Stop Defending the Questions ChatGPT Was Always Going to Take

Based on the research ofXue, Wang, Zheng, Li and Tan, "Can ChatGPT Kill User-Generated Q&A Platforms?," Information Systems Research, 2026

When a technology can answer the same question a platform was built to answer, the intuitive fear is total: the machine substitutes for the community, the community empties out, the platform dies. Junzhi Xue, Lizheng Wang, Jinyang Zheng, Yongjun Li, and Yong Tan take that fear seriously and then dismantle it. Using Stack Overflow and the release of ChatGPT as a natural experiment, they show the substitution is real but selective. Question volume falls by roughly 14% on average, and by as much as 27.9% over time, but the decline is concentrated exactly where you would want traffic to leave: mid- to low-quality content, topics with rich and well-structured existing knowledge bases, and less-experienced users. The complex, context-dependent problems stay. Read through a niche-theory lens, this is not displacement. It is partitioning, and the half the platform keeps is the valuable one.

The 14% was never really yours

The counterintuitive core of the paper is that the migrating questions were the platform's weakest asset. A question whose answer already sits in a well-documented, heavily-trafficked corner of the knowledge base contributes almost nothing new when it is asked for the thousandth time. It consumes moderator attention, clutters search, and irritates the experienced contributors who have answered it before. When ChatGPT absorbs that query instead, the platform loses a page view and sheds a liability. The authors find the substitution is strongest precisely in topics with deeper accumulated answer-side knowledge, which is to say the questions most fully solved already are the ones most cleanly handed off. What remains is the genuinely hard, novel, situation-specific problem that a language model, trained on yesterday's corpus, cannot reliably resolve.

There is a second effect that the panic narrative misses entirely. The paper documents a direct improvement in the quality of the questions that do get asked, a positive spillover from lower search and articulation costs. When the routine askers route around the platform, and when the remaining askers can use a model to sharpen a vague problem into a precise one before posting, the average question that reaches the community gets better. The platform is not just losing its worst traffic. It is upgrading the traffic it keeps.

The managerial reflex, though, runs the other way. Faced with a 14% volume drop, most operators reach for the retention playbook: gamified prompts to ask more questions, SEO to recapture the how-do-I-reverse-a-string searches, engagement metrics that treat every lost page view as a wound. Every one of those moves spends scarce resources re-acquiring the lowest-value activity on the platform. The paper's logic says the opposite is correct. Let the low end go, and pour the freed capacity into ranking, incentives, and moderation that reward high-value contributions and protect the experienced users who answer the hard questions.

Stack Overflow already ran the experiment in public

The reason this reads as counterintuitive is that Stack Overflow's headline numbers look apocalyptic, and the trade press has treated them as a death notice. Similarweb measured a 13.9% year-over-year traffic decline at Stack Overflow in March 2023, only months after ChatGPT's launch, part of a slide it clocked at roughly 6% a month since the start of 2022. By May 2025, as The Pragmatic Engineer documented, the monthly volume of new questions had fallen to levels last seen when the site launched in 2009. On its face this is the community hollowing out. Read through the paper, it is the low end draining away on schedule, and the question is whether management responds by defending the drain or by fortifying what is left.

Stack Overflow's early instinct was to defend. In December 2022, within weeks of ChatGPT's release, its moderators imposed a temporary ban on AI-generated answers, calling them "substantially harmful" to the site because they were plausible-looking and frequently wrong. As a content-integrity measure that was defensible; a flood of confident, incorrect machine answers is a genuine threat to a platform whose entire value is vetted correctness. But as a competitive posture, banning the technology does nothing about the underlying substitution. Users were not posting ChatGPT answers so much as they were quietly not asking the questions at all, and no moderation policy reaches a question that never gets typed.

The machine takes the questions you were tired of answering. The fight worth having is over the ones it cannot answer at all.

The more telling move came later, and it points toward the paper's real prescription. The same knowledge base that a chatbot substitutes for is also the raw material the chatbot was trained on, and that dependency is leverage. The authors describe a self-reinforcing knowledge flywheel between LLMs and platforms: the model needs the community's accumulated, human-vetted answers as an input for future learning, even as it siphons off the community's routine traffic. A platform that grasps this stops treating the model purely as a predator and starts treating it as a customer.

Sell the archive, defend the frontier

That is precisely the turn Stack Overflow made. In May 2024 it announced an API partnership with OpenAI, licensing its OverflowAPI so that OpenAI could use Stack Overflow's vetted technical content to improve model performance and surface attributed answers inside ChatGPT. The company that had banned the chatbot eighteen months earlier was now selling it access to the archive, capturing value as an input to the model rather than being hollowed out as its victim. The flywheel, monetized.

Stack Overflow was not alone in reading the moment this way. Reddit signed a content-licensing deal with Google in February 2024, reported at roughly $60 million a year, granting the search company access to Reddit's corpus for AI training, and the arrangement made Reddit one of the reference cases for the emerging market in human-generated training data. The strategic content is identical across both: a platform sitting on years of user-generated knowledge stops trying to wall that knowledge off from the models and starts charging for it, converting an existential threat into a licensing line item. It is worth naming the tension honestly, because it is real. Reddit's own leadership has since questioned whether the Google deal still serves it as AI-generated search answers cut into the traffic that referral links used to bring, and by mid-2026 the two sides were reportedly in tense renewal talks. Licensing the archive funds the platform, but if it accelerates the substitution of the platform's own front door, the flywheel can spin the wrong way.

Which is why licensing is only half the prescription, and the less important half. The durable asset is not the static archive you can sell once; it is the living capacity to generate answers the archive does not yet contain. A model trained on every existing Stack Overflow answer still cannot resolve the bug in a library released last week, the error specific to one team's undocumented configuration, or the design tradeoff that has never been argued out in public before. Those are the questions the paper finds stubbornly resistant to substitution, and they are generated by exactly the experienced contributors that a retention-obsessed, low-end-defending strategy tends to neglect. Defend the frontier, sell the archive, and never confuse the two.

What the model cannot answer for you

The forward-looking question for any knowledge platform is therefore not "how do we win back the questions ChatGPT took," but "what do we still do that a language model structurally cannot." The paper's niche-partitioning result draws the boundary. On the model's side of the line sit the routine, the well-documented, and the already-solved, questions where the answer is a retrieval problem. On the platform's side sit three things a model cannot manufacture. The first is novelty: problems that postdate or fall outside the training corpus, where a machine confabulates and a human who has hit the same wall does not. The second is community, the reputational and social fabric that makes an expert willing to write a careful answer to a stranger's obscure problem for status rather than pay. The third is trust, the human vetting and adjudication that let a reader believe an answer is correct rather than merely fluent, the exact property whose absence drove Stack Overflow's original AI-answer ban.

A platform that organizes itself around those three things treats the volume decline not as a wound to be stanched but as a filter to be embraced. It reallocates moderation away from policing duplicate beginner questions and toward curating the hard ones. It rebuilds its reputation and incentive systems to reward the contributors who answer what the model cannot, rather than the ones who farm points on questions the model has already absorbed. It licenses its archive to the model developers on terms that fund the community, while guarding the freshness and the trust that make the archive worth licensing in the first place. And it measures its health not by raw question count, a metric now permanently deflated by the migration of the low end, but by the depth and novelty of the questions it still uniquely resolves.

The blunt version, for any operator watching a chatbot eat into the numbers: the traffic you are losing is the traffic you were overpaying to keep. Stop defending it. The platform that survives the language-model era is not the one that clung hardest to the routine question. It is the one that recognized the routine question was never where its value lived, and spent the reprieve fortifying the frontier the model will keep failing to reach.