Back to all findings

Who killed Stack Overflow?

Last month it got 968 new questions. At its peak it got that many before breakfast. I pulled every month since 2008 and lined up the suspects. The obvious one had help.

It’s 2 am in 2013 and something is broken. You don’t know why. You paste the error message into Google, and the first result is a page with an orange logo: a question from someone exactly as lost as you, and under it, with a green tick, an answer from a stranger who fixed it for free.

That page was Stack Overflow. For about ten years it was how the world’s programmers got unstuck. If you’ve ever written code, you’ve been there. If you haven’t, you’ve used something built by someone who was.

Last month, the whole site got 968 new questions.

In March 2014, it got that many before breakfast.

200,000 a month 100,000 March 2014207,169. One every 13 seconds. LockdownEveryone learns to code 30 Nov 2022ChatGPT September 2026968. One every 45 minutes.
2010 2012 2014 2016 2018 2020 2022 2024 2026
New questions on Stack Overflow, by month, July 2008 to October 2026. Pulled from the Stack Exchange API on 30 October 2026; questions deleted since aren't counted. Download the numbers (CSV).

A chart like this went around this week, and I didn’t quite trust it, so I pulled the numbers myself: every month since July 2008, from Stack Overflow’s own API. The chart was right. If anything it was kind.

  • The peak was March 2014: 207,169 questions. One every 13 seconds, all month.
  • September 2026: 968. One every 45 minutes.
  • That’s fewer than August 2008, when the site was a private beta and you needed an invitation to ask. The beta got 3,744.
  • And it’s speeding up. In 2023 the number halved about every ten months. In 2026, about every four to five, though the last three months have held at about a thousand.

March 2014

37

new questions in the eight minutes this post takes to read.

September 2026

0

in those eight minutes. The next one is due in about 45 minutes.

Stack Overflow at each month's pace. With JavaScript on, this counts from the moment you opened the page.

So who did it? Every post I’ve read this week has the same answer: ChatGPT. Case closed.

I don’t think it’s that simple. Four suspects.

Suspect 1: the chatbot

Name
ChatGPT, and everything that followed it.
Arrived
30 November 2022.
Motive
None. It just answers.
Means
Four seconds. Any hour. Never says "duplicate".

The cliff is real. In November 2022, the month ChatGPT came out, Stack Overflow got 109,120 questions. Twelve months later, 50,273: down 54%. By last month, down 99%.

100,00010,0001,000 30 Nov 2022ChatGPT The private beta, August 2008: 3,744 968
2022 2023 2024 2025 2026
The same numbers since 2022, on a scale where every gridline is ten times fewer. A straight line here means the fall never slowed. In January 2026 it dropped below the private beta's month.

What a chatbot does that a forum can’t: it answers at 2 am, it never tells you your question has been asked before, and it never downvotes you. In Stack Overflow’s own survey this October, 64% of developers said they now visit it less for simple questions, and more than 70% said they use AI every day.

This isn’t just two lines that happen to fall together. In 2024, three researchers compared Stack Overflow with its Russian and Chinese equivalents, where ChatGPT was officially unavailable. Within six months, Stack Overflow had lost a quarter of its activity relative to them. They called it “a significant shift of knowledge from public to private domains.”

Stack Overflow saw it coming. Five days after ChatGPT launched, its moderators banned ChatGPT-written answers, because “the average rate of getting correct answers from ChatGPT is too low.” Six months later, its own volunteer moderators went on strike, after the company told them to stop removing answers they suspected were written by AI.

Guilty? Of the blow, yes.

But look at the chart again. The line had peaked eight years before this suspect was born, and had been falling for five.

Suspect 2: the duplicate police

Closed as duplicate

From March 2014 to November 2022, questions fell by 47%. Nearly half the site was gone before ChatGPT existed.

Some of that is natural. After a few years, the common questions have been asked, and a good answer to “how do I undo a git commit” doesn’t need asking twice. That page has been read 16.8 million times. The whole point of the site was to build a library, not a chat.

But the library had bouncers. Ask a question in the wrong shape and it got closed within minutes, downvoted, or marked as a duplicate of a question from 2011 that wasn’t quite the same. By 2014, one study found, nearly 40% of questions were being deleted or going unanswered, up from 22% three years earlier.

Everyone who used the site has a story. Stack Overflow admitted it, in a company blog post in April 2018 titled “Stack Overflow Isn’t Very Welcoming. It’s Time for That to Change.” “Too many people experience Stack Overflow as a hostile or elitist place,” it said. The fixes that followed, a question wizard and more reputation for askers, didn’t turn it around. The people who were told to go away went away, and told their friends.

2011 about 22 in 100

2014 nearly 40 in 100

Of every 100 questions asked, how many were deleted or never got an answer. From Ivan Srba and Mária Bieliková's study of the site (2016).

Guilty? Of the wound the blow landed on.

Suspect 3: the company

Owner
Prosus, since 2021, for $1.8 billion.
2023
Two rounds of layoffs: 10% in May, 28% in October.
2024
Licensed the archive to Google, then OpenAI. Put the public data dump behind a login.
2025
Put its own chatbot on the site as the new front door, with asking as the fallback.
2026
Prosus has written off 87% of what it paid.

