← All postsΣ Society
The internet is running out of people who answer questions
Stack Overflow now gets about 2.5% of the questions it got before ChatGPT. I went through the Stack Exchange data dump, Wikipedia's statistics and archived robots.txt files to see what else is shrinking, and why.
In September 2026, people asked 1,041 questions on Stack Overflow, counting the ones that were still on the site at the end of the month. In October 2022, the last full month before ChatGPT came out, they asked 105,744. August and September 2026 both had fewer questions than August 2008, the site’s first full month. The people who answer are going the same way: 89,654 different people answered at least one question in March 2017, and 1,561 in September 2026.
The text that language models learned from, Stack Overflow and Wikipedia and forums and the rest of the open web, exists because people wrote it, in places where readers and search traffic came back to them. I wanted to know how much of that has shrunk since ChatGPT and AI answers in search arrived and how much of the shrinking AI is responsible for, because that decides what the next models will be trained on.
So I pulled the data: the full Stack Exchange data dump of September 30, 2026, read straight out of its 7z archives, Wikipedia’s traffic and editor statistics, archived robots.txt files of the domains that contributed the most text to C4 (a web dataset that many language models were trained on), and published numbers from Cloudflare, Pew Research and company filings. Monthly series run through October 2026 where the source allows it; the data dump ends with September.
Questions grew until 2014 and then declined slowly, by about 7% a year according to a trend fitted to 2016–2022, with a bump in the spring of 2020 when people were stuck at home. After ChatGPT the line falls off a cliff: 1,334,634 questions in 2022, 787,599 in 2023, 398,365 in 2024, 109,033 in 2025 and 19,041 in the first nine months of 2026. The old trend would have predicted about 82,000 questions for September 2026. The actual number was 1.3% of that.
A data dump only contains posts that weren’t deleted, so I checked the decline against something deletion can’t touch. Every post on the site takes its id from one counter, and the gap between the first ids of two consecutive months counts everything created in between, deleted or not: 355,271 posts in October 2022, 7,379 in September 2026. Deletions don’t change the picture, although more of the little that gets posted now disappears: 60% of everything posted in 2026 so far, against 31% of what was posted in 2022.
How the loop worked
Public Q&A ran as a loop. Someone asks in public, and someone else answers, for reputation points or because they fought the same error last week. Search engines index the page, and for years afterwards everyone who pastes that error message into a search box lands on it. Some of those readers vote or leave a comment that corrects the answer, a few start asking and answering themselves, and the traffic pays for the site through ads and job listings.
A language model takes the output of that loop, the published answers, and returns nothing to its inputs. When someone asks a chatbot instead of searching, the answer is generated in a private conversation that nobody else can find or vote on, and the site that taught the model never sees the visitor. The rest of this post follows the loop around, starting with the people who ask.
Who stopped asking
Stack Overflow had 113,597 different people asking questions in an average month of 2016, 80,983 in 2022 and 1,853 in 2026. The number asking their first question ever went from 28,155 a month in 2022 to 634.
The first studies of the decline found it concentrated among newer accounts (Burtch, Lee and Chen) and casual users (Quinn and Gutt). Del Rio-Chanona, Laurentsyeva and Wachs compared Stack Overflow with its Russian and Chinese counterparts, where ChatGPT wasn’t officially available, and estimated that six months after ChatGPT’s release, activity on Stack Overflow had fallen by 25% relative to them. The longer data fits that picture. The median question got about 40% longer between 2022 and 2025, and the share that includes code went from 83% to 87%, so the easy questions seem to have left first. In the 2026 Stack Overflow Developer Survey, 64% of the developers who answered the question said they visit the site less for simple questions because of AI.
Stack Overflow changed things itself, too. The moderators’ strike of June to August 2023 over AI-generated content doesn’t show up in the series. The changes of early 2026 might: the site extended its experiment with opinion-based questions to all users in February and released the beta of a redesign in March, and over the same months the decline sped up from about 6% a month to about 15%. My data can’t say whether that was the site, better AI tools or both.
Who still answers
For most of the site’s history the people who answer were the smaller group, and they have gone the same way. To see who is left, I grouped every answerer by the year of their first answer on the site.
The inflow of new answerers dried up along with the questions: about 22,100 people a month posted their first answer in 2015, and 625 a month in 2026. Of the people who answered anything in the twelve months to September 2026, 41% gave their first answer before 2017 and 33% started in 2023 or later.
