Does It Matter Where a Correct Answer Came From? AI-Generated Help in Programming Communities, Examined From Both Sides

This is a position paper. It argues both sides before taking one, cites what it can, and says where the evidence is thin. It is written to be disagreed with in public.

1. The question

A member of a programming-language community asks a question. Another member consults a language model, runs the result on their own machine, confirms that it works, and posts it with an explanation. The community has a rule against AI-generated content. The answer is correct. The asker is helped.

What, precisely, has the rule protected?

This paper takes that question seriously from both directions. It is easy to answer from either side with a slogan — "slop is drowning us" or "gatekeepers hate progress" — and both slogans are partly right, which is why the argument never ends. The aim here is to separate the variables that are actually in dispute, examine what evidence exists for each, look at whether technology (watermarking and provenance signals) can settle the matter, and arrive at a position that each side could at least recognise as fair.

Three framing points first. The paper is about question-and-answer communities for programming languages and tools — forums, mailing lists, Discord and Matrix servers, Stack Exchange sites — not about academic authorship, journalism, or art, where the considerations differ. It treats "AI-generated" as a spectrum from "pasted verbatim, unread" to "drafted with a model, then tested, corrected, and explained," because almost every disagreement turns on where on that spectrum a post sits. And it is written by someone who uses these models daily and is not neutral; §8 states that interest plainly.

2. Background: how we got here

The modern version of this argument begins in December 2022, when Stack Overflow imposed a temporary ban on ChatGPT-generated answers. The moderators' stated reason is worth quoting in substance rather than caricaturing: the average rate of correct answers from the model was too low, the answers were easy to produce and looked plausible, and the volume was overwhelming the site's volunteer curation. The ban was a capacity decision, not a philosophical one.^[1]

Six months later the same site's moderators went on strike, and the trigger is instructive: a change in how the company wanted AI-generated content to be detected. Moderators argued that the detection heuristics they were being asked to abandon were what let them enforce the policy at all; the company argued the heuristics produced too many false positives against real people.^[2] Both were right, and that tension — you cannot enforce an authorship rule without a detector, and the detectors are unreliable — is the central practical problem of this paper.

The empirical picture on accuracy has been measured. A Purdue study of 517 Stack Overflow questions found that 52% of ChatGPT's answers contained incorrect information, and — the more important finding — that users preferred the model's answers a substantial fraction of the time anyway because of their comprehensiveness and articulate style.^[3] Plausibility without correctness is exactly the failure a community's trust system is least equipped to catch. Those figures are from 2023-era models; the error rate has fallen since, but the structural point survives any particular number.

Adoption, meanwhile, is not in dispute. Stack Overflow's own developer survey in 2024 found roughly three-quarters of respondents using or planning to use AI tools in their workflow, while well under half said they trusted the accuracy of the output.^[4] The community is therefore in the position of regulating a practice that most of its members engage in and most of its members distrust — which is not a contradiction, it is the problem statement.

Two further facts shape the landscape. First, the same platform that banned model answers signed a data-licensing agreement with OpenAI in 2024, which made explicit what had been implicit: the community's archive is the model's training data.^[5] Second, research on "model collapse" has shown that models trained recursively on model-generated output degrade over generations, losing the tails of the distribution first.^[6] A forum that becomes a mirror of model output is, on this evidence, eating its own seed corn.

3. The case for restriction, argued properly

The strongest version of the restrictionist case does not mention authorship at all. It has three parts.

3.1 The effort signal has collapsed

For the entire history of online help, writing an answer cost the writer more than reading it cost the reader. That asymmetry was the community's immune system. Nobody typed three paragraphs of wrong code for amusement, so the existence of a detailed answer was a weak-but-real signal that someone had engaged with the problem. Language models invert the asymmetry: the answer now costs ten seconds to produce and costs every reader the time to verify. When production is free and verification is not, the scarce resource is no longer answers but reader attention, and a community can be drowned by members who mean well. The December 2022 ban was a direct response to this, and the condition has not gone away.

3.2 Unverified output has a measured error rate

The Purdue figures above are the uncomfortable core. Even granting that current models are better, the answer that looks most authoritative in a thread is not reliably the correct one, and readers demonstrably cannot tell the difference by reading. A community that allows unverified model output at volume is, in effect, running an experiment in which its readers are the error-detection layer. That is a reasonable thing to refuse.

3.3 The community is upstream of the model

This is the argument restrictionists make least often and should make most. The model's fluency in a language is a compression of years of humans arguing about problems that were not yet documented. If the forum becomes a place where model answers are reposted, the next genuinely new question — about the library released last month, the bug nobody has hit before — finds a corpus with no new human knowledge in it, and a model trained on that corpus has nothing new to say. The model-collapse literature gives this intuition teeth. Protecting the forum as a source of novel knowledge is not nostalgia; it is protecting the thing the models themselves depend on.

