In July 2026, Stack Overflow generated only 1,304 questions, a 99% decrease from the 207,000 questions in March 2014.

Article author and source: MachineHeart

Stack Overflow, known as the "Bible of programmers," with over 24 million questions accumulated in its professional community and serving as the largest programming Q&A corpus on the web, is heading toward decline.

In July 2026, programmers worldwide asked 1,304 questions on Stack Overflow.

In March 2014, this number was 207,000.

A few days ago, developer Daniel Lockyer posted a line chart containing this data on X, remarking, "The end of an era."

The data is real and comes from the Stack Exchange Data Explorer, the official open SQL query interface of Stack Overflow.

The next day, independent developer Pieter Levels (levelsio) shared the image, connecting it to another topic: If no one is asking questions on Stack Overflow and Reddit is overrun by promotional bots, how much human-written content remains on the internet? Without new content, what will the next generation of models be trained on?

These two tweets sparked widespread discussion, with netizens expressing a variety of emotions—some moved, some excited, some saddened, and more.

The Rise and Fall of Stack Overflow

Carefully examine the monthly question volume curve on Stack Overflow above; it can be clearly divided into three segments.

The first segment is growth. In July 2008, during the private beta month, there were only four questions; in August, this jumped to 3,744, and after the public launch on September 15, the number surged to 14,024 for that month. By January 2010, monthly questions had reached 44,923; in 2011, annual questions surpassed one million, and by 2013, they exceeded two million. This is a textbook hockey stick phenomenon.

The second segment is a plateau. The highest monthly figure of 207,000 occurred in March 2014, followed closely by 201,000 in both March 2017 and March 2016. These three peak months are spread across four years, indicating a plateau rather than a sharp spike. Annual totals remained consistently above two million between 2013 and 2017, peaking at 2.18 million in 2016. Afterward, activity eased but did not collapse: 1.88 million in 2018, 1.75 million in 2019, and a rebound to 1.85 million in 2020, with 180,000 alone in April 2020.

The third phase is the downturn. In 2022, there were 1.34 million inquiries; in 2023, 790,000; in 2024, 400,000; and in 2025, 110,000. The total number of inquiries for the entire year of 2025 is less than 53% of the number recorded in a single month—March 2014.

Looking even closer, the compression has become more severe: 17,935 in December 2024, 3,312 in December 2025 (a year-over-year decrease of 81.5%), 2,910 in January 2026, 1,903 in March, 1,229 in May, and 1,072 in June. Excluding the private beta launch month of July 2008, which had only four issues, June’s 1,072 represents the lowest full-month total in the platform’s history. July 2026 saw a slight increase but remained very low at just 1,304.

Totaling 12,000 over the first seven months of 2026, this is approximately the volume of questions from just two days in 2016, based on the average daily rate that year.

Stack Overflow's decline is not entirely due to AI

As shown in the chart, Stack Overflow's decline actually began eight years before ChatGPT's.

The turning point occurred in 2014, when Stack Overflow began systematically improving review efficiency by more aggressively closing duplicate questions and those that violated guidelines. The platform’s intent was to preserve signal-to-noise ratio, at the cost of excluding new users. A 2022 study found that, in a sample of 968 new user posts, 49% experienced at least one of the following: closure, no responses, or unexplained downvotes.

https://www.scitepress.org/PublishedPapers/2022/110811/110811.pdf

This culture has been criticized for a decade. When reporting on the January data, DevClass quoted a developer’s reaction, suggesting that while AI has indeed accelerated the decline, the root cause lies in the community’s long-standing punishment of those who try to participate—people have finally found a tool that won’t say, “That’s a stupid question.”

Below Daniel Lockyer’s tweet, we can also see many similar criticisms:

The issue is that hostility alone isn’t enough to drop annual question volumes to the hundred-thousand level. Before 2022, this number consistently remained above 1.3 million. The real cliff edge resulted from the combination of two factors: ChatGPT reduced the marginal cost of asking questions to zero, and AI assistants within IDEs eliminated the act of asking questions altogether.

Developers no longer need to leave their editor, write out a minimal reproducible example that a stranger can understand, or wait for hours.

Stack Overflow's value was once built on turning one person's question into an asset for everyone. AI assistants do the opposite: they turn public questions back into private conversations.