Since 2025 there have been more people answering than asking, 2,757 against 1,853 a month in 2026, but look at what they answer. In 2014, 13% of all answers went to questions that were more than a month old. In 2025 and 2026 it’s 44%. A large part of what is still written on Stack Overflow is new answers to old threads, not help for someone who is stuck today. New questions also wait longer: the median time to a first answer went from 25 minutes in March 2014 to 70 minutes in October 2022 and 2 hours 15 minutes in September 2026, and the share answered within a day fell from 76% in 2014 to 58% in 2026.
Which questions went first
Del Rio-Chanona and her coauthors found that the first six months hit the most widely used languages hardest, Python and JavaScript among them. Four years in, nearly everything has collapsed, so the interesting part is the order. I counted the questions per tag in the year before ChatGPT, the year to September 2024 and the year to September 2026.
By the year to September 2024, questions about basic constructs (loops, lists, strings, regex) were down to 17% of their pre-ChatGPT level, while questions about platforms and tools (Docker, Azure, Git, iOS) were still at 42%. Two years later they were at 0.7% and 3.1%. What emptied first is the kind of question that has one well-known answer, which a model has seen thousands of times. What held on longer depends on a specific environment, like a cloud account or a build that only fails on one machine. A tag’s size doesn’t predict how far it fell: across all 259 tags with at least 2,000 questions in the year before ChatGPT, the rank correlation between size and decline is −0.05. I grouped the tags in the figure after seeing the numbers, so read the groups as a description, not as a test.
The rest of Stack Exchange
If people abandoned Stack Overflow only because chatbots got good at code, the other Stack Exchange sites should look different, at least the ones about cooking or Jewish law. Some of them were already shrinking faster than Stack Overflow, so for each of the 96 sites with at least 50 questions a month in the year before ChatGPT, I fitted the same kind of trend to 2016–2022 and compared what happened afterwards with what it predicted. TeX is missing, because the posts in its archive are corrupt halfway through.
Every group fell. Of the 96 sites, 94 are below their own trend, 90 of them below its whole 90% interval, and the median site gets 17% of the questions it got before ChatGPT. Del Rio-Chanona’s comparison sites went the same way: Russian Stack Overflow is at 4% of its old volume and Mathematics Stack Exchange at 16%, so the comparison that worked in 2023 doesn’t work anymore.
The timing is the interesting part. Stack Overflow, its versions in other languages and the programming, math and science sites broke away from their trends within months of ChatGPT’s release. The sites about languages, humanities and everyday life (English, History, Philosophy, Cooking, Travel, DIY and so on) stayed close to their trends through most of 2024 and then fell: the median hobby site was at 79% of its trend in the third quarter of 2024 and at 27% at the end of 2025. Google had launched AI Overviews in the US in May 2024 and expanded them to more than 100 countries at the end of October, and chatbots kept getting better and more widely used. The data fits a first wave in which chatbots replaced help with code and math and a second one in which AI answers reached everything else, but it can’t tell those mechanisms apart, and it can’t rule out other changes in how Google ranks Stack Exchange pages.
A few sites still look like they did before. Philosophy, Judaism, Biblical Hermeneutics and Puzzling get between 74% and 89% of their old volume, and MathOverflow, where research mathematicians ask each other questions, gets 47%. My guess is that askers there want other people’s interpretation, or a problem worked through by someone who enjoys it, more than a fact.
Wikipedia
Wikimedia publishes Wikipedia’s traffic split by who is reading: people, crawlers that identify themselves, and bots that its heuristics catch pretending to be people. I pulled the monthly numbers for English Wikipedia, together with the number of editors who made five or more edits to articles in a month.
Human pageviews averaged 7.68 billion a month in 2016–2019 and 7.63 billion in both 2023 and 2024. From January to October 2026 they averaged 6.56 billion, 14% less than in the same months of 2024, most of it on the mobile web, and every month since May 2024 has been below the same month a year earlier. The Wikimedia Foundation points at two causes at once: search engines answering directly with generative AI, and younger people looking things up on social video platforms. There is a measurement caveat it has been open about. After bots built to evade detection inflated the “human” numbers in 2025, it reclassified the traffic from March to August 2025 and found human pageviews down about 8% from 2024, but earlier months were counted with the older rules, so the real decline is probably somewhat smaller than 14%. The traffic it does catch as automated rose from 15% of all pageviews in 2016–2019 to 36% in 2025.