3.4 What this case does and does not establish

Taken together, these three arguments establish that a community is entitled to govern unverified volume. They do not establish that a community should govern authorship. None of the three harms is caused by a model having drafted the first version of a post; each is caused by a post being unverified, or by too many posts arriving, or by the forum ceasing to add anything. A rule aimed at authorship hits these harms only by accident, and hits a great deal else.

4. The case against restriction, argued properly

The strongest version of the case against also has three parts, and it, too, does not really turn on authorship.

4.1 Provenance was never the standard

No community has ever policed whether an answer came from a search engine, a textbook, a colleague, or a memory of a bug fixed in 2019. The standard was always whether the answer was right and whether the poster would stand behind it. "No AI" quietly substitutes a different standard — how was this drafted? — and the substitution is rarely defended on its merits because it is hard to defend. An answer that was model-drafted, then run, corrected, and explained has met every requirement the community ever had of a human answerer. A rule that rejects it has changed the standard without saying so.

4.2 The rule lands hardest on those with the least

Consider who is on the asking end of most of these threads: not a senior engineer with an employer-paid subscription to the best model, but someone on an old laptop, in a place where the good tools cost a week's wages, who has been stuck for two days on something the people who could answer it in thirty seconds have not bothered to look at. The common outcome is silence, or a reply of the form "try this, try that" — guesses typed without running anything, which cost the replier nothing and cost the asker another evening. That reply was written by a human and passes every AI-content rule in existence. It is also worth less than nothing: confidently wrong, and it closes the thread.

Now the answer that was model-drafted, tested, and explained arrives — and under an authorship rule, that is the post that gets flagged. The rule, in this case, does not protect quality. It protects the comfort of those who already have access to everything at the expense of those who have the least. Any serious defence of restriction has to answer this case directly, and most do not.

4.3 The rule cannot be enforced, and the attempt does harm

An authorship rule requires an authorship detector. §5 examines the technical state of the art in detail; the summary is that detectors produce false positives against real people at rates that are unacceptable for a punitive rule, that they are biased against non-native writers of English,^[7] that their vendors have withdrawn them for low accuracy,^[8] and that the 2023 moderator strike was, at bottom, about this. A rule that can only be enforced by guessing will be enforced against the people who write in an unfamiliar style — which, in a global programming community, is a description of the people the community most needs.

4.4 What this case does and does not establish

These arguments establish that governing authorship is unjustified in principle, inequitable in effect, and impractical in enforcement. They do not establish that "anything goes." The harms in §3 are real, and a community that ignores them because authorship rules are bad will find out, as Stack Overflow did, that unverified volume is a capacity problem regardless of one's philosophy.

5. Can watermarking and provenance settle this?

If the practical obstacle to an authorship rule is that authorship cannot be reliably detected, a natural question is whether the model vendors can solve that for the communities. Two families of technique are relevant, and it is worth being precise about what each can and cannot do.

5.1 Statistical text watermarks

The dominant approach, introduced in academic form in 2023 and shipped by Google as SynthID-Text in 2024, biases the model's token choices at generation time toward a pseudo-random "green list" so that a detector holding the key can later test whether a passage has an improbably high green-token fraction.^[9][10] It is a genuine advance: the watermark is invisible to readers, does not measurably degrade quality at the settings studied, and the detection is statistical rather than a guess about style.

Its limitations, however, fall exactly where programming communities would need it to hold:

• Short, low-entropy text. The watermark's statistical power grows with the number of tokens where the model had a real choice. A twelve-line code snippet is both short and heavily constrained — there are few ways to write a correct for loop — so the signal is weak precisely in the content that answers are made of. The SynthID authors are explicit that low-entropy text limits detectability.^[10]
• Editing and paraphrase. Research in 2023 showed that paraphrasing model output substantially degrades watermark detection, and that in the limit a sufficiently capable paraphraser defeats it.^[11] The good-faith poster in §1 — who tested, corrected, and rewrote the explanation — removes the watermark in the course of doing exactly what the community wants.
• Key custody. Detection requires the vendor's key. A forum cannot check a SynthID watermark itself; it would have to send members' posts to Google, or to each vendor in turn, which is both a privacy problem and a coverage problem.
• Open-weight models. A watermark is applied at sampling time by the party running the model. Anyone running an open-weight model locally — which is the entire population of cost-constrained users in §4.2 — applies no watermark unless they choose to. Open weights are, structurally, the end of any watermark-based enforcement regime.

