ChatGPT Didn't Kill Stack Overflow. It Fired the Easy Questions.

Based on the research ofJunzhi Xue, Lizheng Wang, Jinyang Zheng, Yongjun Li, and Yong Tan, "Can ChatGPT Kill User-Generated Q&A Platforms?," Information Systems Research, 2026

A Natural Experiment Nobody Had to Design

Stack Overflow's core product, crowd-sourced programming answers, overlaps almost completely with what a general-purpose language model can produce on demand. That overlap makes ChatGPT's November 2022 launch an unusually clean shock to study. Junzhi Xue, Lizheng Wang, Jinyang Zheng, Yongjun Li, and Yong Tan treat it exactly that way, tracking question volume on Stack Overflow before and after the release and finding an average 14 percent decline, widening to nearly 28 percent by May 2023 as adoption spread. That headline number is the one most "AI is killing the internet's knowledge commons" takes lean on. It is also, on its own, the least interesting number in the paper.

What Actually Disappeared, and What Didn't

The decline was not evenly spread across the platform. It hit hardest among mid- to low-quality questions, in topics that already had deep, well-structured answer archives, and among the newest, least experienced users, exactly the population most likely to ask something ChatGPT could answer competently from its training data alone. Complex, context-dependent problems, the kind tied to a specific codebase, a specific error trace, an obscure version interaction, held up far better. Among the questions that still got asked, average quality rose, which the authors read as a genuine productivity spillover: when ChatGPT absorbs the "how do I reverse a string in Python" layer of demand, the humans who remain spend less time rephrasing basics and more time stating the actual problem well. Niche theory gives this pattern a name: not displacement but partitioning, with ChatGPT occupying the broad, shallow center of the demand distribution and Stack Overflow's surviving core retreating toward the narrow, deep edges it was always better suited to serve. That vocabulary comes from organizational ecology, where resource-partitioning theory was built to explain why specialist firms often thrive, not perish, once a market concentrates around a handful of generalists: the generalist's scale advantage covers the broad, homogeneous center of demand, but it leaves the heterogeneous edges of the resource space uncovered, and a focused competitor can occupy those edges precisely because the generalist has no incentive to chase them. Applied to Stack Overflow, the narrow, deep edges are not abstract. They are the dependency conflict that only shows up in one build pipeline, the race condition that only reproduces under a specific load pattern, the deprecated API behavior that ChatGPT's training data never captured because the question was never common enough to generate a worked example anywhere on the open web. Those are also, not incidentally, the questions a model without live access to a specific codebase structurally cannot answer well, which is exactly why they were the ones left standing.

The Platform Felt It Before the Paper Measured It

Independent tracking corroborates the direction, even where exact magnitudes are hotly contested. One widely circulated analysis of Stack Overflow's own page-view data found traffic roughly halved between April 2022 and the end of June 2023, with daily "all posts" volume falling from roughly 20,000 to roughly 7,000 a day over a similar window, a claim that went viral enough that Stack Overflow felt compelled to respond directly. The company told The Pragmatic Engineer in August 2023 that the real decline was closer to 14 percent, not 50 percent, based on data it shared with the newsletter, broadly consistent with the ISR paper's own average effect even as the company's public statements elsewhere downplayed any material impact. SimilarWeb separately recorded a double-digit drop in site visits in December 2022 alone, the month after ChatGPT's launch, though Stack Overflow attributed part of that to seasonal effects. A later Pragmatic Engineer analysis, drawing on Stack Overflow's own Data Explorer figures shared by a longtime top contributor, found that monthly question counts had fallen by 2025 to levels last seen around the site's 2009 launch, a decline that had actually started creeping in as early as 2014, years before ChatGPT existed, which complicates any story that treats the drop as a single, clean AI shock rather than a slide that AI adoption sharply accelerated. Whatever the precise attribution, the strategic response was unambiguous: Stack Overflow had already trimmed about 10 percent of staff earlier in 2023, then in October 2023 cut roughly 28 percent of its workforce in the same stretch that it shipped OverflowAI, a generative layer bolted onto the existing Q&A product rather than a replacement for it, an implicit bet that humans are still needed for the part the model cannot do alone.

Everyone Else Is Hedging the Same Bet, Differently

Other user-generated platforms facing the same functional overlap chose different trades. Reddit, rather than fighting substitution, decided to get paid for feeding it: it struck a data-licensing arrangement with Google reported at roughly 60 million dollars a year in February 2024, followed by a similar deal with OpenAI announced that May, and told investors ahead of its IPO that the aggregate contract value of its licensing arrangements ran to roughly 203 million dollars over terms of two to three years, with at least 66.4 million dollars of that revenue expected to land within 2024 alone. That is Reddit monetizing the raw material that trains the very models competing with it for attention, rather than fighting the substitution Stack Overflow is absorbing for free. Both the Google and OpenAI deals are due for renewal in 2026, which will be the first real test of whether the licensing price holds now that the novelty has worn off and both sides know what the archive is actually worth. Quora took a third path, launching Poe in December 2022 as a multi-model chat aggregator sitting on top of OpenAI's, Anthropic's, and other companies' models, rather than building something meant solely to out-answer the questions Quora itself used to host. That choice treats the underlying model layer as a commodity any vendor can supply and bets Quora's own value instead on being the neutral place users go to compare and switch between them, in effect becoming a storefront for the technology instead of racing it, a strategy only available to a platform whose core asset was always traffic and user habit rather than a proprietary answer archive. None of the three responses, Stack Overflow's bolt-on AI layer, Reddit's licensing play, or Quora's aggregator pivot, is a bet that the old model of "ask a stranger on the internet" simply keeps working as before. Each is a bet about which sliver of that old model is still worth defending.

The questions that survived the launch of ChatGPT were not the ones that mattered most to casual users. They were the ones that mattered most to experts.

The Flywheel Cuts Both Ways

The paper's authors close on a phrase worth sitting with: a self-reinforcing knowledge flywheel. Stack Overflow's archive, like Reddit's and every other forum's, is also raw training material for the next generation of models, including the one doing the substituting. If niche partitioning holds and platforms shrink toward their hard, high-expertise core, that core still has to keep producing newly solved problems, for new frameworks, new bugs, new edge cases, or the flywheel has nothing fresh left to learn from. A platform that successfully specializes down to a smaller, sharper, more expert audience, which the paper's framing suggests is already happening at Stack Overflow, is not merely surviving ChatGPT. It is quietly becoming more valuable to the next model's training run than the broad, shallow version of itself ever was, provided it survives being substantially smaller long enough to matter. The operator's job is not to stop the shrinkage. It is to make sure what remains keeps being worth asking, and worth training on.