AI bans are a symptom, not the disease
Open source AI code bans are emerging as a defensive response by maintainers who are overwhelmed by low-quality, AI-generated pull requests that erode mentoring, clog review queues, and threaten the long-term health of collaborative projects by rewarding quantity over understanding. Godot Engine’s decision to largely outlaw AI-generated contributions is not a crusade against tools, but against “vibe coding slop” that turns their GitHub into a dumping ground. The team is rewriting its contribution policy to prohibit almost all use of AI from contributors, after an overwhelming number of pull requests, many apparently AI-authored, flooded in. That flood does not mean a wave of new, committed developers; it means more review work for the same small group of people already carrying the project.
Godot’s hard line: saving mentors from vibe-coded slop
Godot’s maintainers are blunt: “AI contributions have the added pain of being demoralizing.” They say AI cannot take responsibility, and they cannot trust heavy AI users to understand their code enough to fix it. Review has always been tedious, but it doubled as mentoring—feedback went to a human who might become a future maintainer. Once the other side of the pull request is a coding agent or a user pasting uncomprehended output, that payoff disappears. So Godot is updating its AI contribution policies to bar any substantial AI-generated code, whether it comes from a bot or a human using an assistant, even if the output is reviewed and disclosed. New contributors, defined as those with three or fewer merged pull requests, must now get explicit permission before submitting new features or large refactors, a move aimed squarely at keeping “vibe coders and AI agents” out while nurturing people who engage with the codebase.
Burnout on GitHub: more PRs, not more people
The harsh stance is driven by maintainer burnout, not a collapse in newcomer interest. GitHub reports merged pull requests have grown from about 25 million per month in early 2023 to roughly 90 million a month today. However, a study of 1,888 repositories that adopted AI coding agents found newcomer participation held steady or ticked slightly upward, with the worst estimate being a statistically insignificant 1.5% dip. In other words, AI tools are not pushing newcomers out the door in established projects at current adoption levels. The real pain point is review volume and AI generated code quality: pull request volume has nearly quadrupled, the code is more complex, and some contributors “may not fully grasp what they’re proposing.” For overworked volunteer maintainers, a backlog full of half-understood AI output is not innovation—it is unpaid triage duty that saps morale.
Compilers, AI, and the real fight over quality
Linus Torvalds’s view exposes the core paradox. He argues that boasting “ninety-nine percent of our code is written by AI” misses the point, because “one hundred percent of their code is written by compilers.” Compilers have translated human-readable code into machine instructions for decades, yet no one claims that GCC “writes” their C++. AI sits as another abstraction layer above the compiler: humans express intent, AI generates source, then compilers and assemblers do their usual work. The fundamentals of programming—problem decomposition, logic, architecture—do not vanish. Godot’s ban is not a rejection of this abstraction stack; it is a demand that contributors still understand the intent and structure behind their changes. When AI-generated code increases cyclomatic complexity by 3–4% and cognitive complexity in Python by about 11%, as the agent-adoption study reports, maintainers are right to ask whether today’s tools are making their code harder to keep reliable tomorrow.
The future of collaborative development: intent over output
Zero-tolerance AI contribution policies are spreading; the Zig language adopted one in April, explicitly arguing that the point of reviewing a pull request is investing in the person behind it. Godot’s current rules are already conservative and the Foundation expects to revisit them as tools change, but their direction is clear: protect maintainers first. Contributors can still use AI for narrow tasks—code completion, regex, or mechanical find-and-replace—if they disclose it, but autonomous agents and vibe-coded garbage trigger an automatic ban. This is the new bargain for developer workflows: AI can help, but only where humans remain accountable and present. For ordinary contributors, it means slower “spray and pray” coding and more focus on understanding the patch they submit. For open source leaders, it is a reminder that healthy projects are built on relationships and shared judgment, not on limitless, low-effort output.
