Add AI policy to contributor guide.

[Dirbaio]

Initial proposal of AI policy. Feedback welcome, I'll leave this up for a few days.

[leftger]

I agree almost 100% on everything in the policy. The one issue I think would need more clarification is what counts as "low effort". I myself have sometimes introduced 1-2 minor bugs that other individuals introduce 1-2 line PRs to fix them. That's not low effort. In those cases a lot of times people have referenced the reference manuals to prove their point. I think adding some clarification like that would avoid scaring people away from not contributing actual fixes to the codebase (regardless of AI/human-origination)

[Dirbaio]

Good point, what wording would you use?

[bugadani]

Not sure if this should be so strict, a may could also work - sometimes something can be trivial enough, but useful enough, to just accept, no matter the packaging.

[Bluefinger]

Stuff like this is to set a clear boundary for maintainers to be able to act without having to weigh every PR for "usefulness" if it fails to meet the requirements set by the contributor's guide. Being willing to enforce one's code of conduct is vital to maintain community health.

If something is that trivial & also useful enough, then the code can easily be redone by a maintainer or picked up by another contributor.

[leftger]

To your ask dirbaio:
"Blind AI-Generated PRs are NOT allowed. Submitting direct and unedited LLM outputs without human oversight or local testing is strictly prohibited. This rule does not apply to small fixes; single-line bug corrections backed by reference manuals or tests are highly encouraged. The requirement is simple: you must personally run, test, and understand the logic of your contribution."

[ivmarkov]

You must review and understand every single line of AI-written code"

A bit too strong? And for a largish PR I don't think that would be the case anyway. Even if the human claims he did it, probably he didn't. A code review theater.

How about:
"You must thoroughly understand the AI-generated code, and be prepared..."
or
"You must understand all AI-generated code in the PR, and be prepared..."

[bugadani]

How do you understand something that you did not review?

[ivmarkov]

"...every single line"?

[nponsard]

Google translate now enables its "advanced" mode by default when translating some languages, this mode uses an LLM and suffers from the same LLM flaws...

This means hallucinations could occur. I sadly don't have a good alternative to recommend.

[0e4ef622]

I can recommend DeepL as a potential alternative.

[jordens]

I would add a short preamble that sets out the intention and the goal before listing the individual rules.

(patch display might end up being crappy here. this is not a rewrite of the rules, mostly a reordering/regrouping)

Suggested change

	## AI Policy
	- Using AI / LLM to write code is **allowed**.
	- You **must** disclose AI usage in the PR description.
	- You **must** review and understand every single line of AI-written code before submitting the PR, and be prepared to answer questions and make fixes to it.
	- Fully autonomous agents like OpenClaw are **NOT allowed**. Every issue, every PR must have a real human behind it.
	- Using AI / LLM to write PR descriptions, commit messages, comments, or replies to reviewers is **NOT allowed**. You must write them yourself. AI must not replace human-to-human communication.
	- If you need help writing English, consider using a traditional translator like Google Translate instead of LLMs. They do a great job without the pitfalls of LLMs (overly verbose text, hallucinations, etc.).
	- Low-effort pull requests are **NOT allowed**. Don't simply paste the issue text into an LLM and submit the result. It will most likely be a bad contribution. If fixing the issue was so easy, the person opening the issue or a maintainer would've already done it.
	- Consider the time it takes to review your PRs. Maintainer time is the most scarce resource in the Embassy project. Keep diffs minimal, ask first before doing thousands-of-lines refactors, split PRs if you can. This has always been good practice, but it's especially relevant now that LLMs make it too easy to generate mountains of code.
	- Pull requests not following this policy will be closed.
	## AI Policy
	Embassy allows honest AI/LLM use as long as it keeps the burden of responsibility with the human PR author and appreciates the valuable human PR reviewer time. AI must not replace human-to-human communication.
	Consider the time it takes to review your PRs. Maintainer time is the most scarce resource in the Embassy project. Keep diffs minimal, ask first before doing thousands-of-lines refactors, split PRs if you can. This has always been good practice, but it's especially relevant now that LLMs make it too easy to generate mountains of code.
	Pull requests not following this policy will be closed.
	- Using AI / LLM to write code is **allowed**.
	- You **must** disclose AI usage in the PR description.
	- You **must** review and understand every single line of AI-written code before submitting the PR, and be prepared to answer questions and make fixes to it.
	- Fully autonomous agents like OpenClaw are **NOT allowed**. Every issue, every PR must have a real human behind it.
	- Using AI / LLM to write PR descriptions, commit messages, comments, or replies to reviewers is **NOT allowed**. You must write them yourself.
	- If you need help writing English, consider using a traditional translator like Google Translate instead of LLMs. They do a great job without the pitfalls of LLMs (overly verbose text, hallucinations, etc.).
	- Low-effort pull requests are **NOT allowed**. Don't simply paste the issue text into an LLM and submit the result. It will most likely be a bad contribution. If fixing the issue was so easy, the person opening the issue or a maintainer would've already done it.