The editorial argues Stack Overflow's public Q&A site is now closer to a museum exhibit than a product, with question volume down ~90% from its 2014 peak. Prosus has pivoted hard into data licensing deals with Google and OpenAI, monetizing the 58-million-question archive plus the Teams B2B product — meaning the forum's death doesn't threaten the company's profitability.
The submitted Sherwood News piece frames Stack Overflow as a forum that has collapsed under AI competition while the parent company keeps the lights on through licensing deals. The framing emphasizes the irony that the same AI tools draining the forum's user base are now the company's primary customers for its historical data.
The editorial argues the more interesting story isn't 'AI killed Stack Overflow' but that frontier code models — GPT-5, Claude, Gemini, Llama derivatives — were all trained on Stack Overflow scrapes captured at the site's peak. That corpus of human-graded, accepted-answer-marked code reasoning has no equivalent replacement, and now that humans have stopped asking and answering new questions, the feeding tube for future model training is being pinched off.
The editorial notes that moderators staged a strike in 2023 over the company's AI content policies, signaling that the community side has not gone quietly. The core human-expert proposition — ask a question, get an answer in minutes — has been eaten by tools that respond in seconds and don't close questions as duplicates, suggesting moderation culture itself contributed to the exodus.
The editorial points out that while Stack Overflow itself is functionally a ghost town, pockets of the broader Stack Exchange network — Mathematics, TeX, Server Fault — remain active. This suggests AI substitution hits hardest in domains where LLMs are most competent (general programming) and leaves specialized expert communities relatively intact.
Stack Overflow, the site that defined how a generation of programmers learned to debug, is functionally a ghost town. The Sherwood News piece picked up on Hacker News this week put hard numbers on something practitioners have felt for two years: question volume is down roughly 90% from its 2014 peak, and the decline accelerated sharply after ChatGPT launched in late 2022. Traffic followed. The forum that once handled millions of monthly questions now sees a trickle.
The interesting twist is that the company is still profitable — just not from the forum. Prosus, which acquired Stack Overflow for $1.8 billion in 2021, has pivoted hard into data licensing. The deals with Google (announced 2024) and OpenAI give the AI labs structured access to the 58-million-question archive: tagged, voted, accepted-answer-marked, and human-curated in a way no other code corpus on the open web is. Stack Overflow Teams, the private B2B version, also keeps the lights on. The public Q&A site is now closer to a museum exhibit than a product.
The community side has not gone quietly. Moderators staged a strike in 2023 over the company's AI content policies. The Stack Exchange network outside the flagship site still has pockets of life (Mathematics, TeX, Server Fault). But the core proposition — ask a question, get an answer from a human expert within minutes — has been eaten by tools that respond in seconds and don't close your question as a duplicate.
The surface-level read is "AI killed Stack Overflow," and that's true but boring. The more interesting story is that the same AI that killed it depends on it to function — and the feeding tube is being pinched off.
Every frontier code model — GPT-5, Claude, Gemini, the Llama derivatives — was trained on Stack Overflow scrapes that captured the site at its peak. That corpus is the single largest source of high-quality, human-graded, accepted-answer-marked code reasoning on the internet. GitHub gives you code; Stack Overflow gives you code *plus* the explanation of why it works, *plus* the explanation of why the three other approaches don't, *plus* a vote from thousands of practitioners on which is best. Nothing else has that shape.
Now the well is drying. New language features (anything post-2023), new framework versions, new error messages from new tools — these used to generate hundreds of Stack Overflow threads within weeks of release. They don't anymore. The questions are being asked, but they're being asked in private to LLMs, and the answers (correct or not) leave no public trace. This is a classic feedback-loop failure: the AI consumes the substrate it grew on, and the substrate doesn't regenerate. Derek Lowe would call this a catalyst-inhibitor relationship; SREs would call it cache poisoning by way of starvation.
The community reaction on the HN thread split predictably. One camp argues this is fine — LLMs are strictly better than searching SO, the duplicates-and-downvotes culture was hostile, good riddance. The other camp points out that LLMs are confidently wrong about new APIs precisely *because* there's no public corrective signal anymore. Both are right. The first camp is describing the present; the second is describing the trajectory.
Licensing the archive to AI labs is a smart Q4 move and a terrible decade move. It monetizes a depreciating asset. Each year, the archive is a smaller fraction of the universe of code knowledge it once covered. Prosus is selling a snapshot of 2014-2022 programming knowledge to companies that need it to keep working *now*. The deal makes sense for Stack Overflow's shareholders and is structurally unsound for the AI industry that just signed the check.
First, the practical: if you're shipping code that uses any framework released after 2023, treat your AI assistant's confidence as inversely proportional to the recency of the API. Claude will lie about Next.js 15 features with the same tone it uses to correctly explain Express middleware from 2016. The difference is what training data existed. Build a habit of cross-checking against the actual library's GitHub issues, changelog, and (where they still exist) the smaller niche forums where the actual library maintainers hang out — Discord servers, GitHub Discussions, Bluesky, the dying embers of subreddits.
Second, if you maintain a library or framework, your documentation just got promoted. It is now the only authoritative source of truth your users' AI assistants can be retrieved against. Invest in it accordingly. Write the FAQ entries. Document the gotchas inline. Tag releases with detailed migration notes. The world where users could Stack Overflow their way through your weird API quirks is over; you eat the support burden directly through docs or through GitHub issues, and docs scale better.
Third, and this is the one nobody is acting on yet: the company that figures out how to capture and curate the millions of private Q&A interactions happening inside ChatGPT and Copilot today will own the next Stack Overflow-shaped data moat. That's the structural opening. Discord servers, GitHub Discussions, and the AI labs themselves are all sitting on it. Whoever extracts a high-signal public corpus from those private conversations — with attribution, voting, and verification — wins the next training-data round.
Stack Overflow the forum isn't coming back. The duplicates-closing, rep-farming, accept-the-answer culture was killed by a faster, more patient substitute, and you can't out-friendly an LLM. But Stack Overflow the *dataset* is going to keep getting more valuable in absolute dollar terms even as it gets less complete in relative terms — which is exactly the kind of unstable equilibrium that ends in either a price spike, a new entrant, or a quiet realization across the AI industry that they need to start funding the human Q&A layer they hollowed out. Watch for the first AI lab to pay developers directly to answer questions in public. That's the tell that someone has done the math on the substrate problem.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.