The editors haven’t moved. In the year to October 2026, about 30,500 people a month made five or more edits to articles, against 30,300 in 2016–2019. It’s the only place in this post where the people who produce the material are still there in the same numbers, and Wikipedia works differently from a Q&A site: nobody edits an article because a stranger asked them something that morning. The crawlers are a cost of their own. In April 2025 Wikimedia wrote that at least 65% of its most expensive traffic came from bots, and in January 2026 it announced Amazon, Meta, Microsoft, Mistral AI and Perplexity as customers of Wikimedia Enterprise, its paid API, next to Google.
The open web
Websites can turn crawlers away with robots.txt, and since OpenAI documented GPTBot in August 2023, the large AI companies have published crawler names that can be blocked. For the Data Provenance Initiative, Longpre et al. found that in the year to April 2024, robots.txt restrictions made about 5% of all tokens in C4, and at least 28% of the tokens from its most actively maintained, critical sources, fully off limits. In October 2025, Bouchaud and Ramaciotti found that a quarter of the thousand most visited websites restrict AI crawlers.
I wanted to continue that series to October 2026 with the Initiative’s own domain list and archived copies of each domain’s robots.txt from the Wayback Machine. After a few hundred requests the Wayback Machine started answering with 403 errors and refusing connections, and it kept doing so even at one request every five seconds. So I have a complete history only for the eleven domains with the most text in C4, 9.2% of the text of the top 2,000. They are at least the largest sources there are.
The news sites block almost every AI crawler, most of them by April 2024; the New York Times already blocked Common Crawl’s CCBot in July 2023. Wikipedia, Google Patents and PLOS block none. Two blocks went away: the Guardian dropped its GPTBot block between April 2024 and January 2025 and announced a content partnership with OpenAI in February 2025, and Scribd dropped its block too. None of them blocks Googlebot, because that would mean leaving Google Search. Google offers a separate token for opting out of Gemini training, Google-Extended, but according to its documentation that token “does not impact a site’s inclusion in Google Search,” and AI Overviews are part of Search. Cloudflare reported in September 2026 that 17% of the sites behind it block AI training in some way.
Blocking a crawler only matters if the crawler would have sent something back. Cloudflare measures that as a crawl-to-refer ratio: the HTML pages a company’s crawlers requested, divided by the page visits that arrived with a referrer from that company’s products.
In July 2025, Google’s crawlers fetched 5.4 pages for every visit Google sent back. OpenAI’s fetched 1,091, and Anthropic’s 38,066, down from 286,930 in January. Even if the missing referrers hide a factor of ten, the AI companies take orders of magnitude more than they return. Cloudflare classified 80% of AI crawling in the year before that post as collecting training data.
Search engines are turning into answer engines too. Pew Research tracked the browsing of 900 adults in the US in March 2025: on Google searches that showed an AI summary, people clicked an ordinary result in 8% of visits, against 15% when there was none. Google, whose AI Overviews have more than 2.5 billion monthly users, wrote in August 2025 that organic clicks to websites were “relatively stable year-over-year”. Reddit told its shareholders on October 29, 2026 that its search referral traffic was “down meaningfully” and that in October it had lost more than a million search referrals a week.
Who pays for the answers
Stack Overflow the company is doing better than Stack Overflow the site. In 2024 it gave Google and OpenAI access to its content through its OverflowAPI. Prosus, which bought the company in 2021, reported revenue of US$189 million for Stack Overflow and GoodHabitz in its year to March 2026, up 11% and “primarily driven by Stack Overflow”, and in the same report US$482 million of impairment losses, primarily on the two, because they “are not performing in line with expectations when they were acquired following the disruption caused by generative AI”. Reddit’s IPO filing disclosed data licensing contracts worth US$203 million over two to three years. The money goes to the platforms. The answers were written by volunteers under a Creative Commons license, and none of these deals sends the next asker back to the page.
For a company that sold answers directly, it went much worse. Chegg sells homework help written by paid experts. Its revenue fell from US$766.9 million in 2022 to US$376.9 million in 2025, and in October 2025 it cut 45% of its staff, writing that “the new realities of AI and reduced traffic from Google to content publishers have led to a significant decline”.
What the next models learn from
The scenario people usually reach for here is model collapse. Shumailov and colleagues showed in Nature in 2024 that a model trained only on text from the previous generation of models loses the rare parts of the distribution first and then degenerates. In their language-model experiment, keeping 10% of the original human data in every generation reduced the damage to a minor loss, and later work found that adding synthetic data to the real data, instead of replacing it, keeps the error bounded (Gerstgrasser et al., 2024; Kazdan et al., 2025). Nobody has a reason to throw away the human text collected before 2023, so the dramatic version isn’t the likely one.
