Godot Just Banned AI-Written Code, and the Reason Is Better Than the Headline

TL;DR

On June 30, the maintainers of Godot, the largest open-source game engine, rewrote their contribution policy to ban AI-authored code. Substantial AI-generated code is out. Autonomous agents and "vibe coding" against the repo earn an automatic ban. AI-generated text in human-to-human discussion is out too. What survives is narrow: code completion, regex, find-and-replace, and machine translation of text a human actually wrote. The reason is not technophobia. It is arithmetic. AI made opening a pull request nearly free, review stayed exactly as expensive as it always was, and the maintainers finally ran the numbers on who pays that bill.

What the Policy Actually Says

Strip the outrage headlines and the new policy is precise, not a tantrum. It draws a hard line between AI as an autocomplete and AI as an author. On the allowed side sit the menial mechanics you already use without thinking. On the forbidden side sits anything that lets a model, or a human hiding behind one, ship code nobody can vouch for.

• Banned: substantial AI-generated code, and any autonomous AI agent or "vibe coded" pull request against the repository. Using an agent on the repo is an automatic ban from the GitHub project.
• Banned: AI-generated text in human-to-human communication, meaning issues, proposals, and PR discussion. Godot frames this one as basic respect for the person reading it.
• Allowed: "menial" AI help like code completion, regex, and find-and-replace, plus machine translation of text a human originally wrote.
• Required: if you used AI to author code in any capacity, you disclose it in the PR. Every PR still gets reviewed and approved by a human before it merges.

The Load-Bearing Line: "AI Can't Take Responsibility"

The sentence doing all the work in the announcement is this one:

"AI cannot take responsibility, and we can't trust heavy users of AI to understand their code enough to fix it."

That is not a vibe, it is an operating constraint. An open-source project runs on a chain of accountability: a contributor stands behind a change, a reviewer trusts that contributor, and when the change breaks in two years someone who understands it shows up to fix it. A model breaks every link in that chain. It cannot be on the hook, it cannot learn from the specific feedback a maintainer leaves, and a contributor who only pasted its output often cannot explain what they submitted.

Godot's second worry is quieter and, if you maintain anything, more damning: the mentorship pipeline. Reviewing a newcomer's messy-but-human PR is how you mint the next maintainer. Reviewing a model's output teaches nobody anything, because the model will make the identical mistake tomorrow no matter how patiently you explained it today.

This Is an Economics Problem Wearing an Ethics Costume

Here is the mechanism, and it is the part every maintainer already feels in their bones. The cost of producing a pull request collapsed. The cost of reviewing one did not move a millimeter. Godot says it plainly: the effort to open a PR went down, the number of PRs went up as a result, and the pool of people able to review them stayed exactly the same size.

Think of it like a professor who suddenly gets ten times the term papers because students can generate them in seconds, but still has to grade every one by hand, alone, for free, after work. The bottleneck did not disappear. It just got shoved entirely onto the people least able to absorb it, and the maintainers called that "increasingly draining and demoralizing" back when they started deliberating this in February.

The 3-PR Rule Nobody Is Quoting

Buried under the AI headline is a second change with teeth. Contributors with three or fewer merged PRs can no longer open new features or significant refactors without a maintainer's explicit go-ahead. New folks earn trust on bug fixes and docs first, then graduate to the big stuff. It is an unglamorous, sensible gate, and it is aimed squarely at the drive-by "here is a 2,000-line rewrite from a model I ran once" PR that used to eat a reviewer's entire afternoon.

It Is Not a Blanket AI Ban, and Godot Wants You to Know That

Worth saying clearly, because the headlines flattened it: Godot did not ban AI. It pushed back on the "vibe coded" framing itself, insisting it still tolerates disclosed, human-owned assistance. The rule is about authorship and accountability, not about which editor plugin you run. You can use a model to help you write code you understand, disclose it, and put your name on the result. You cannot let a model be the author and expect a stranger to clean up after it.

And yes, there is a delicious inevitability to announcing an AI-slop ban and then watching a wave of AI-summarized takes about it roll across the timeline within hours. The maintainers, one suspects, are too tired to be surprised.

Why This Matters Beyond One Game Engine

Godot is not a fringe project. It is one of the most-watched repositories on GitHub, with a contributor base big enough that its policy reads as a template other maintainers will copy-paste. This is the first heavyweight open-source project to write down, formally, that the constraint on merging is human review capacity, and that AI throughput without matching review throughput is not a gift but a denial-of-service.

If you build with agents, the signal is not "AI is unwelcome." It is that the finish line for open-source contribution moved. Generating a plausible diff was never the hard part, and now that it is trivial, it counts for even less. The scarce thing is a human who can stand behind the change, answer for it, and fix it at 2am when it breaks. Bring that, disclose your tools, and you are fine. Bring an agent and a shrug, and Godot just made you the cautionary tale.

Key Takeaways

• The change: on June 30, Godot's contribution policy banned substantial AI-authored code, banned AI-generated text in project discussion, and made using an autonomous agent on the repo an automatic ban.
• What is still allowed: disclosed, menial assistance (completion, regex, find-and-replace) and machine translation of human-written text. Every PR still needs human review.
• The real reason: AI collapsed the cost of producing a PR while review capacity stayed flat, and the maintainers refused to eat the difference.
• The load-bearing quote: "AI cannot take responsibility," plus the point that models cannot learn from a reviewer's feedback the way a human contributor does.
• The quiet second rule: contributors with three or fewer merged PRs cannot ship new features or big refactors without maintainer sign-off.
• The takeaway for builders: generating code is no longer the bar. Accountability is. The scarce resource in open source is human review, not more diffs.

Sources: Godot Engine: Changes to our Contribution Policies, PC Gamer, 80.lv, Game Developer