Committed Raw

DHH, the creator of Ruby on Rails, recently wrote that open-source projects blocking AI contributions are gatekeeping. When he speaks, people notice. He has been right about a lot of things. This time, he is half right. He points to real resentment in the community. But he mistakes where it comes from.

The problem is not who gets to contribute. It is what happens when anyone can. Zig, the systems language built by Andrew Kelley, has one of the strictest no-LLM policies out there. No AI-generated code or prose. No paraphrasing LLM output. No LLMs for editing, translation, or even finding bugs. Talking about chatbots is banned. When I first heard about this policy, I thought it was overkill. Then I read why.

Zig runs on a small core team with a big backlog of pull requests. An AI-generated PR does not just sit there harmlessly. It takes review time. And review time is the most limited thing the project has. Kelley calls these contributions negative value. Not zero. Negative. They steal attention from patches written by people who actually know the code. This is not gatekeeping. This is survival. Several well-known open-source projects have shut down public contributions entirely because of this. The low-quality flood makes real progress impossible.

Kelley has given a second reason in interviews that goes further. Zig also tries to train new engineers. The project wants contributors who can form their own ideas, argue about tradeoffs, and solve hard problems. If that is the goal, letting an LLM do the thinking defeats the whole point. You end up with people who can prompt but cannot think.

NetBSD has a different worry. Licenses. An LLM can spit out copyrighted code from its training data. If you commit that code without knowing its origin, the project could be on the hook. This argument is not as strong as Zig's, but it is real. Calling it elitism ignores actual legal risk.

So where is the middle? Mitchell Hashimoto, who built Vagrant and Terraform, created a system called Vouch. The rule is simple. You cannot open a pull request until someone the project already trusts vouches for you. It does not ask what tools you used. It asks whether a human with a track record says your work is worth reading. This flips DHH's whole argument. The restrictions are not about which tools you use. They are about whether your output can be trusted. Vouch moves the gate from tool choice to reputation. It is a trust network, not a tool ban.

I have been coding since 2009. PHP first, then C# .NET, then Node.js and Python. Now I build sovereign AI applications at a startup. I use AI tools. And I know the feeling of losing myself to them. There is a pattern I fell into more than once. You let the AI make a few changes. Those changes pile up. Soon the codebase feels foreign. You lean harder on the AI to keep moving. At some point the project stops being yours. It becomes a blob you talk to in English and hope it does what you mean. I have watched this happen to junior engineers. I have lived it myself.

I got out by setting a hard rule. AI is an assistant, not the pilot. I make every decision. I read every line before it lands. The tools keep getting better. But fundamentals still come first. Always. The future is not for engineers who reject AI, and it is not for engineers who hand over the wheel. It is for people who know what correct looks like, and who use whatever tools get them there faster, without ever losing their own judgment.

The projects drawing lines around AI are not defending status. They are defending attention. Their attention. The judgment of the people who keep the project alive. That is not elitism. That is how things actually work. Ship with protection. Keep the basics solid. Then layer AI on top.