It's hard to believe, but in December 2025, the number of questions posted by developers on StackOverflow (opens in new tab) had drastically dropped to 3,862.
To put that into perspective, just three years earlier, in November 2022, the month ChatGPT launched, the site was still getting over 100,000 questions per month.
At its peak in 2014, that number was 200,000. I still remember seeing the December 2025 numbers for the first time and thinking they had to be some kind of mistake.
But the truth is, they're not. The word "decline" doesn't even begin to describe what happened - it's too gentle. A decline is something that happens gradually, over time. What we're talking about here is more like a platform falling off a cliff. A 96% drop in three years, leaving everyone wondering what could have caused such a drastic change.
The collapse is steeper than most people realize.
| Year | Monthly Questions | Context |
|---|---|---|
| 2014 | ~200,000 | Peak activity |
| Early 2023 | ~100,000 | ChatGPT had launched 3 months prior |
| Late 2024 | ~25,000 | Steady decline accelerating |
| December 2025 | 3,862 | 78% drop year-over-year |
Traffic cratered on the same schedule. StackOverflow recorded 161 million monthly visits in March 2025; by early 2026, roughly 55 million. A 66% drop.
The business fallout followed quickly. Layoffs of 10% of the workforce (opens in new tab) in May 2023, then another 28% (opens in new tab) in October 2023. Over 150 people gone from a company of 500 in five months.
What really stands out is how fast it happened - just three years to go from being on top to being almost irrelevant.
I don't think anyone, including StackOverflow's own leadership, expected this to happen so fast.
It's easy to think that ChatGPT is the reason StackOverflow is struggling, but that's not entirely fair. The truth is, StackOverflow had already started to decline long before AI came into the picture.
The moderation culture is the obvious culprit. Aggressive downvoting, questions closed as duplicates of something only tangentially related, a general hostility toward anyone who wasn't already fluent in the platform's norms. A new user would post a question, watch it get downvoted within minutes, and never come back. StackOverflow knew this was happening. They had optimized for a clean, searchable knowledge base at the expense of being welcoming, and that tradeoff held up exactly as long as there was no alternative. The moment AI offered the same answers without the judgment, developers walked. Not because AI answers are better, but because AI doesn't make you feel stupid for asking.
So there was this issue with outdated information. By 2022, a lot of the most popular answers on StackOverflow were really old. For example, a highly-voted answer about React from 2018 might still be talking about class components and lifecycle methods, which was correct back then but not really relevant or helpful for someone learning about it today. The problem was made worse by the fact that the most popular answer always stays at the top, no matter how old it is, which made the whole site feel like it wasn't keeping up with the times. This made it hard to trust the information on the site, because you never knew if what you were reading was still accurate or not.
Large language models like GPT-4 were trained on enormous datasets of human-generated text, and StackOverflow was among the richest sources in those datasets: millions of precise, technical Q&A pairs, peer-reviewed by domain experts, covering virtually every programming language and framework in active use.
That pipeline is now drying up. Developers ask AI instead of posting on StackOverflow, which means no new questions, which means no new answers, which means no fresh human-generated training data for the next generation of models. Researchers have a name for what happens next: model collapse. When AI systems are trained primarily on AI-generated content instead of human content, they degrade. Each successive generation drifts a little further from reality, loses a few more edge cases, becomes a little more generic. Diversity narrows, nuance flattens.
For all its shortcomings, StackOverflow was the real deal - it was genuine and raw, like a honest conversation between developers. It showed how actual programmers think and work through problems, with all the mistakes and wrong turns that come with it. Even with its flaws, it felt human and authentic, and that's something that's still missing today. Nothing has been able to fill the gap and provide the same kind of honest, unfiltered insight into the way developers really work. It's not clear what could replace it, or if anything even can.
It seems like people are looking at this the wrong way, they're saying it's just the loss of a question and answer site, but that's not even close to what really happened.
Let's look at things from a different angle. Sometimes the most helpful information isn't the one that's marked as correct. It's the one that's wrong, but has a lot of people discussing it and pointing out why it doesn't work. This kind of conversation can be really valuable because it shows you how to think about a problem, not just what the solution is. For example, if someone gives a wrong answer, but then 12 other people explain why it's wrong, what would happen if you tried it that way, and what you should do instead, that's really useful. You learn a lot more from that than from just seeing the right answer. It's like having a whole conversation about the topic, rather than just getting a quick fix. And that's what makes it so helpful. You get to see the reasoning behind the answer, and that's what really matters.
When AI handles the same question, you get one confident response. StackOverflow gave you five conflicting ones and forced you to figure out which was right. Those are fundamentally different learning experiences.
