AI Ate the Well It Drank From. Stack Overflow Just Showed Us the Receipt.

Stack Overflow's monthly question volume has fallen roughly 75% — from peaks above 200,000 a month down under 50,000, erasing about fifteen years of growth. The forum that taught a generation of developers how to fix things is, functionally, a ghost town. AI killed it: why wait for a stranger to answer when Claude or Copilot answers in your editor in two seconds?

Fair enough. I haven't posted a question on Stack Overflow in years either. But here's the part nobody's sitting with: every model you're asking instead was trained on Stack Overflow. We just watched an AI drink a well dry — and the well was the thing that fed the AI.

The Feedback Loop Nobody's Naming

There's a loop here so obvious it's somehow invisible. The models got good at coding by ingesting an enormous corpus of humans publicly working out hard problems: Stack Overflow threads, GitHub issues, blog post-mortems, mailing-list arguments — the whole messy archive of people thinking out loud where a scraper could see it.

Then the models got good enough that the humans stopped publishing. Why write the painstaking Stack Overflow answer when you can paste the working code into your own repo and move on? Why blog the fix when the only "reader" is a crawler that'll launder your work into an answer with no link back? The incentive to externalize knowledge — to put what you learned somewhere public — is collapsing, because the audience and the credit both evaporated.

So the corpus that made the models stops growing. Worse, the public web increasingly fills with AI-generated content, which means the next training run is partly the model eating its own output. That's the model-collapse failure mode researchers keep warning about — except it's not a lab hypothetical anymore. Stack Overflow is the field demo.

The Knowledge That Calcifies

Think about what actually lives in a Stack Overflow answer. It's not the easy stuff — nobody asks how to write a for-loop. The valuable threads are the weird, specific, undocumented landmines: this library's version 3.2 breaks on this OS in this exact config, and here's the one-line workaround a stranger found at 2 a.m. so you didn't have to.

That knowledge gets created at the frontier of new tools — new framework versions, new APIs, new failure modes that didn't exist last year. Historically it flowed: human hits the wall, human solves it, human posts it, everyone else (and the next model) learns it. Cut the "posts it" step and the flow stops at the moment of the new and the hard.

Which means the models' practical knowledge starts to calcify around the world as it was when people still bothered to write things down publicly. They'll stay sharp on React 18 forever and get progressively shakier on whatever ships in 2027, because the connective tissue — humans publicly debugging the new thing — is being severed. The AI is great at the problems we already solved in the open. The genuinely novel ones increasingly have no answer for it to have learned, because we solved them in private and told no one.

That's also, quietly, where your value goes up. Being the person who can solve the undocumented thing — the problem with no Stack Overflow answer because it's too new or too niche — is exactly the skill the models can't shortcut, precisely because we stopped feeding them the data that would let them.

A Commons Got Exhausted in Real Time

Here's my take: we treated public knowledge like it was free and infinite, and it was neither. It was a commons, and commons get exhausted when everyone extracts and nobody replenishes. Stack Overflow's collapse isn't a story about a clunky old forum losing to slick AI. It's the first visible crack in the supply chain of machine intelligence — the human-generated, freely-published knowledge that the whole edifice quietly assumed would keep showing up forever.

I don't think this ends the models. They'll buy data, license it, generate synthetic training sets, pay humans to produce it in controlled settings. But notice what that does: it turns a free commons into a metered, owned, paid pipeline. The knowledge doesn't stop existing — it stops being public. The same migration-behind-the-paywall I expect for blogs, happening to the raw material of AI itself.

So the practical move, if you're a developer or a team lead: keep an internal commons alive. The post-mortems, the "why we chose this," the gnarly-bug write-ups — capture them somewhere your own people and your own tools can learn from, because the public version of that habit is dying and you can't assume the open web will carry it for you anymore. We spent fifteen years building a shared brain in public. We're about to find out what it costs to rebuild it in private.

Sources: Stack Overflow traffic collapses as AI tools reshape how developers code | PPC Land · Stack Overflow Traffic Collapses 75% as AI Replaces Developer Q&A | byteiota · AI Killed the Stack Overflow Star: The 76% Collapse in Developer Q&A | Allstacks · Dramatic drop in Stack Overflow questions as devs look elsewhere for help | DevClass