Zero. Or close to it, anyway, if you look at the delta of new, high-quality, niche technical queries being posted on Stack Overflow compared to three years ago. The line on the graph doesn’t just dip; it craters. For a decade, the site was the undisputed sovereign of the developer’s workflow, the place you went when the documentation was written by someone who hated humans and your code was throwing an error that looked like a random string of Wingdings. Now, the traffic is evaporating.
The data from the Stack Exchange Explorer shows a collapse in views and questions that aligns perfectly with the release of ChatGPT. It is the most honest visualization of the “AI effect” we have seen yet. While the industry loves to talk about “copilots” and “augmentation,” this graph shows a replacement. We aren’t using AI to help us find answers on Stack Overflow; we are using AI to avoid the site entirely.
(And honestly, who cares about the “duplicate” police anymore?)
The friction of the old way was immense. You had to format your question perfectly, provide a minimal reproducible example, and then pray you wouldn’t get downvoted into oblivion by a moderator who woke up on the wrong side of the bed. LLMs removed that social tax. You can be as messy, incoherent, and desperate as you want in a chat prompt, and the model will still try to help you. It is a frictionless experience that makes the traditional Q&A forum look like a trip to the DMV in the 1970s.
Here is the problem: LLMs are not magic. They are statistical mirrors. For years, they have been eating Stack Overflow for breakfast, training on the millions of hours of human frustration and correction that comprise the site’s archive. The model doesn’t “know” how to fix your Python memory leak; it knows how ten thousand humans fixed that leak between 2008 and 2022.
By migrating the entire developer base away from the public forum and into private chat windows, we have effectively cut the feed. We are burning the library to keep the heater running. It is like a chef burning his only cookbook because he bought a microwave that claims to taste the same. The microwave is convenient, sure, but it cannot invent a new recipe.
When the flow of new, human-verified solutions stops, the models begin to eat their own exhaust. We are entering the era of model collapse, where AI is trained on AI-generated content, leading to a degradation of nuance and an increase in confident hallucinations. If no one is posting the “weird edge case” on a forum because they’re asking a bot, that edge case never gets documented for the next generation of weights.
Does this actually matter for the average dev? Probably not in the short term. Most of us are just happy to have the boilerplate written for us while we drink our coffee. But for the field as a whole, it is a suicide pact. We’ve traded the long-term health of the knowledge commons for the short-term convenience of a chat interface.
It is a suicide pact.
The industry will likely try to pivot. We will see “verified” human data markets or paid contributions to training sets, but that is a far cry from the organic, chaotic brilliance of a public forum. The incentive structure is gone. Why spend an hour debugging a problem and writing a clear explanation for the public when you can just get a “good enough” answer from a bot in three seconds?
By Q4, we will see the first major coding model release that explicitly admits to a performance plateau due to the lack of fresh, human-generated training data. The plateau is coming because we’ve stopped contributing to the source. We’ve optimized for the answer and forgotten about the process of discovery. The graph is a warning, but we’re too busy prompting to notice the cliff.
