When an AI-Generated Answer Actually Works, Who Is the Rule Protecting? An Open Question for Programming Communities

This week I learned that a programming-language community I post in keeps an eye on whether members are posting AI-generated answers. The person who told me was kind about it. The guidelines exist, I had read them, and I understand why they were written. But the conversation left me with a question I could not put down, and I want to work through it here rather than in a comment thread where it will turn into a fight.

The question is this. If someone is stuck, and I ask a model, and the answer it gives me works — I ran it, it compiled, it fixed the thing — and I post that answer, who was harmed? The person who asked got unstuck. The thread now has a correct answer on it for the next person. What, exactly, is the rule protecting?

I have an answer, and it is not the one either side of this argument usually gives.

What the rule is actually protecting

The honest case for these rules has nothing to do with where an answer came from. It has to do with cost. For the whole history of forums, a written answer cost the writer more than it cost the reader. That asymmetry was the community's immune system: nobody typed out three paragraphs of wrong code for fun, so volume was a weak but real signal of effort, and effort was a weak but real signal of correctness.

Language models broke that. An answer now costs the poster ten seconds and costs every reader the time to verify it. When posting is free and verification is not, the scarce resource flips from answers to reader attention, and a community can be drowned by members who mean well. That is what happened to Stack Overflow in late 2022, when it banned ChatGPT-generated answers: not because moderators were proud, but because they could not keep up with confident, fluent, wrong posts arriving faster than anyone could check them. That was a real problem. It still is.

So the rule is protecting reader trust against unverified volume. I agree with protecting that. I do not agree that "AI-generated" is the right thing to ban to protect it.

Why "no AI" is the wrong rule

Nobody polices whether an answer came from a search engine, a textbook, a colleague, or a memory of a bug you fixed in 2019. The community has never cared about provenance; it cared about whether the answer was right and whether the person posting it would stand behind it. "No AI" quietly replaces that standard with a different one — how was this drafted? — which is both unenforceable and beside the point.

It is unenforceable because detection does not work and will work less every year. Every editor, IDE, and browser is growing a model inside it. Within a couple of years, asking whether a model touched a piece of text will be like asking whether spellcheck did.

It is beside the point because it bans the good case along with the bad one. A member who asks a model, runs the code, understands what it does, fixes the part that was wrong, and posts the working result with an explanation has done exactly what the community wanted from a human answerer. Banning that answer because of how the first draft was produced does not protect anyone. It just leaves the person who asked without help, and it tells the people who verify that their verification does not count.

And there is an uncomfortable part of this that I think has to be said plainly: some of the enforcement is not about quality. It is about identity. "I suffered to learn this, and you should too" is a feeling I understand — I have it — but it is not a policy, and dressing it up as one is how communities become clubs.

Who the rule actually lands on

Here is the part that made me want to write this, and it is not about me.

Think about who is on the other end of most of these threads. Not a senior engineer with a company subscription to the best model. Someone on a six-year-old laptop, somewhere the good tools cost a week's wages, who has been stuck for two days on a problem that the people who could answer it in thirty seconds cannot be bothered to look at. They post. Nobody replies. Or worse, someone replies with "try this, try that" — three guesses typed without running anything, which cost the replier nothing and cost the asker another evening. That answer was written by a human. It passes every AI-content rule on the books. And it is worth less than nothing, because it is confidently wrong and it closes the thread.

Now someone else comes along, asks a model, runs the result, sees it work, and posts it with an explanation. That person just did more for the asker than the whole community managed. Under a "no AI" rule, that is the post that gets flagged.

I cannot see how that is justifiable. A rule that lets the untested guess through and stops the tested answer is not protecting quality. It is protecting the comfort of people who already have access to everything, at the expense of the people who have the least. The asker on the cheap laptop does not have the model. The answerer does. Standing between them and calling it standards is the opposite of what a community is for.

Why the community still needs to exist

The other side of my own question deserves a straight answer too. If the model already knows the answer, why have a forum at all?

Because the model only knows what the forum wrote down. Every language model's fluency in a programming language is a compression of years of humans arguing in threads about exactly the problems that were not yet documented. The community is not a slower version of the model. It is the model's upstream. The day the forum becomes a mirror of model answers, there is nothing new for the next model to learn, and the next genuinely new question — the one that is not in any training set, because the library was released last month — has nobody to answer it.

That is the strongest argument for these communities, and it is also the strongest argument against flooding them. A forum full of regurgitated answers does not just waste readers' time; it poisons its own future.

Where the line actually belongs

I think the workable rule has four parts, and none of them mention AI.

• Verification is the bar, not authorship. If you ran it, it works, and you can explain why, post it. If you did not run it, do not post it, no matter who or what wrote it. This is the rule communities have always had. It just needs to be said out loud now.
• Accountability stays with the poster. Use any tool you like. You cannot outsource responsibility for being wrong. "The model said so" is not a defence, any more than "Google said so" ever was.
• Disclose, and own the changes. "Drafted with a model, tested on version 1.8, I changed the error handling" is acceptable almost anywhere, and it is more honest than a human pretending. Disclosure should be normal and un-shameful, the way "sponsored" labels became normal.
• Volume is a separate problem from quality. One verified, well-explained answer a day is a contribution. Twenty a day is a flood even if each one is fine, because nobody can review them. Communities are within their rights to rate-limit, and that is a better tool than trying to guess provenance.

That is the balance. Regulate unverified content. Regulate volume. Require disclosure. Leave authorship alone.

My own position, with the disclosure that goes with it

I should say where I stand, because I am not a neutral party. I use language models every day. Some of the writing on this site is drafted with them, and I say so. The work I am proudest of here is the work where the model was the least of it — the SQL engine I benchmarked on my own hardware, the microcontroller I ran a model on and timed, the series where I probe a model live before I write a word about it. Those posts are defensible on any platform because they add something a model could not have produced on its own: a measurement, on a real machine, on a real day.

I think that is the test that will survive. Not "did you use AI?" but "what did you add?" If the answer is "I ran it and here is what happened," you belong in the thread. If the answer is "I pasted what it said," you do not, and it does not matter how good the paste was.

Where this is heading

In the short term, more monitoring, more bans, and a great deal of collateral damage to people who verify their work and are treated the same as people who do not. That is where we are now, and it is why I wrote this.

In the medium term, I expect the detection arms race to be abandoned as unwinnable, disclosure to become ordinary, and the rules to converge on the one they should have started with: you are responsible for what you post.

In the long term, the communities that survive will be the ones where people post what the models cannot — new problems, measured results, things they actually built and broke. The ones that become AI-answer mirrors will fade, not because of any rule, but because nobody needs a slower copy of the model.

If your community is writing these guidelines right now, I would ask one thing. Write down what you are actually protecting. If it is reader trust, say so, and build the rule around verification and disclosure. If it is the feeling that answers should be earned the hard way, say that too — and then ask whether the person stuck at two in the morning agrees with you.

I do not think they will. I think they want the answer that works. The job of a community was never to make that harder. It was to make sure it was right.