A single SQL query on the Stack Exchange Data Explorer, plotted as a graph, became the most-discussed post on Hacker News this weekend: 273 points and 319 comments in seven hours. The query counts questions asked per month on Stack Overflow since July 2008. The graph shows the site’s question volume has fallen back to levels last seen when Stack Overflow launched. The argument in the comments is not about whether the decline happened — it’s about who killed it.
What is Stack Overflow? It’s the website where programmers go to ask coding questions and get answers from other humans. For 15+ years it was the default first stop when you got stuck — you’d search your problem, and someone had usually already asked the same question and gotten an answer. It was a free, public, community-built knowledge base. This article is about a graph that went viral showing that Stack Overflow is dying — and the debate about why.
🔍 THE BOTTOM LINE: Stack Overflow is not dead, but it is no longer where developers go to get unstuck. The peer-reviewed estimate is a 25% drop attributable to ChatGPT alone in the first six months, but the longer arc shows decline starting in 2014 — eight years before ChatGPT. AI accelerated a fall that the site’s own moderation culture started. The open question for the open web is what replaces the knowledge commons that Stack Overflow’s Q&A used to generate for free.
What the Graph Shows
The Data Explorer query is one line of SQL: count questions grouped by month, ordered ascending. The output is 217 rows starting July 2008 with 4 questions, climbing to a peak around 2014, plateauing, spiking during the pandemic lockdown in March 2020, and then declining steadily from mid-2020 onward. The post-ChatGPT decline (from November 2022) is visible but is a steepening of an existing slope, not a new direction.
The Pragmatic Engineer’s May 2025 analysis put numbers to the same data: by May 2025, monthly questions were “as low as when Stack Overflow launched in 2009.” Marc Gravell, a top-10 all-time contributor, shared the graph that became the basis for that piece. The pattern is a classic bell curve — slow build, peak, long decline — with ChatGPT’s release date marked as the point where the decline accelerated sharply.
The Commentariat’s Verdict: It Wasn’t Just AI
The 319 comments on the HN thread are remarkably consistent on one point: blaming ChatGPT alone is tunnel vision. The top-rated comments, in order:
- “Interesting that you can see COVID in the graph.”
- “Except for covid, it seems the decline was already there.”
- “Stackoverflow did it to themselves by having incredibly unhelpful users.”
- “AI and ridiculously aggressive moderation. If it had been a more welcoming place it probably would have lasted longer.”
- “SO did that all to themselves when they decided they didn’t want a community to form and that only question and answers mattered. The moment something else allowed to have a better way to get your answers, there was no reason to go there, because there was no community.”
- “Seing a bell curve and singling out a factor that appears only for the 15% of the total time demonstrates some pretty extreme tunnel vision.”
The through-line: Stack Overflow’s moderation culture — closing questions as off-topic, downvoting beginners, treating duplicates as crimes — hollowed out the community years before ChatGPT existed. ChatGPT just provided a polite, always-available alternative that didn’t make you feel stupid for asking.
One comment that got traction described the workplace version of the same shift: “The company my friend works for has a slack channel for help with code, like an internal stackoverflow. It’s almost inactive now. I asked to see one of the questions from 2024 — it could have been solved with one LLM search. We have eliminated a whole genre of peer to peer communication.”
The Peer-Reviewed Estimate: 25% From ChatGPT Alone
The peer-reviewed paper by Maria del Rio-Chanona and colleagues, published in PNAS, put a number on the ChatGPT-specific effect. Using a difference-in-differences model against Russian and Chinese programming Q&A sites where ChatGPT access was limited, they estimated a 25% drop in Stack Overflow posting activity attributable to ChatGPT in the first six months after its November 2022 release. They explicitly interpret this as a lower bound.
The same paper found the decline was larger for posts related to the most widely-used programming languages — the ones LLMs are most knowledgeable about — and that the effect was not concentrated among new or low-quality users. “We find no significant change in post quality, measured by peer feedback, and observe similar decreases in content creation by more and less experienced users alike. Thus, LLMs are not only displacing duplicate, low-quality, or beginner-level content.”
That finding matters. The naive defense of AI substitution is “it only kills the junk questions.” The data says it’s killing the good ones too.
The Deeper Problem: The Training Data Loop
The PNAS paper raises the point that should worry anyone building AI systems: “This substitution can present a significant problem for the training data needed to develop future models if it leads to a reduction of human-generated content.” Stack Overflow’s Q&A archive is a major component of the training corpora for coding models. As those models improve, they reduce the incentive to post on Stack Overflow, which reduces the flow of new human-generated coding knowledge into the public corpus, which weakens the training data for the next generation of models.
This is a slow-motion version of the model collapse problem. It doesn’t happen in a single training run — it happens across the ecosystem, as the open web’s knowledge production infrastructure atrophies. The Pragmatic Engineer piece frames the same concern: “interactions with AI models are not added to the shared pool of online knowledge.”
