The submitter framed the query explicitly as 'What AI did to stackoverflow in a graph,' pointing to the sharp cliff starting in late 2022 as evidence that ChatGPT's release caused developers to stop asking questions on Stack Overflow. The visualization emphasizes the post-ChatGPT collapse as the central narrative.
The editorial argues that questions peaked in 2014 and drifted downward for eight years before ChatGPT arrived, so blaming AI alone is 'emotionally satisfying and analytically lazy.' The interesting question is what a Q&A site peaking in 2014 reveals about how developers learn and why the reputation-for-answers bargain broke long before LLMs.
Commenters ran their own variants of the query — filtering out closed questions, restricting to answered questions, and slicing by individual tags. JavaScript, Python, C#, Java, and even historically stable tags like SQL and regex all show the same 70%+ decline from peak, ruling out moderation policy or a single-language story as the cause.
A Stack Exchange Data Explorer query published to Hacker News plotted monthly question volume on Stack Overflow from 2008 through mid-2026. The shape is not subtle. Questions climbed steadily through the 2010s, peaked around 2014, drifted down through the late 2010s, dropped in 2020, and then fell off a shelf starting in late 2022. By mid-2026, monthly new questions are back to volumes last seen in 2009 — the site's first full year online.
The query itself is unremarkable: a `COUNT(*)` grouped by month over the `Posts` table, filtered to `PostTypeId = 1` (questions). The chart is the point. Commenters on the HN thread ran their own variants — filtering out closed questions, filtering to answered questions, filtering to specific tags — and the shape stays the same. JavaScript, Python, C#, and Java tags all show the same curve. Even the historically resilient ones like SQL and regex are down 70%+ from peak.
The November 2022 inflection is the obvious story, and everyone in the thread names it: ChatGPT shipped, and developers stopped asking. But the more honest reading is that Stack Overflow was already in structural decline. Questions peaked in 2014. The 2015-2022 slope was gentle but consistent. ChatGPT didn't start the fire — it poured accelerant on one that had been smoldering for eight years.
The framing that dominates most discussions of this data — "AI killed Stack Overflow" — is emotionally satisfying and analytically lazy. The more interesting question is what a Q&A site that peaked in 2014 tells you about how developers actually learn, and why the model broke before LLMs arrived.
Stack Overflow's original bargain was: strangers on the internet will answer your question in exchange for reputation points, and everyone else who Googles the same problem later gets the answer for free. That bargain worked spectacularly for the class of question that has one correct answer, is asked frequently, and is annoying to look up in docs. "How do I flatten a list in Python?" "Why is my CORS header being ignored?" "What does `error: expected primary-expression` mean?" The site's canonical answers to these questions have been viewed tens of millions of times. That's real value, delivered at zero marginal cost.
But somewhere around 2015 the moderation culture hardened. Duplicate-closing, low-quality-closing, and off-topic-closing all ramped up. The community's stated goal was signal preservation. The practical effect was that asking a novice question became socially expensive. Several HN commenters revisit this: a whole generation of developers learned to Google Stack Overflow but never learned to *post* on it, because posting meant risking a public downvote from someone with 200k rep. The site trained users to consume its content and punished them for contributing to it — which worked as long as the existing corpus stayed relevant, and stopped working the moment a competing interface offered answers without the social tax.
ChatGPT — and later Claude, Copilot, Cursor, and the rest — offered exactly that competing interface. You type your broken CORS config into a chat box and get an answer in three seconds, with no risk of a mod closing your question as a duplicate of something from 2013 that doesn't quite match your situation. The models are wrong sometimes. Stack Overflow is also wrong sometimes, and the highest-voted answer is often a decade out of date. For a working developer trying to unblock themselves in the next 90 seconds, the tradeoff isn't close.
The part nobody in the thread wants to sit with: the models are answering your CORS question using Stack Overflow's corpus. Every major coding model was trained on scraped Stack Overflow data. The site is, in a very real sense, subsidizing the tool that's replacing it. And if the input flow of new questions has collapsed to 2009 levels, the corpus is going stale in real time. The tacit knowledge that trained GPT-4 and Claude 3 on how to debug a webpack config in 2021 isn't being refreshed for how to debug a Vite 6 config in 2026, because nobody's asking the question in a place the models can learn from.
First, stop assuming the docs are the fallback. For a lot of libraries — especially the ones with high churn and low corporate backing — the *real* documentation was Stack Overflow's top answers, curated over a decade by people who were tired of answering the same thing twice. That layer is not being replenished. If you maintain an open-source project, the odds that new user questions land somewhere your future contributors can find them are lower than they were three years ago. GitHub Discussions, Discord servers, and vendored LLM sessions all fragment the knowledge instead of aggregating it.
Second, if you're building tooling that depends on Stack Overflow scrapes — training data, semantic search over developer Q&A, RAG pipelines — assume the corpus is frozen at roughly its 2022 state, with a long tail of low-volume additions. Anything post-2023 is thin. Anything post-2024 is close to nonexistent for many stacks. Vite 5 → Vite 6, React 18 → React 19, Node 20 → Node 22 — the transitions are barely represented. Your model's confident-sounding answer about a 2026 API is very possibly hallucinated from an older version.
Third, be careful what you replace it with. Discord communities are searchable in theory and unsearchable in practice. GitHub Discussions are gated behind a repo you have to know about. The good answers are increasingly locked inside private LLM conversations that never get published. The public knowledge commons that made the last 15 years of software development possible is being quietly converted into private query-response transactions, and it's not clear who's building the replacement.
Stack Overflow the company will survive — the archive is still valuable, and the enterprise team-knowledge product has customers. But Stack Overflow the community-driven Q&A site is not coming back. The graph doesn't bend. The interesting question over the next two years isn't whether the model shifts, but what fills the vacuum: whether the LLM vendors start hosting their own public Q&A layers to keep training data flowing, whether a new pseudonymous forum culture emerges around AI-assisted debugging, or whether we all just quietly accept that the collective debugging knowledge of the 2020s lives in nobody's archive.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.