When an AI-Generated Answer Actually Works, Who Is the Rule Protecting? A Fair Look at Both Sides — and a Rule Both Could Sign

This week I learned that a programming-language community I take part in keeps an eye on whether members post AI-generated answers. The person who told me was kind about it, the guideline is public, and I had read it. I still found myself turning the question over for days, and I want to work through it here, honestly, from both sides — because I think the communities writing these rules and the members on the wrong end of them want the same thing, and are talking past each other about how to get it.

The question is simple to state. Someone is stuck. Another member asks a model, runs the result, confirms it works, and posts it with an explanation. The community has a rule against AI-generated content. The answer is correct. What should happen?

What the rule is protecting — and it is real

It is easy to dismiss these rules as gatekeeping. That would be unfair, and wrong. The people who wrote them were responding to something that actually happened.

For the whole history of forums, writing an answer cost more than reading one. That asymmetry was the community's immune system: nobody typed three paragraphs of wrong code for fun, so a detailed answer was a weak but real signal that someone had engaged with the problem. Language models inverted that. An answer now costs ten seconds to produce and costs every reader the time to verify. When Stack Overflow banned ChatGPT answers in December 2022, the moderators' stated reason was not pride — it was that fluent, plausible, wrong answers were arriving faster than volunteers could check them. A 2023 study of 517 Stack Overflow questions found that roughly half of ChatGPT's answers contained errors, and that readers often preferred them anyway for their confident style. That is precisely the failure a community's trust system is worst at catching.

There is a second argument moderators make less often and should make more. The model only knows what the forum wrote down. Every model's fluency in a language is a compression of years of humans arguing about problems nobody had documented yet. If the forum fills with model answers, the next genuinely new question — about the library released last month — finds a corpus with nothing new in it. The community is the model's upstream, and protecting that is not nostalgia.

And there is a third argument that is rarely said out loud but is the strongest one: a simple bright-line rule is a deterrent, and deterrents work without detection. "No AI answers" is easy to understand, most honest members will simply obey it, and that alone cuts the flood at the source. A moderator does not need a detector for the rule to do most of its work. Any fair critique of these rules has to reckon with that.

Who the rule also lands on

Now the other side, which is just as real.

Think about who is asking most of these questions. Not a senior engineer with an employer-paid subscription to the best model. Often it is someone on an old laptop, somewhere the good tools cost a week's wages, who has been stuck for two days on something that the people who could answer it in thirty seconds have not got round to. The usual outcome is silence, or "try this, try that" — guesses typed by a human without running anything. That reply passes every AI rule in existence, and it is worth less than nothing, because it is confidently wrong and it closes the thread.

Then a tested, explained, model-assisted answer arrives, and under a blanket rule that is the post that gets flagged. That outcome is not what the rule's authors wanted either. It is a side effect of writing the rule around the one thing that is easy to name — how the first draft was produced — rather than the things the community actually cares about.

Two more facts sit on this side. First, communities never policed provenance before: nobody asked whether an answer came from a search engine, a textbook, or a colleague; the standard was whether it was right and whether you would stand behind it. Second, the detectors that a blanket rule eventually needs do not work well — they have documented false-positive rates, they are measurably biased against people writing in a second language, and the company with the most famous one withdrew it for low accuracy. A rule that can only be enforced by guessing will be enforced against the members who write in an unfamiliar style, which in a global community is a description of the people it most needs.

Both sides want the same three things

When I put the two cases side by side, something became clear: neither is really about authorship. Strip the slogans away and each side is protecting the same three things.

• Correctness. Moderators want fewer wrong answers; askers want an answer that works. Nobody on either side wants the untested guess.
• Honesty. Moderators want to know what they are reading; askers want to know how much to trust it. Nobody wants concealment.
• Capacity. Moderators cannot review a flood; askers drown in one. Nobody wants twenty posts a day from one account.

Authorship is a proxy for all three. It is a reasonable proxy when you have nothing else. But it is a proxy that punishes the careful member and lets the careless human through, and there is a better option.

A rule both sides could sign

Here is the guideline I would propose to my own community — offered as draft language, tiered so a community can adopt the minimum or the whole. I have tried to write it so that a moderator who believes in the current rule could accept it as a strengthening, not a loosening.

Tier 0 — the floor, for every community

• You are responsible for what you post. Use any tool you like; responsibility for correctness cannot be delegated. "The model said so" is not a defence, any more than "a search result said so" ever was.
• Post only what you have run. If you did not execute it, do not post it as an answer. This is the rule communities always had. It simply needs saying aloud now, because the tools made it possible to skip.

Tier 1 — disclosure, recommended

• Say what you used and what you did with it. "Drafted with a model; tested on 1.8.2; I changed the error handling" is enough. Disclosure should be normal and un-shameful, the way "sponsored" labels became normal. Scientific publishing reached exactly this position in 2023 — Nature, arXiv and the ACM all ruled that a model cannot be an author, use must be disclosed, and the human is responsible — without banning the tool. A community that punishes disclosure gets concealment instead.
• Quote, do not launder. If part of an explanation is verbatim model output, mark it, so readers can apply their own discount.

Tier 2 — capacity, where it is strained

• Rate limits by account age and reputation, applied to everyone. This addresses the 2022 flood directly and without guessing at provenance. One verified, explained answer a day is a contribution; twenty is a flood even if each is fine, because nobody can review them.
• A cooling-off period on brand-new questions. For questions about a library or version released in the last few weeks — the ones where the community genuinely is the model's upstream — ask that the first answers come from people who have actually hit the problem. This keeps the forum producing what no model can yet, which is the moderators' best argument, honoured rather than ignored.

Tier 3 — what not to rely on

• Do not treat a detector as evidence. Not a classifier, not a stylistic hunch. The false-positive profile is documented, and the careful poster is the one it misses.
• Sanction the observable thing. Unverified answers, undisclosed verbatim pasting, and volume abuse can all be seen and acted on. How the first draft was produced cannot be, and need not be.

Why is this stricter, not looser? Because under the current rule a human who posts three untested guesses is fully compliant. Under this one, they are not. The rule moves the line from who wrote it to did you check it — and that is a line that catches more bad answers, not fewer.

The conversation I would like to have

If you moderate a community that has one of these rules, I do not think you are wrong about the problem. I think the rule is aimed at the easiest thing to name rather than the thing you care about, and I think a version aimed at verification, disclosure and volume would give you more to enforce, not less. I would genuinely like to hear where that proposal breaks.

If you are the person on the old laptop who finally got an answer that worked: the rules are being written right now, by whoever shows up. Show up, and say what would have helped.

The job of a community was never to make help harder to get. It was to make sure the help was right. Those two goals only conflict if we let the rule be about the wrong thing.