Stack Overflow’s content is under a CC-BY-SA license. Your private ChatGPT conversation is not. Every question that moves from public Q&A to private LLM chat is a piece of the commons that doesn’t come back.
The Timeline, Annotated
The Pragmatic Engineer’s annotated timeline is the clearest summary of what happened and when:
- 2014: Questions started to decline. Stack Overflow significantly improved moderator efficiency — questions closed faster, more closed, “low quality” removed more aggressively. “This tallies with my memory of feeling that site moderators had gone on a power trip by closing legitimate questions. I stopped asking questions around this time because the site felt unwelcome.”
- March 2020: Big jump in traffic. Pandemic lockdowns, forced remote working — instead of asking colleagues, developers Googled and visited Stack Overflow.
- June 2020: Questions start to decline again, faster than before. Two and a half years before ChatGPT.
- November 2022: ChatGPT launches. Questions decline rapidly. “ChatGPT is faster and it’s trained on StackOverflow data, so the quality of answers is similar. Plus, ChatGPT is polite and answers all questions, in contrast to StackOverflow moderators.”
- 2025–2026: Monthly questions at 2009 launch levels. The site is still operational but no longer the default first stop for a stuck developer.
NZ Angle
New Zealand’s developer community is small enough that the Stack Overflow decline has a specific local shape. NZ dev Slack channels, Discords, and the long-running NZ Tech mailing lists were already the first stop for many Kiwi developers before ChatGPT — Stack Overflow’s moderation culture was particularly hostile to the kind of “I’m new to this stack and the docs are unclear” question that junior devs need answered. ChatGPT has replaced both: the hostile public Q&A and the friendly-but-slow private channel.
The NZ-specific worry is the same as the global one, just sharper: where does the next generation of NZ developer knowledge come from? The seniors learned by answering on Stack Overflow. The juniors ask ChatGPT and never build the public artifact. The compounding knowledge base that used to be a public good is becoming a private one, owned by whichever model provider the junior happens to use.
For NZ’s growing AI-coding-tools sector — the Cursor-clones, the agent frameworks, the local-model dev tools — this is both an opportunity (the demand for private, local knowledge generation is rising) and a constraint (the training data those tools need is thinning out at the source).
❓ FAQ
Is Stack Overflow shutting down? No official announcement. The Pragmatic Engineer’s read is that the question is “when Stack Overflow will wind down operations, or the owner sells the site for comparative pennies, not if it will happen.” Stack Overflow was acquired by Prosus in 2021 for $1.8 billion. The asset is shrinking.
Did ChatGPT kill Stack Overflow? Partially. The peer-reviewed estimate is 25% of the drop in the first six months is attributable to ChatGPT, interpreted as a lower bound. The longer decline started in 2014 and is attributable to moderation culture. ChatGPT accelerated a fall that was already underway.
Why does this matter if I’m not a developer? Stack Overflow was the largest functioning knowledge commons on the open web — a CC-BY-SA-licensed archive of human problem-solving at scale. Its decline is the case study for what happens to public knowledge infrastructure when a private, non-copyleft substitute arrives. Wikipedia, open textbooks, public scientific datasets — all face versions of the same substitution risk as LLM coverage of their domains improves.
What’s the fix? There isn’t a clean one. Possible directions: (1) license requirements on AI outputs derived from copyleft corpora, (2) public investment in Q&A infrastructure that doesn’t depend on volunteer labor and aggressive moderation, (3) model providers paying into a fund that finances open knowledge production, (4) accept the decline and build new commons on different foundations. None of these are currently being pursued at scale.
Where did the graph come from? A public Stack Exchange Data Explorer query — one line of SQL, anyone can run it. The Data Explorer is one of the few parts of the Stack Exchange ecosystem that remains actively useful.
🔍 THE BOTTOM LINE: A single SQL query and a graph lit up 319 comments because it visualizes something developers have been feeling for years: the place they used to go for help is emptying out, and the thing that replaced it is private, polite, and doesn’t give anything back to the commons. The 25% ChatGPT number is real and peer-reviewed. The longer decline is real and self-inflicted. The training-data feedback loop is the part nobody has a fix for yet. Read this alongside our earlier piece on what makes a developer when Claude writes the code — the Stack Overflow graph is the infrastructure-side answer to the same question.
📰 Sources
- Stack Exchange Data Explorer, Query 1953768 (stackoverflow.com)
- Hacker News discussion, “What AI did to stackoverflow in a graph” (48956949)
- The Pragmatic Engineer, “Stack overflow is almost dead” (May 2025)
- Maria del Rio-Chanona et al., “Large language models reduce public knowledge sharing on online Q&A platforms” (PNAS)
- Singularity.kiwi prior coverage: If Claude writes the code, what makes a developer?, AI coding agents cost more than developers