Here’s the part most posts skip. In 2024, Stack Overflow sold access to its archive to the two companies whose chatbots were emptying it: Google in February, OpenAI in May. When one user edited his best answers in protest, they were restored within the hour and he was suspended for a week.

Everything ever posted there is licensed so that anyone can reuse it, forever, as long as they credit the author and share alike, and the site’s own terms add that you can’t take that permission back. That same summer, the company stopped posting the public data dump to the Internet Archive, put it behind a login, and made downloaders promise not to train models on it.

Free for the labs that pay. Not free for you to train on.

Then, in December 2025, it put its own AI answer box on stackoverflow.com, billed as the new entry point to the site, with asking the community as the fallback. In November, 5,347 questions. In January, 3,614: a 32% drop across a winter that had cost the site 5% to 13% in each of the three winters before. It kept falling until this autumn, when it settled at about a thousand.

2019/20−1%
2020/21+4%
2021/220%
2022/23−12%ChatGPT
2023/24−5%
2024/25−13%
2025/26−32%
Questions in January against the November before, every winter since 2019. Before ChatGPT, winter barely made a dent. From the Stack Exchange API, 30 October 2026.

In fairness, the company doesn’t hide any of this. Its head of community wrote in September that “people don’t need to come directly to Stack Overflow anymore in order to get answers to their programming questions.” The day after the chart went viral, it published a post titled “The reports of Stack Overflow’s death are greatly exaggerated”. It doesn’t mention the chart. The site, it says, is turning into “knowledge infrastructure”: a private version for companies, and a new site where AI agents post and humans check their work.

Guilty? Of selling the X-rays while the patient was on the table. Though the patient was already dying, and the public ward was being paid for by the private one.

Suspect 4: us

The strangest number I found is this one. In September 2026, the site got 968 questions and 2,347 answers. That’s 2.4 answers for every new question. At the 2014 peak it was 1.5.

The answerers are leaving too, but a step behind the askers, and the ones still there answer more than one question each. The librarians are still at the desk. The library is empty.

We didn’t stop asking questions. We stopped asking them where anyone else could see. The same question that would have been a public page in 2014, with a green tick and the next ten thousand visitors with the same bug, is now a private chat that disappears when you close the tab. Twenty to thirty million visits a month still land on Stack Overflow. Almost nobody writes.

One month on Stack Overflow: about 25 million visits and 3,315 new questions and answers.

Each dot is one visit: 7,500 of them, and one new post. Visits are Similarweb's and Semrush's estimates for September 2026, 20 to 30 million; posts are September's 968 questions and 2,347 answers.

Guilty? Of preferring to be answered over being corrected. Which is to say, of being human.

The twist

The killer learned to code from the victim.

When Meta published the recipe for LLaMA, its first widely used language model, in 2023, Stack Exchange, the family of sites Stack Overflow belongs to, was 2% of the training data. In The Pile, the dataset many early models were built on, it was 5%. Twenty-four million questions and thirty-six million answers, written by volunteers, for free, under a licence that lets anyone use them. One man, Jon Skeet, wrote 35,806 of those answers himself.

LLaMA Meta, 2023

  • Web pages67%
  • More web pages15%
  • Code, from GitHub4.5%
  • Wikipedia4.5%
  • Books4.5%
  • Science papers2.5%
  • Stack Exchange2%

The Pile EleutherAI, 2020

  • Web pages18.1%
  • Medical papers14.4%
  • Books12.1%
  • Pages linked from Reddit10%
  • Science papers9%
  • Code, from GitHub7.6%
  • Court opinions6.1%
  • Stack Exchange5.1%
  • 14 smaller sources17.6%
What two training sets were made of, by share. Stack Exchange is the family of sites Stack Overflow belongs to. From Table 1 of each paper.

Every time a chatbot fixes your Python error at 2 am, somewhere in what it learned from is a stranger’s answer from 2013. The thing that replaced Stack Overflow is, in part, Stack Overflow.

Case file: Stack Overflow, 2008 to 2026

  • The chatbot delivered the blow. 99% gone in under four years.
  • The duplicate police had already cost it half its questions, and most of its goodwill.
  • The company sold the archive to the companies emptying it, then built its own chatbot too.
  • We took the faster answer, every time. We'd have done it in 2014 too.

Why I care

Here’s what the chart doesn’t show. When a new programming tool ships next year, people will get stuck on it. In 2014 those questions became public pages, searchable by the next ten thousand people with the same problem, and, as it turned out, readable by the machines that learned to code from them.

Now those questions go into chat windows. Nobody else can search them. No future model can learn from them, and a 2024 paper in Nature found that models trained mostly on other models’ output degrade, generation by generation. The library mostly stops at 2023, and every answer after that lives in a tab somebody closed.

Questions asked each year, one book per year, from July 2008 to October 2026. From the Stack Exchange API, 30 October 2026.

I don’t know how that ends. Maybe the models are good enough that it doesn’t matter. Maybe in five years, a stranger with a green tick is something you explain to your kids.

What I actually took away:

  • Stack Overflow didn’t die of one thing. It was bleeding for five years, then it got hit by a truck.
  • The answerers are leaving a step behind the askers. The last ones out are still answering.
  • The best free record of how programmers get stuck was sold to the people who ended it, legally, with the authors’ permission in the small print.
  • If you learn something hard this year, write it down somewhere public. Someone is at 2 am right now.
Sources (17)

Up next

All findings