The real value of StackOverflow wasn't just in the answers themselves, but in the discussions that took place in the comments section. That's where engineers would dive into the specifics, pointing out potential issues and limitations. For example, someone might say, "This solution works, but it's not efficient for large datasets because it's O(n²)." Or, "This approach is fine, but it won't work if you're using an older version of PostgreSQL, like anything before version 12." Others might chime in, saying, "This is okay for a simple application, but it's not suitable for a distributed system." It's this kind of nuanced, contextual knowledge that AI systems struggle to replicate. They can provide general information and answers, but they often miss the exceptions and caveats that are crucial in real-world applications. Those exceptions, which can make or break a project, were often buried in the comment threads, where engineers would engage in lively debates and share their expertise.
There's something more, something that's not so easy to put into words. Let's call it the habit of learning through browsing around. Think about it, back in 2020, if a junior developer wanted to learn about database indexing, they'd search for it on StackOverflow. They'd find a few related questions, read the answers and the comments, and then follow some links to other related questions. After about half an hour of doing this, they'd come out with a really deep understanding of the topic. It wasn't just about getting a simple answer - they were also taking in all the context that came with it.
In 2026, that same developer asks ChatGPT, gets a response, pastes it into their code, and moves on. The answer might be perfectly correct.
But the learning didn't happen.
Something really caught my eye in StackOverflow's 2025 Developer Survey: a whopping 82% of people who took the survey said they still use StackOverflow pretty regularly - at least a few times a month. And what's even more interesting is that about 35% of them said they go there because they're having trouble with code that was generated by AI and needs to be fixed.
Think about what that means. StackOverflow is becoming an AI fact-checker, which is a product its founders never envisioned building. Developers use AI for the first pass, then go to StackOverflow to verify whether the AI got it right. That's "ask a question, get an answer" turned inside out. The platform that AI was supposed to kill is becoming necessary precisely because AI is unreliable. There's a real irony in that.
Whether StackOverflow can build a sustainable business around this new role is unclear. Prosus reported that Stack Overflow and GoodHabitz achieved 12% revenue growth to $95 million combined in the half-year to September 2025, driven primarily by enterprise products. The community that built StackOverflow's value was never the revenue engine; the data it produced was.
If you're a senior engineer, your value just went up, because evaluating a solution is becoming scarcer than generating one. AI can produce ten approaches to a problem in seconds. Knowing which one will buckle under load, which one has security implications, which one creates maintenance debt six months from now requires judgment that only comes from experience. StackOverflow's comment sections used to build that kind of judgment almost as a side effect. That training ground is gone.
If you're earlier in your career, the temptation is to let AI handle every question and skip straight to working code. The engineers who will stand out are the ones who resist that, who treat AI output as a first draft and then dig into documentation, understand the tradeoffs, and occasionally break things on purpose to learn how they fail. Developing your own sense of why something works matters more now than it did five years ago, not less.
And if you're hiring, it's probably time to rethink what you're testing for. " Implement a binary search tree" is a useless interview question now. " This system is dropping 2% of messages under load; walk me through how you'd diagnose it" tells you what you actually need to know about a candidate.
To be honest, I'm not sure I have a straightforward answer to this, and I think it's wise to be skeptical of anyone who says they do.
The StackOverflow model, which was a public knowledge base created and maintained by volunteers, had a good run for fifteen years and achieved something really amazing. However, it's not going to be back in the same way it was before.
So, what can help fill the gap a bit? It's likely that content curated by experts and reviewed carefully will become more valuable as the internet gets flooded with stuff made by AI. This is because when everything is generated, it's the human touch that makes things stand out. Smaller online communities, like some Discord servers or specialized forums, might take over some of the expert discussions that used to happen on StackOverflow. Also, AI tools that show where they got their information and explain how they arrived at their answers, instead of just presenting them with too much confidence, could help solve the problem of trust.
But the core loss is real. There was a public, searchable space where engineers argued about the right way to solve things, and those arguments were preserved for the next person who hit the same wall. That space is effectively gone now. We'll be feeling its absence for a while.
Data sources: SimilarWeb (opens in new tab), WebProNews (opens in new tab), DevClass (opens in new tab), The Verge (opens in new tab), TechCrunch (opens in new tab), StackOverflow Blog (opens in new tab), 2025 Developer Survey (opens in new tab), NextInData (opens in new tab).
Entry-level software engineer jobs dropped 67% in one year. Stanford's study confirms AI is replacing junior developers, while senior developer employment grew 9%. Here's what the data actually shows.
Cloudflare used Vercel's open-source test suite and AI to rebuild Next.js as Vinext in one week for $1,100. What this means for open source and software.
Vibe coding promises 10x productivity. The data says otherwise: 40% of AI code has security flaws and developers score 17% lower on comprehension.
