and the cosmic-flatpak exception. Everything else in the discourse is interpretation.

Frequently Asked Questions

Did System76 ban all AI use in Pop!_OS development?

No official statement says that. What is published is a COSMIC PR template rule: no LLM-generated content in the PR, including code, comments, and descriptions. Jeremy Soller also carved out cosmic-flatpak for upstream Flatpak manifests reviewed mainly for sandboxing.

When did the policy appear?

Soller announced it on FLOSS.social on 30 September 2026 and linked the revised cosmic-epoch PR template. Broader Linux coverage followed around 2 October 2026.

Can maintainers still close PRs that look AI-written even if the box is checked?

The template already says incomplete checkboxes can lead to closure. Reviewers can still reject changes for quality, copyright, or maintainability, which Linuxiac notes was already common for large LLM-written PRs before the hard ban.

Is cosmic-flatpak included in the ban?

Soller said there is an exception for cosmic-flatpak, because upstream projects manage their own manifests and COSMIC mainly checks sandboxing.

Does this ban using AI tools to learn about the codebase?

The published checklist bans LLM-generated content inside the PR. It does not spell out every offline use of AI tools. If you contribute, the safe reading is: do not put model-generated code, comments, or descriptions into the PR you ask humans to merge.