What is happening is slower. Epoch AI estimated the stock of public human-written text at about 300 trillion tokens and expected training sets to reach that size between 2026 and 2032. New text keeps coming, but Pew Research found signs of AI authorship on more than a third of the pages in Common Crawl’s July 2026 crawl that were published after ChatGPT’s release. And the kind of text Stack Overflow produced, a concrete problem with a concrete fix that strangers checked, voted on and corrected, has nearly stopped being written in public. New releases show it most clearly, because nobody can have written about them before.
Python 3.8 collected 630 questions tagged with its version in its first ten months, and Python 3.14, released in October 2025, collected 12. .NET 6 got 2,675 and .NET 10 got 130; iOS 14 got 832 and iOS 26 got 131. Version tags were already used less before ChatGPT, but not at this speed. Whatever the next models know about the releases of 2025, they will learn it mostly from documentation, code, synthetic examples and the conversations people have with models, not from strangers arguing in public about the right answer.
The questions didn’t stop. They moved into chat windows, and those conversations can become training data too: in August 2025, Anthropic started asking its consumer users whether their chats may be used to train future models. The difference is who gets to learn from them. A public answer, with its votes and corrections, was there for every model and every person who came later. A chat log is there for one company.
Getting the data
Stack Overflow’s archive in the data dump is 69 GB, but it keeps each table in its own 7z block, so I downloaded only the 23.6 GB block with the posts and decoded its LZMA2 stream on the fly, without writing the 105 GB of XML to disk. The other sites are small enough to fetch whole. The trend fits, the tags, the Wikipedia series and the robots.txt checks all run from the parsed tables in a few minutes.
Sources
- Stack Exchange, Inc., data dump of 2026-09-30 (CC BY-SA), community re-upload on the Internet Archive: stackexchange_20260930. Counts for October 2026 from the Stack Exchange API.
- Maria del Rio-Chanona, Nadzeya Laurentsyeva and Johannes Wachs, Large language models reduce public knowledge sharing on online Q&A platforms, PNAS Nexus 3(9), 2024.
- Gordon Burtch, Dokyun Lee and Zhichen Chen, The consequences of generative AI for online knowledge communities, Scientific Reports 14, 2024.
- Michael Quinn and Dominik Gutt, Heterogeneous Effects of Generative AI on Knowledge Seeking in Online Communities, Journal of Management Information Systems 42(2), 2025.
- Stack Overflow, 2026 Developer Survey: Community, and What’s new at Stack Overflow: March 2026; the moderators’ strike letter, 2023.
- Wikimedia Foundation: Analytics API; Marshall Miller, New user trends on Wikipedia, 2025; How crawlers impact the operations of the Wikimedia projects, 2025; Wikimedia Enterprise partners, 2026.
- Shayne Longpre et al., Consent in Crisis: The Rapid Decline of the AI Data Commons, NeurIPS 2024, with its domain list.
- Paul Bouchaud and Pedro Ramaciotti, Web Crawler Restrictions, AI Training Datasets & Political Biases, 2025.
- Cloudflare: The crawl-to-click gap, 2025; crawl-to-refer ratio on Radar, 2025; accountable mixed-use AI crawlers, 2026.
- Google: common crawlers and Google-Extended; AI features and your website; AI Overviews launch (May 2024) and expansion (October 2024); Liz Reid, AI in Search is driving more queries and higher quality clicks, 2025; I/O 2026 keynote.
- Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, 2025, and How much of the internet is written with AI?, 2026.
- Stack Overflow press releases on the Google Cloud and OpenAI partnerships, 2024; Prosus, financial results for the year ended 31 March 2026.
- Reddit, Form S-1, 2024, and Q3 2026 shareholder letter.
- Chegg, results for 2022, Form 10-K for 2025 and the October 2025 restructuring.
- Ilia Shumailov et al., AI models collapse when trained on recursively generated data, Nature 631, 2024; Matthias Gerstgrasser et al., Is Model Collapse Inevitable?, 2024; Joshua Kazdan et al., Collapse or Thrive?, ICML 2025.
- Pablo Villalobos et al. (Epoch AI), Will we run out of data? Limits of LLM scaling based on human-generated data, 2024.
- Anthropic, Updates to consumer terms and privacy policy, 2025.