The platform has not been inactive. In October 2023, it cut 28% of its workforce (following the layoff of 58 employees in May), launched OverflowAI, shifted toward licensing data to companies like OpenAI, and implemented anti-scraping measures to compel platforms back to the negotiating table.

https://stackoverflow.blog/2023/10/16/stack-overflow-company-announcement-october-2023/

Complete full cloud migration by 2025; decommission all approximately 50 servers from the New York (actually New Jersey) data center; retire the Colorado disaster recovery facility in June; year-end article “The Great Unracking” closed this 16-year chapter of physical infrastructure with the subtitle: “Goodbye, thank you for all your bits.”

The site was redesigned and received a new logo in February 2026. CEO Prashanth Chandrasekar has remained in position since 2019, and the company is still owned by Prosus. The $1.8 billion acquisition in June 2021 was completed exactly eighteen months before the steepest phase began.

Not just Stack Overflow

Stack Overflow appears very prominent because it clearly publishes statistics on question data. However, the same curves are appearing in many other places with different slopes.

Chegg is widely regarded as the one that died the fastest. This online education company, founded in 2005 as a textbook rental service and going public in 2013, relied on the same model: turning homework answers into searchable, archived assets, driving traffic through search engines, and charging monthly fees.

After ChatGPT launched, students have no reason to pay this fee. Needham’s survey shows that 62% of students planned to use ChatGPT in November 2024, up from 43% in spring 2023; during the same period, Chegg’s usage dropped from 38% to 30%.

In 2023, it attempted a自救 by creating CheggMate using GPT-4 combined with over 100 million of its own academic content, aiming to prove that professional AI is worth paying for—but students did not accept it.

In May 2025, 248 employees (22%) were laid off and the U.S. and Canadian offices were closed; in October, another 388 employees (45% of remaining staff) were laid off. By the end of April 2026, the stock price was around $1.07, with a market capitalization of approximately $125 million—down about 99% from its peak. It took roughly 39 months for the value to decline from its peak to nearly zero.

The freelance platform is the second example, with a very clear segmentation. Fiverr’s Q2 2026 earnings report released on July 29: revenue of $97.8 million, down 10% year-over-year; marketplace transaction fees fell 15.5% to $63.1 million. Most striking is the number of buyers: 2.7 million annual active buyers as of June 30, a 21.9% decline from 3.4 million a year earlier. However, average spending per buyer rose 15.6% to $368 during the same period, the take rate increased from 27.6% to 28.0%, and the number of clients completing projects over $1,000 grew 13% year-over-year. Management provided the most straightforward explanation on the earnings call: on a TTM basis, writing and translation categories saw the steepest declines, exceeding 24%. Full-year guidance was lowered to $356–372 million, representing a 14% to 17% year-over-year decrease, and they stated that the effects of this transformation may take at least six quarters to materialize.

The same logic applies on Upwork and Freelancer.com.

In late 2025, Upwork had approximately 785,000 active clients, a decrease of about 47,000 from the approximately 832,000 in 2024, marking its largest contraction since going public.

The parent company of Freelancer reported a group GMV of AUD 881.5 million for FY25, a 7.1% year-over-year decrease.

The original rationale for these platforms was to commodify human skills to reduce costs; now, a new technology has further commodified "output." Research in 2024 by Manav Raj and others at the Wharton School of the University of Pennsylvania has quantified this shift: following the advent of ChatGPT, there has been a measurable substitution in freelance demand for writing, translation, and basic coding.

The situation with Wikipedia is more nuanced, as what declined wasn't contributions, but readers. In May 2025, the Wikimedia Foundation updated its bot detection system and reclassified a group of crawlers disguised as humans, revealing that genuine human traffic had decreased by approximately 8% year-over-year.

Marshall Miller, Senior Director of Products, said: Fewer visits mean fewer volunteers growing into editors and fewer individual donors. Wikipedia’s contributions have always been highly concentrated: research shows that 77% of articles come from just 1% of editors. The Foundation’s draft annual plan for fiscal year 2026–2027 states that this is not a temporary phenomenon, but a structural shift.

Quora took a different path: instead of clinging to Q&A, it bet on becoming an AI gateway. It poured its resources into Poe, raising $75 million from a16z in January 2024 to expand this multi-model aggregation platform. A Q&A community ultimately staked its future on something meant to replace Q&A...

