The problem is not that members use AI. It is that writing became free while reading stayed expensive. A post that cost nothing to produce still costs everyone else their attention, and a community is the place where that imbalance shows up first.
Almost every policy written about this fails for the same reason: it tries to regulate how something was written rather than whether it was worth reading.
You cannot tell, and the attempt does damage
Start here, because everything else depends on it.
Detection tools do not work reliably, and their errors are not random. They flag non-native English writers, people who write formally, autistic members whose prose is precise, and anyone who uses a template. A community that acts on detector output will accuse innocent members disproportionately from those groups, and each wrong accusation costs you a person who did nothing.
Your own instinct is not better. Confidently identifying machine text is a skill almost nobody has, including people certain they have it — which is why the rest of this post is about behaviour you can observe rather than authorship you cannot.
What is actually going wrong
Name the real complaint, because "it feels like AI" is a proxy for three different problems that need different fixes.
Volume without stake. Someone posts six long answers an hour on topics they have no history with. The issue is not the prose, it is that nobody is behind it.
Confident wrongness. A fluent, well-structured, plausible answer that is incorrect. This is worse than an obviously bad answer, because it takes an expert to notice and a beginner to believe.
Padding. Four paragraphs where one would do, restating the question before answering it. Members read this as disrespect even when they cannot say why.
Four policy approaches
Only one of these survives contact with a real community.
| Approach | What it catches | What it costs you |
|---|---|---|
| Ban AI use outright | Nothing reliably — it is unenforceable | Honest members who used it to translate or tidy |
| Require disclosure | Only the conscientious, who were not the problem | Little, but it achieves little |
| Use detector tools | False positives, weighted toward non-native writers | Trust, and the members you wrongly accuse |
| Judge effort and accountability | The behaviour you actually object to | Moderator judgment, which is the real cost |
The fourth is the only one that both works and is fair, and it has the useful property of already covering low-effort human posts, which were a problem long before any of this.
The test that works: can they answer a follow-up?
Accountability is observable. Authorship is not.
If someone posts a detailed answer and cannot respond to "why would you do it that way rather than the other way", the post was not theirs in the sense that matters — regardless of what wrote it. Someone who used a model to draft an answer they genuinely understand will handle the follow-up easily, and that person is not your problem. They may be one of your better members.
This test is also fair to state publicly, which the others are not. You can put "be able to stand behind what you post" in a code of conduct. You cannot put "do not sound like a machine" in one.
Banning AI bans your non-native speakers first
This consequence is predictable and consistently overlooked.
For a member writing in their third language, a model is the difference between contributing and staying silent. The same is true for people with dyslexia, and for anyone writing outside their professional register. A prohibition on AI removes those contributions before it removes a single bad-faith poster, because bad-faith posters ignore rules.
If your community spans languages, treat fluency-assistance as legitimate and say so explicitly — otherwise your rule quietly means "post in confident English or not at all", which is the outcome discussed in running a multilingual community.
Raise the cost of posting, slightly
The structural fix is to ask for things a model cannot supply on its own.
Prompts that require specifics — what you tried, what happened, what your constraints are — produce answers only a person with the experience can write. Generic questions invite generic answers, and always did; the flood just made it obvious.
Reward replying over posting. A community that celebrates volume gets volume, and volume is now free. One that visibly values the person who answered a hard question well gets fewer, better posts — the mechanics are in a community engagement strategy.
When it is just spam, treat it as spam
A good portion of what owners call an AI problem is an old problem with new fluency.
Promotional posts, link farms, and accounts posting across nine unrelated spaces are spam whether a person or a model wrote them. These are the easy cases: rule-based handling, no judgment required, and no need to reason about authorship at all — see AI moderation for communities for where automation genuinely helps.
Separating these out matters, because it leaves you with a much smaller pile of genuinely ambiguous posts, and that pile is small enough for a human to read.
Hold yourself to the same rule
Owners write announcements with models too, and members can tell more easily there because they know your voice.
The standard is the same one you are applying to them: you have to stand behind it. A machine-drafted update you edited and mean is fine. A machine-drafted apology is not, for reasons covered in how to moderate an online community — the whole value of an apology is that a person chose the words.
Never let a model answer a member's direct question as if you had. If it turns out to be wrong, you own the answer either way, and you will have spent trust on something you did not read.
Write the rule down in behaviour terms
Three lines cover almost every case, and none of them mentions AI.
Post things you can stand behind and discuss. Do not flood — a few considered posts beat many quick ones. Do not present something as your experience if it is not. Written that way, the rule applies identically to a person padding a post and a person pasting one, which is exactly right, because the community experiences both the same way.
If you also want a disclosure norm, make it social rather than enforced: people mentioning that a model helped them draft something is healthy, and policing whether they mentioned it is not — for how to introduce it deliberately, see how to use AI in your community.
The bottom line
Stop trying to identify machine text, because you cannot, and the attempt lands hardest on members who did nothing wrong.
Judge posts by whether someone can stand behind them, treat obvious spam as spam, ask questions that need real specifics, and apply the same rule to your own announcements. The communities that come through this well will be the ones that were already about accountability rather than output — the rest were vulnerable to a flood of cheap words long before anything could generate them.