Your Questions Are Not Getting Lazier Because of AI. They Are Getting Harder.

Based on the research ofRao, Shan, Rivera, Qiu and Wang, "How Generative AI Transforms Questioning Behavior on Q&A Platforms: Evidence from a Natural Experiment with Pilot Usage of ChatGPT," MIS Quarterly, 2026

The fear: an answer machine should empty the question box

The intuition behind "AI will kill the forum" is simple enough to feel obvious. If a user can get a correct-looking answer from a chatbot in three seconds, why would they spend ten minutes writing up a Stack Overflow question, waiting hours for a human to notice it, and risking a downvote or a "duplicate" close from a moderator along the way? Every incentive point toward substitution: AI replaces the ask, not just the answer. That is the story told in most of the post-ChatGPT retrospectives on developer forums, and it is not wrong about the long run. Stack Overflow's own public question-volume data, tracked independently by engineers who pull it from the Stack Exchange Data Explorer, shows a rapid decline setting in right around November 2022, the month ChatGPT launched, compounding a longer slide that had already started years earlier. By early 2026 the platform was posting a few thousand new questions a month, a small fraction of its 2014 peak of roughly 200,000 a month, according to reporting on the site's continuing collapse.

So the paper by Rao, Shan, Rivera, Qiu, and Wang is not fighting that long-run decline, and it is not arguing Stack Overflow is thriving. It is doing something narrower and more useful: isolating a specific, short natural-experiment window, and asking what generative-AI answers did to human questioning behavior while the AI answers were still a legitimate, permitted part of the platform, before the ban removed that option and before the broader multi-year decline overtook everything. In that narrow window, the quantity-destroys-quality story does not hold up.

The natural experiment: a pilot, then a ban, inside one platform

The setting is almost a textbook natural experiment. In the weeks after ChatGPT's public release, Stack Overflow users could and did post AI-generated answers, since no rule yet forbade it. Then, on December 5, 2022, Stack Overflow's moderators posted a sitewide policy: "Overall, because the average rate of getting correct answers from ChatGPT and other generative AI technologies is too low, the posting of content created by ChatGPT and other generative AI technologies is substantially harmful to the site and to users who are asking questions and looking for correct answers." The moderators called it a temporary measure at the time, and the underlying ban on AI-written content has since become Stack Overflow's standing policy.

That abrupt, dated, platform-wide cutoff is what makes the paper's difference-in-differences design possible. Users who had been exposed to ChatGPT-generated answers in their communities form a treated group; users in corners of the platform less touched by the pilot form a comparison group; the ban date gives a clean "after." The authors track four concrete outcomes around that boundary: the number of questions a user posts per week, how long each question is, how novel its content is relative to what has been asked before, and how many upvotes it collects.

What actually happened: more questions, not fewer, and better ones

The result runs against the substitution story point for point. Exposure to ChatGPT-generated answers is associated with an increase in the number of questions users post, and those additional questions are longer, more novel, and receive more upvotes than before. Users with access to AI-generated answers did not quietly stop asking. They asked more, and they asked better.

The paper's authors test two explanations for why, and both point the same direction. The first is a complexity shift: ChatGPT absorbs the routine, previously-answered-a-thousand-times questions, the "how do I reverse a string in Python" tier of query, leaving the human community to field what is left over, which skews toward the genuinely hard and unresolved. The second is a skill-transfer effect: users learn from watching AI-generated answers how to frame a clearer, more precise question, and they carry that skill into the questions they ask on the platform. Both mechanisms predict exactly the pattern in the data, longer and more novel questions, without requiring that users ask more out of pure habit or boredom.

The paper also finds the effect is not uniform. It is weaker among already-experienced answerers, users who did not need an AI assist to write a crisp question in the first place, and stronger among high-reputation users, the people most invested in the platform's status economy and most likely to redirect freed-up effort into a harder, more upvote-worthy question rather than walking away entirely.

The chatbot did not make the questions disappear. It made the easy ones disappear, and left the hard ones standing out.

The counterargument the paper takes seriously: reputation, not just curiosity

It would be convenient to read this as proof that AI makes everyone a better, more curious learner. The paper is more careful than that, and the moderator effect it finds is a useful check on over-claiming. The questioning boost is strongest among high-reputation users, the people with the most to gain from Stack Overflow's badge-and-score status economy, not among casual users with the least platform history. That is consistent with a less flattering but equally real reading: freed from spending effort on routine questions, status-motivated users reinvest that freed effort into asking the kind of question that builds reputation, a harder, more novel, more upvotable question. The behavior change is real and measurable either way; the paper's data cannot fully separate "users became better thinkers" from "users had more bandwidth to chase status," and both stories point toward more and better questions rather than fewer.

That distinction matters operationally. If the lift comes from status-seeking, it will only persist as long as the platform's reputation system still rewards human-authored, hard questions, which is exactly the kind of incentive Stack Overflow's later full ban on AI-written content was designed to protect.

What this means if you run a knowledge platform

The lesson is not "generative AI is safe for your community," because the long-run Stack Overflow numbers say plainly that it is not, whatever the cause. The lesson is narrower and more actionable: the mechanism by which AI answers interact with human questions is not simple substitution, and treating "will usage go up or down" as the only question you ask will miss the more important one, which is what kind of usage survives. A platform that lets AI handle the bottom of its question distribution while protecting the incentive to ask something genuinely novel at the top can see engagement concentrate rather than evaporate, at least for a while and at least among its most committed users. A platform that lets AI-generated content substitute for human-authored answers across the board, with no reputation system left to reward the harder question, has no mechanism left to protect, and the traffic data from 2023 onward is what that looks like once the forcing function is gone.

Before writing off a question-asking community because an AI tool now answers the easy version of its questions, check which users are still showing up, and what they are asking now that they used to skip.