Beyond that, the entire information-content business is at stake. A randomized controlled trial involving 1,065 desktop Chrome users found that when AI Overviews appear, off-site organic clicks drop by 39.8%, zero-click searches rise by 34.5%, and there is no measurable improvement in user experience ratings. Ahrefs’ February 2026 study estimated a 58% decline in click-through rates for top-ranking pages. The consequences have already manifested in layoffs and site closures: Bauer Media Group announced in April 2026 the shutdown of its German digital subsidiary and the elimination of 160 jobs; the nonprofit reference site Overfishing.org, operational for 21 years, has shut down.

In summary, the pattern is this: categories where machine-generated answers are functionally equivalent to the original collapse the fastest. This includes reference materials, technical documentation, definitional content, and how-to tutorials. Conversely, work that requires long-term relationships, contextual understanding, and accountability remains largely unaffected. Fiverr’s financial reports also confirm this: low-value transactions are declining, while complex projects exceeding $1,000 are growing.

Worse still, the experts are retreating.

If users were simply asking questions from a different location, the situation would still be straightforward. However, a working paper published in July 2026 by Kenny Ching at the Auckland University Business School highlights a more problematic mechanism.

He tracked the behavior of 24,304 contributors on Stack Overflow over 17 months and found that, starting in 2022, the rate at which high-reputation users left the platform began to accelerate, gradually catching up to the rate of low-reputation users, who had always left more quickly. In other words, it wasn’t just askers leaving—answerers were departing too, and those who had spent years building expertise and earning recognition in the community were leaving with the greatest determination.

Ching named this mechanism "signal compression." He explained that the reason these people left is that their hard-earned expertise had become indistinguishable from the responses of a chatbot.

When everyone can use AI to produce something that looks decent, putting in genuine effort is no longer a signal that earns recognition. In the interview, he extended this logic beyond platforms: the same dynamic is unfolding in classrooms, companies, and scientific communities, with the long-term risk being that “AI may sever the development of future human expertise by undermining the incentives to demonstrate genuine effort.”

This aligns with Stack Overflow’s own survey data. A 2025 survey covering 177 countries and approximately 49,000 developers found that 84% of respondents are using or plan to use AI tools (up from 76% in 2024), but only 29% trust the accuracy of AI outputs, 46% explicitly expressed distrust, and just 3% reported “high trust”—among experienced developers, this figure was 2.6%.

https://survey.stackoverflow.co/2025

The biggest complaint isn't that AI can't write code, but that it produces code that's "almost right, but not quite"—66% of people are troubled by this, and 45% say they spend significant time debugging AI-generated code. When they don't trust AI's answers, 75.3% say they turn to a human for help.

The problem is that person is leaving.

Conclusion

Now let’s put several things together and take a look.

Stack Overflow's existing corpus is openly available under the CC BY-SA license, making it one of the most extensively harvested technical corpora. Models trained on it are now precisely why new contributions have ceased.

The existing corpus is timestamped. Answers on Stack Overflow about jQuery, Python 2, and Angular.js are gradually becoming outdated. New frameworks, breaking changes in new versions, compilation errors in a library used by only three thousand people under a specific CUDA version—this kind of knowledge used to be documented on public web pages by the first person to encounter the issue, but now it occurs in private conversations between an individual and a model, and disappears once resolved.

The platform's answer is "a trusted human intelligence layer." Chandrasekar repeatedly emphasized this term: AI carries risks of misdirection and may lack complexity and relevance, making a curated knowledge base and responsible data usage essential. This is logically sound, but it requires people willing to continuously supply that layer; however, Ching's data shows that these individuals are declining.

There is also a reverse possibility. Models don't necessarily need forum-style Q&A to learn new knowledge—they can learn from code repositories, official documentation, and user interaction logs.

In 2026, Google opened a document API covering approximately 40 million documents with 24-hour reindexing, clearly designed for machine consumption. Framework maintainers are also catching up: Next.js documentation now includes a "Common Mistakes" section, written in a style nearly identical to Stack Overflow answers; the Svelte team even trained their own chatbot using documentation and source code. The places where knowledge is produced may simply be shifting, not disappearing.

But there is an unanswered distinction between these two prospects: forum-based Q&A has a byproduct—it is public, searchable by third parties, and not owned by any single company. Documentation is written by maintainers, while interaction logs belong to the model providers. When programmers’ first response to a problem shifts from “search to see if anyone asked this before” to “just ask,” what is lost is not merely traffic to a website, but a public record accessible and editable by anyone.

In the comments under Lockyer’s tweet, developer Mayberry offered a somewhat optimistic prediction: blogs, niche communities, and personal accounts may regain importance, because “individuals with reputation and sites with a reputation for original content will become valuable again.”