Quick Facts
- A University of Auckland study tracked 24,304 Stack Overflow contributors over 17 months and found high-reputation users withdrawing at accelerating rates since 2022.
- Stack Overflow’s monthly question volume has collapsed from 200,000 in 2014 to under 50,000 by late 2025, a retreat to 2008 levels.
- 84% of developers now use AI in their work, and 68% rarely or never participate in Stack Overflow Q&A despite 82% still visiting the platform.
Expert developers are leaving online communities, and generative AI is the reason. A working paper from Dr. Kenny Ching at the University of Auckland Business School documents the departure of high-reputation contributors from Stack Overflow, the world’s largest developer Q&A platform, tracing the acceleration to 2022 when generative AI tools went mainstream.
Ching tracked 24,304 Stack Overflow contributors over 17 months. Less established users left first. But the departure rate among veterans kept climbing, closing the gap over time.
Signal Compression
Ching’s working paper, titled “When Effort Stops Signalling: AI, Signal Compression, and the Withdrawal of Authentic Performance,” introduces the concept of signal compression. When AI can produce expert-looking answers instantly, human expertise becomes indistinguishable from automated output.
“They aren’t leaving because they can’t compete with the technology,” Ching said. “They’re leaving because their hard-earned expertise is no longer distinct from a chatbot’s answer.”
The community mechanics that built Stack Overflow depended on reputation as a proxy for trust. A complex, well-explained answer earned upvotes and status. AI erased that reward structure.
The Numbers Behind the Collapse
Stack Overflow’s own Data Explorer shows posts in April 2025 were down 64% from April 2024 and down more than 90% from the 2020 peak. Questions have dropped 76.5% since ChatGPT launched. By late 2025, monthly submission volumes had fallen to 2008 levels, erasing 15 years of growth.
By March 2023, ChatGPT had already driven a 12% reduction in average daily web visits to Stack Overflow. A Stack Overflow developer survey from 2025 found that 82% of developers still visit the platform at least a few times per month, but 68% don’t participate or rarely participate in Q&A.
Separately, Cornell University researchers found that AI-generated content threatens online communities on multiple levels. Sixty percent of moderators cited degraded content quality. Sixty-seven percent cited disruption to social dynamics and authentic human connection.
A Feedback Loop With Long-Term Consequences
The business risk extends beyond one platform. ChatGPT trained on Stack Overflow’s data. That training helped reduce Stack Overflow’s traffic. Stack Overflow is now selling that declining dataset back to AI companies, having signed partnerships with both OpenAI and Google Cloud.
Researchers call the downstream risk model collapse: a degenerative process where AI models trained on AI-generated data produce narrower and lower-quality outputs over successive generations. A February 2026 article in Communications of the ACM stated that model collapse is not a theoretical risk but is occurring in production systems today.
“Within a few generations, original content is replaced by unrelated nonsense,” one analysis noted.
Ching sees the problem spreading beyond developer forums. “This isn’t just about coding platforms,” he said. “I argue this exact same dynamic is discouraging genuine effort in classrooms, corporate workplaces, and scientific communities. The long-term risk is that by crushing the incentive to show genuine effort, AI might actually truncate the formation of future human expertise.”
What This Means for Software Leaders
For founders and engineering executives, the departure of expert contributors from open knowledge communities has a direct cost. Stack Overflow has long served as a free, crowd-sourced support layer for software development. As that resource degrades, teams will rely more heavily on AI tools whose training data is itself becoming thinner and less reliable.
The irony is structural. The same AI tools that reduced the incentive for experts to share knowledge online are now dependent on that shared knowledge to remain accurate. Companies building on top of large language models should consider what happens when the human-generated data that powers those models stops growing.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.