Stack Overflow Is a Ghost Town. The Company Is Doing Fine.

When was the last time you actually opened Stack Overflow? Not Googled into it, not scraped an old answer — actually went there to ask something. For two decades it was the developer’s Wikipedia. Then ChatGPT showed up, and the lights started going out.

A Forum Without Questions

The inflection point is almost too clean to be a coincidence. New questions on Stack Overflow started dropping sharply in November 2022 — the same month ChatGPT launched. By 2024, monthly new questions were down roughly 75% from pre-ChatGPT levels. The site is now posting question volumes last seen in 2008, back when it was still a scrappy beta.

What’s left isn’t just smaller. It’s different. The easy questions are gone, the beginners are gone, and the community energy that came with them is gone too. What remains is the residue: gnarly edge cases ChatGPT couldn’t solve, obscure version conflicts, weird interactions between libraries nobody’s documented. Stack Overflow has quietly become the place you go when the AI gives up.

Why People Left

The reason isn’t accuracy. ChatGPT hallucinates plenty. The reason is that ChatGPT is nice to you.

Stack Overflow’s culture was its original sin. Ask a beginner question and you got “marked as duplicate,” “off-topic,” or the classic RTFM treatment. The site optimized for a curated knowledge base, not for the person on the other end of the question. Generations of developers learned to fear posting there.

ChatGPT inverted that experience overnight. You can ask anything. You can ask it badly. You can ask it three times in a row. It explains line by line, never sighs, never closes your thread. As one developer put it on Hacker News: Stack Overflow made you feel stupid; ChatGPT makes you feel smart. For most developers, that’s the whole game.

So Why Is the Company Still Here?

Here’s the twist. The forum is hollowing out, but Stack Overflow the company is doing fine — arguably better than it was during the forum’s golden years. Two reasons.

First, data licensing. In 2024, Stack Overflow cut deals with OpenAI and Google to license its Q&A archive for AI training. The irony writes itself: the product that killed Stack Overflow’s traffic was trained on Stack Overflow’s data, and now Stack Overflow gets paid every time a new model wants more of it. The corpse pays rent.

Second, Stack Overflow for Teams. The enterprise version — a private, in-house Q&A platform — is humming along at Google, Microsoft, Bloomberg, and others. Companies still need somewhere to capture institutional knowledge that ChatGPT doesn’t know: internal codebases, undocumented services, the reason that one Kafka cluster is configured weirdly. The public forum is dying. The private forums are billable.

The Real Problem Isn’t the Site. It’s the Pipeline.

Here’s where it gets uncomfortable. LLMs are trained on human-generated content. Stack Overflow was one of the richest, highest-signal sources of that content for code. If that well runs dry — if developers stop asking each other questions in public — what does the next generation of AI learn from?

New languages will ship. New frameworks will ship. New bugs will surface in libraries that don’t exist yet. Historically, humans argued about these things in public, converged on answers, and that corpus became training data. Strip out the humans, and you get AI trained on the output of older AI, drifting further from ground truth with every generation.

Researchers already have a name for this: model collapse. The hypothesis is that successive training on synthetic data accumulates subtle errors until diversity and accuracy both degrade. Stack Overflow’s decline may be the clearest field evidence we have that the loop is starting to close.

The Takeaway

Stack Overflow isn’t really a story about one website. It’s a story about what happens when an AI tool eats the ecosystem that made it possible in the first place. The public square where developers compared notes is emptying out, and its archive has been digested into model weights.

So a question worth sitting with: where did you ask your last hard coding question, and where do you think you’ll ask it five years from now? The honest answer might decide what kind of internet we end up with.