5.2 Provenance metadata

The second family, represented by the C2PA "content credentials" standard, attaches signed metadata to a file recording how it was made.^[12] It is the right tool for images and video distributed as files. It is nearly useless for forum text: a post is copied into a text box, not uploaded as a signed file, and the metadata does not survive the clipboard. Nothing in the current standard addresses a paragraph typed into a web form.

5.3 The regulatory push

Regulators are nonetheless leaning on these techniques. The EU AI Act's transparency article requires providers of generative systems to mark output in a machine-readable way so that it is detectable as artificially generated, with obligations phasing in from August 2026; China's labelling measures took effect in 2025.^[13][14] It is important to read these correctly: they place obligations on providers of models to mark output, not on forum members to disclose use, and they are explicitly framed around what is technically feasible. They will raise the floor for verbatim, unedited output from hosted commercial models. They will not, on the evidence above, reach edited text, short code, or locally run open-weight models — and they do not pretend to.

5.4 Assessment

Watermarking is worth having. It will make the worst case — verbatim, unread, high-volume pasting from a hosted model — somewhat more detectable, and that is the case everyone agrees is harmful. It cannot do what an authorship rule needs: distinguish the tested, edited, explained post from the lazy one, because the tested post is the one whose watermark is gone. A community that builds its rules around watermark detection will catch the careless and miss the careful, which is the opposite of the incentive it wants to create. The conclusion is that technology does not rescue the authorship rule. It confirms that the wrong variable is being measured.

6. Separating the variables

Set the two cases side by side and something becomes clear: neither actually cares about authorship. Each is built from three other variables, and those can be governed directly.

Variable	Harm if ungoverned	Can it be observed?	Rule that governs it
Verification	Wrong answers that look right (§3.2)	Yes — did it run, on what version, what was changed	Post only what you have run; state how
Disclosure	Readers calibrate trust wrongly; norms erode	Yes — the poster knows and can say	Say what tools were used and what you did with the output
Volume	Curation capacity overwhelmed (§3.1)	Yes — trivially, per account	Rate limits; stricter for new accounts
Authorship	None directly — harms arrive via the three above	No — detection unreliable, biased, defeated by editing (§5)	—

The table is the paper's central claim. Every harm either side can name is a harm of verification, disclosure, or volume. Authorship is the one variable that produces no harm on its own and the one variable that cannot be reliably observed. Rules have converged on it because it is easy to write, not because it is the thing that matters.

7. A framework communities could actually use

What follows is offered as draft guideline language, tiered so that a community can adopt the minimum or the whole.

Tier 0 — The floor (every community)

• You are responsible for what you post. Any tool may be used; 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 have not executed it, do not post it as an answer. This is the rule communities always had; it needs saying aloud now.

Tier 1 — Disclosure (recommended)

• State the tools and what you did with the output. "Drafted with a model; tested on 1.8.2; I changed the error handling" is sufficient. Disclosure is to be treated as normal and un-shameful, the way "sponsored" labels became normal. A community that punishes disclosure will get concealment instead.
• Quote, do not launder. If part of an explanation is verbatim model output, mark it as such, so readers can apply their own discount.

Tier 2 — Volume (where capacity is strained)

• Rate limits by account age and reputation, applied uniformly. This addresses the December 2022 failure mode directly and without guessing at provenance. One verified, explained answer a day is a contribution; twenty a day is a flood even if each is fine, because nobody can review them.

Tier 3 — What not to do

• Do not adopt a detector as evidence. Not a commercial classifier, not a stylistic heuristic, not a watermark check. The false-positive profile is documented, the bias is documented, and the careful poster is the one it misses.
• Do not make authorship a sanctionable offence. Sanction unverified answers, undisclosed verbatim pasting, and volume abuse — all of which are observable — and leave how the first draft was produced alone.

A community that adopts Tier 0 and Tier 1 has, in this paper's view, solved the problem it actually has. Tier 2 is for sites at scale. Tier 3 is the part most existing guidelines get wrong.

8. The author's interest, stated

The author uses language models daily, publishes writing drafted with their assistance and says so, and runs pipelines that measure models on real hardware before writing about them. That is an interest, and it cuts both ways: the author benefits from permissive rules, and the author's measured work is the kind that survives strict ones. The test this paper ends up endorsing — what did you add? — is one the author expects to be held to. If the answer is "I ran it, and here is what happened," the post belongs. If the answer is "I pasted what it said," it does not, however good the paste.

9. What would change the conclusion

A position paper should say what evidence would move it.

