Stack Overflow's questions have dropped 30-fold since AI assistants arrived, and the knowledge archive that trained those AIs is now going stale

The Machines Ate Stack Overflow. Now They're Starving.

Stack Overflow's parent network has seen question volume fall roughly 30-fold from its peak. The decline began well before ChatGPT, but AI coding assistants turned a slow bleed into a collapse. And the same tools that emptied the archive were trained on that archive's verified answers. Nobody is refilling it.

The scale is stark. Across the entire Stack Exchange network, question volume peaked at 286,300 posts in March 2017 and fell to 9,883 by March 2026, a 29-fold drop 1. On Stack Overflow specifically, only 3,862 questions were posted in December 2025, down 78 percent from the prior year 2. The Pragmatic Engineer, the top software engineering newsletter on Substack, declared the site "almost dead" 3.

The immediate cause is no mystery. When a developer hits a wall, asking ChatGPT or Copilot takes seconds. Posting to Stack Overflow takes longer, risks a dismissive rejection from moderators, and often returns the same answer an AI would have given anyway. One developer captured the sentiment in a community discussion: "AI certainly accelerated the decline, but this is the result of consistently punishing users for trying to participate in your community" 2.

But the 30-fold collapse in new questions is the surface story. The deeper problem is recursive, and nobody is solving it.

Stack Overflow's archive of verified, community-vetted answers is the training corpus that made AI coding assistants useful. ChatGPT learned to debug a React hook or configure a Kubernetes pod by ingesting millions of human-validated solutions, and the Pragmatic Engineer confirms the platform's data was a training source for the models that replaced it 3. That archive is now going stale. Frameworks ship new APIs. Libraries deprecate methods. Cloud providers change their console layouts and pricing models. The canonical answers that powered the AI coding revolution are frozen near the moment contributions began to dry up, and no equivalent volume of human expertise is replacing them.

This is not an abstract concern. When a developer asks Copilot how to handle authentication in a framework that shipped a breaking change six months ago, the assistant may confidently recommend a deprecated approach drawn from a Stack Overflow answer no human has updated since 2022. The verified-knowledge pipeline that made AI coding tools trustworthy in the first place is narrowing, and what flows through it is increasingly dated.

The industry has no ready replacement for human-curated Q&A. AI assistants can generate answers, but they were trained on the very archive that is now decaying. The question one developer posed in response to the latest statistics lands directly: "SO was by far the leading source of high quality answers to technical questions. What do LLMs train off now?" 2.

The decay is already measurable in adjacent systems. The RedMonk programming language ranking uses Stack Overflow for half its data. Analyst Rachel Stephens noted that this data is "increasingly stale and of questionable value on a going-forward basis," with no replacement public dataset available 2. If a respected industry benchmark is already feeling the rot, every system that quietly depends on Stack Overflow's knowledge layer will eventually feel it too.

Stack Overflow's owner, Prosus, reported in its half-year results to September 2025 that Stack Overflow and GoodHabitz together achieved revenue growth of 12 percent to $95 million for the first half of fiscal 2026, primarily driven by Stack Overflow's performance 2. The platform is not commercially dying. The public knowledge commons is eroding while the business around it persists.

Stack Overflow itself has launched AI Assist, a feature that answers questions using its own archive 2. The irony is structural: the platform banned AI-generated answers in 2022, then built its own AI layer on top of the archive that AI helped hollow out.

The Pragmatic Engineer noted that Stack Overflow was already losing momentum before LLMs arrived. Tighter moderation policies starting around 2014 pushed contributors away, and the pandemic created a temporary traffic spike that masked the underlying trend 3. AI did not cause the decline alone. But it accelerated the fall past the point of recovery, and in doing so, it severed the supply of fresh human knowledge that made AI coding assistants worth using in the first place.

This is the feedback loop nobody planned. AI assistants devoured Stack Overflow's traffic by answering questions faster and more politely than the community ever did. But those assistants were nourished by the same human contributions they replaced. As the archive goes stale, the quality of AI-generated code help will drift further from what developers actually need. The machines ate Stack Overflow. Whether they can keep feeding themselves is the question nobody has answered.