docs(ai-policy): ask for a rewrite before closing over AI-written text

[MiMoHo]

I'm not a native English speaker, and one of my PRs (nextcloud/server#61827) was closed on 23 September with the argument "AI-written description". The footer already disclosed the tool use. After looking at what else was closed that afternoon, I want to suggest a rule change that doesn't cost maintainers more time.

Between 15:27 and 15:47 UTC, ten PRs in nextcloud/server were closed with the same comment. By the next day, seven of them had been reopened:

• fix(accounts): enforce per-property scope restrictions via API server#60348, fix(comments): dismiss mention notifications and clear unread badge when viewed in activity sidebar server#60617, test(db): replace addToAssertionCount checkbox tests in DB server#60750 and feat(ci): add /explain-ci bot command for PR failure help server#60821 were opened in May, before this policy existed.
• In feat(recent-files): allow grouping search results by mime type server#61164 the author wrote the text and used AI only for grammar and to format a results table. The policy allows grammar help.
• In fix(dav): allow ExternalCalendar as schedule-default-calendar-URL server#61602 the author rewrote the description within minutes once asked, and got a review the same day.

Most of them came back within a day, so closing mostly added a round trip and some frustration on both sides. Asking for a rewrite first would have ended in the same place. Not everyone comes back, though: nextcloud/server#62490 was its author's first PR in server and is still closed. In a survey by Steinmacher et al. (ICSE 2018), around 30% of developers whose PRs had not been accepted said this demotivated them or kept them from opening another one.

I understand why the rule exists. By my rough count, people (not bots) opened 1009 PRs in server between June and August this year, compared to 635 in the same months of 2025. Nobody wants to spend time on junk text, but text written with AI help should be accepted when the author has read, understood, and edited it before publishing.

Translated and generated text are hard to tell apart. Liang et al. (Patterns, 2023) tested seven GPT detectors, and on average they flagged 61% of essays by non-native English writers as AI-generated. Jakesch et al. (PNAS, 2023) found that people judge AI text by cues that don't hold up.

Other projects handle this in a similar way. @llvm asks maintainers to first request changes with a standard reply, and @curl recommends mentioning translation tools, so that maintainers don't wrongly dismiss translated text as AI slop.

What this PR changes:

1. The ideas and decisions in a text still have to come from the author. AI may help to draft or phrase it if the author has read, understood, and edited the result and discloses that AI helped.
2. A PR description should be understandable for both the author and the maintainers. If not, maintainers shall ask for a rewrite, by requesting changes or converting the PR to a draft. Without a reply within a month, the PR may be closed.
3. Contributors who write in another language can add a short translation note.
4. PRs opened before a policy change are asked to adapt instead of being closed.
5. CONTRIBUTING.md and the agent snippet (AGENT_MD_SNIPPETS.md) now say the same. Right now the snippet only tells agents not to write descriptions.

Some objections I expect:

• "This lets AI junk in." The bar for the text stays the same: it has to be understandable and the ideas have to be the author's. Maintainer Discretion is unchanged, so extractive PRs can still be closed right away. Only the unclear cases get a question before they get closed.
• "fix: Extend AI communication server#63211 goes the other way." It extends "own words" to code comments and commit messages. In its review, @joshtrichards pointed out that "in your own words" may not be provable and that the final product is what matters, and @AndyScherzinger agreed. @susnux noted that well guided models write good comments. This PR follows the same line, so it might make sense to decide both together.
• "Experienced readers can spot AI text." Russell et al. (ACL 2025) found that people who often use LLMs for writing are very good at it. Their most frequent cue was typical AI vocabulary, and a quarter of their explanations pointed to grammar that was too perfect. Removing grammar mistakes is exactly what the allowed grammar tools are for. The study also only used professionally written American English articles, not text that was translated or only corrected with AI, which is the case the translation exception is about.
• "A month is too long." Converting a PR to a draft is one click and keeps it out of the way until the author replies. The month only decides when an abandoned PR gets closed.

nextcloud/server has its own copies of these rules in .github/CONTRIBUTING.md and AGENTS.md which is why I suggest @AndyScherzinger to link to this repository.

Related open PRs, each of which can be merged on its own:

• docs(ai-policy): use AI-Assisted-By with model and reasoning effort #817 uses AI-Assisted-By: with the model and reasoning effort instead of the agent name.
• docs(contributing): accept a known identity in the sign-off #816 accepts a known identity instead of a real name in the sign-off.
• fix(ai-policy): do not fail fork pull requests on the labelling step #815 fixes the AI Policy check for pull requests from forks (AI workflow fails on forks #767). Until it is merged, the check may fail on this PR.

🤖 AI (if applicable)

• The content of this PR was assisted using AI

AI-Assisted-By: Claude Opus 5.5 (xhigh)

The model helped me collect the cases and sources and drafted the diff. I reviewed and edited everything. The commit carries the same line as a trailer.

[AndyScherzinger]

Hi @MiMoHo

Thank a lot for the PR and the referenced ones in this one. I'll check with others at Nextcloud and get back to you. I do think the changes are going into the right direction and I am very sorry for the bad experience the current policy has led to for you. This should not be the case and hasn't been the idea behind the policy. Thank you for your elaborate and thought through responses.

[skjnldsv]

Seconded! Thank you for taking the time. I think this is nice!

[susnux]

I am fine with AI use there as well, especially commit messages often are better than no commit message at all, leaving you clueless after years digging git blame why something was changes.

A big problem I see with this is that AI wording is just left without rewording.
This is really hard to read in issue or pull request descriptions.
Meaning the AI style language used by default make people quit reading things, not sure this is enforceable because also badly written human text is not nice 😅