• A watermark that survives editing and works on short code, with detection available to forums without vendor key custody, would make authorship observable and reopen §5. No such technique exists in the literature the author has read; one that did would be a significant result.
• Evidence that disclosed, verified, model-assisted answers have a higher error rate than human answers of equal length would support restricting them on quality grounds. The Purdue study measured unverified verbatim output; the relevant comparison has not, to the author's knowledge, been run.
• Evidence that disclosure norms fail in practice — that members do not disclose, and that non-disclosure is common enough to make Tier 1 meaningless — would push toward Tier 2 volume controls as the primary tool, though not toward authorship rules, which would fail for the same reason.
• Evidence that model-assisted answering measurably reduces the forum's production of novel knowledge (§3.3) would be the strongest case for restriction, and would justify rules specifically protecting new questions — for instance, a cooling-off period before model-assisted answers on questions about recently released versions.

10. Open questions

• Who verifies the verifier? "I ran it" is a claim. Communities have always taken such claims on trust; whether that trust scales when making the claim is costless is unclear.
• Does disclosure change reader behaviour in the direction intended, or does a "drafted with AI" label simply cause good answers to be discounted?
• What is owed to the model's upstream? If the community's archive is the training data, the licensing agreements of 2024 suggest that value flows one way. Whether that changes members' obligations is a question this paper raises and does not answer.
• Where should the line sit for questions? This paper has been about answers. A model-drafted question that has not been tried is a different and possibly worse problem.

11. Conclusion

The harms that restrictionists name are real: the effort signal has collapsed, unverified model output has a measured error rate, and a forum that mirrors the model stops producing what the model needs. The objections that critics raise are equally real: provenance was never the standard, the rules land hardest on those with the least, and the detection they depend on does not work and will not be rescued by watermarking. Both sides are right about the harms and wrong about the variable. Authorship is the one thing in dispute that causes no harm on its own and cannot be observed; verification, disclosure, and volume are the three things that cause every harm named and can each be governed directly.

The question this paper began with — when a model-drafted answer is correct and helps the person who asked, what has the rule protected? — has, on this analysis, an answer. Nothing. And the person on the old laptop who was finally helped would like to know why that was the post that got flagged.

That is not an argument for no rules. It is an argument for the rules a community always had, said out loud: run it, say what you did, and do not flood the place. The conversation about how to phrase those is worth having. The conversation about how to detect who drafted a sentence is, on the evidence, one the communities cannot win, and should stop trying to.

References

1. Stack Overflow Meta (December 2022). "Temporary policy: Generative AI (e.g., ChatGPT) is banned." The moderators' statement of the reason: low average correctness, ease of production, and curation overload.
2. Stack Exchange Meta (June 2023). Moderator strike announcement and the company's response regarding AI-content detection heuristics and false-positive rates.
3. Kabir, S., Udo-Imeh, D. N., Kou, B., Zhang, T. "Is Stack Overflow Obsolete? An Empirical Study of the Characteristics of ChatGPT Answers to Stack Overflow Questions." Proceedings of CHI 2024 (preprint arXiv:2308.02312, 2023). 52% of answers incorrect; user preference driven by comprehensiveness and style.
4. Stack Overflow (2024). Developer Survey 2024 — "AI" section: adoption and trust figures.
5. Stack Overflow and OpenAI (May 2024). Joint announcement of an API partnership and data-licensing agreement.
6. Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., Gal, Y. "AI models collapse when trained on recursively generated data." Nature 631, 755–759 (2024).
7. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., Zou, J. "GPT detectors are biased against non-native English writers." Patterns 4(7), 2023.
8. OpenAI (July 2023). Withdrawal of the "AI Text Classifier" citing a low rate of accuracy.
9. Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., Goldstein, T. "A Watermark for Large Language Models." ICML 2023.
10. Dathathri, S., et al. "Scalable watermarking for identifying large language model outputs." Nature 634, 818–823 (2024). The SynthID-Text paper; see the discussion of low-entropy text.
11. Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., Feizi, S. "Can AI-Generated Text be Reliably Detected?" arXiv:2303.11156 (2023). Paraphrase attacks against watermark and classifier detection.
12. Coalition for Content Provenance and Authenticity. C2PA Technical Specification. c2pa.org.
13. Regulation (EU) 2024/1689 (the AI Act), Article 50 — transparency obligations for providers of generative systems, including machine-readable marking of output; application from August 2026.
14. Cyberspace Administration of China (2025). Measures for Labelling AI-Generated Synthetic Content, effective 1 September 2025.

Disclosure

This paper was drafted with the assistance of a language model (Claude Fable 5), at the author's direction and from the author's stated position, then argued with, edited, and signed by the author, who is responsible for every claim in it. References were checked by the author against the cited sources as of 23 August 2026. Nothing published after that date is asserted. That this disclosure appears at all is the paper's argument, practised.