KDE has closed a contentious discussion over proposed guidelines for using large language models in project contributions after developers and community members clashed over how AI-assisted work should be handled. The proposal, opened by KDE Plasma developer Nate Graham on September 19, attempted to define acceptable uses of LLMs across coding, documentation and communication. Graham closed the discussion on September 21 and said another attempt would follow after participants had time to cool down.
The draft centered on what it called a “Don’t be lazy” principle. Contributors were told not to replace their own judgment, communication or learning process with an LLM, and not to take shortcuts that would leave maintainers responsible for low-quality work. It also proposed that contributions showing obvious or careless LLM use could be ignored or closed rather than requiring maintainers to spend time repairing them.
For code contributions, the proposed rules emphasized keeping a human actively involved rather than accepting model output with little scrutiny. The draft discouraged submitting “vibe-coded” changes that contributors did not understand or could not reasonably produce themselves, along with throwaway AI-generated work presented as a first draft for maintainers to finish. Contributors would remain responsible for understanding, checking and refining anything produced with AI assistance.
The proposal took an even more restrictive position on LLM-generated communication. It discouraged using models to generate commit messages, merge request descriptions, replies to other contributors or reorganized prose that would simply be copied into KDE discussions. One explicitly accepted example was machine-translating text that a contributor had written in their own language into English without changing its tone or style.
Other uses were treated more permissively. The draft allowed contributors to use LLMs for debugging, research and understanding technical material, provided they independently verified the resulting information. That distinction reflects a broader concern in software development that AI tools can assist investigation without necessarily replacing the human responsibility required when code or documentation is ultimately submitted to a project.
The argument did not remain limited to code quality. Participants raised questions about environmental costs, ethics, transparency and whether contributors should disclose AI assistance. Some objected to an LLM policy that did not make environmental concerns a central requirement, while others questioned whether mandatory disclosure would meaningfully improve review when model use can be difficult to detect reliably.
Those disagreements contributed to the thread being closed. Graham said KDE would try again after taking more feedback into account and suggested that a future proposal could restrict participation to KDE developers, who would ultimately be responsible for applying and enforcing whatever guidelines are adopted. The closed proposal therefore should not be treated as finalized KDE policy.
The discussion follows earlier debate inside KDE over fully LLM-generated merge requests. In September, developers discussed whether maintainers should be explicitly allowed to close submissions that appeared to have been generated almost entirely by AI. Graham’s earlier draft drew from policies used by projects including Jellyfin and Fedora while attempting to adapt those approaches to KDE’s own contributor culture.
KDE’s debate is part of a wider software-industry struggle over how generative AI should be used in development. Organizations are increasingly separating assisted workflows such as debugging or translation from cases where generated material is submitted with little human understanding. Similar concerns have appeared outside open-source development as companies and creators decide how much AI-generated work is acceptable, including disputes such as Level-5’s recent controversy over generative AI use.
No final KDE-wide LLM contribution policy has been announced. For now, the closed discussion represents another draft in an ongoing effort to set clearer boundaries around AI-assisted contributions without creating additional work for maintainers. The next version will need to reconcile practical review requirements with the broader ethical and environmental objections that turned the latest proposal into a much larger community dispute.
