Stack Overflow Is Dying, New Questions Are Fewer Than During the Beta Period

marsbitPublished on 2026-08-17Last updated on 2026-08-17

Abstract

The article discusses the decline of Stack Overflow, the prominent programming Q&A community, noting that new question submissions in July 2026 dropped to just 1,304—far below its peak of 207,000 in March 2014. Official data reveals a steady decline in activity since 2014, exacerbated by the rise of AI coding assistants like ChatGPT, which offer instant, private solutions and reduce the need for public forum posts. While a restrictive community culture that often penalized new users contributed to the downturn, the widespread adoption of AI tools has dramatically accelerated the platform's stagnation. Similar declines are observed in other knowledge-based platforms like Chegg, Fiverr, and Wikipedia, where AI-generated content is displacing human-generated queries and contributions. Research indicates that not only are fewer users asking questions, but high-reputation experts are also leaving the platform at increasing rates—a phenomenon termed "signal compression," where the value of demonstrated expertise diminishes when AI can produce superficially competent answers. This threatens the formation of future human expertise. Although AI models have been trained extensively on Stack Overflow's open corpus, the reduction in new, timely public Q&A content may limit the knowledge available for future AI training. Some suggest that trustworthy human-curated knowledge layers, blogs, and niche communities may regain importance as sources of reliable, attributed information. The sh...

Stack Overflow, hailed as the "programmer's bible," a professional community that has amassed over 24 million questions and the largest programming Q&A corpus on the internet, is dying.

In July 2026, programmers worldwide asked a total of 1304 questions on Stack Overflow.

In March 2014, that number was 207,000.

The other day, developer Daniel Lockyer posted a line chart containing this data on X, lamenting "the end of an era."

The data is genuine, sourced from Stack Overflow's official public SQL query interface, the Stack Exchange Data Explorer.

The next day, independent developer Pieter Levels (levelsio) reposted this chart, linking it to another topic: if no one asks questions on Stack Overflow anymore, and Reddit is overrun by promotion bots, how much real human-written content remains on the internet? With no new content, what will the next generation of models train on?

These two tweets sparked widespread discussion. Netizens' attitudes varied greatly, ranging from lamentation and excitement to sorrow... and more.

The Rise and Fall of Stack Overflow

Looking closely at the monthly question volume curve for Stack Overflow above, we can clearly divide it into three segments.

The first segment is growth. In July 2008 during the private beta, there were only 4 questions; 3744 in August; and after the public beta on September 15, it jumped to 14,024 that month. By January 2010, it was already 44,923 per month. The annual question count surpassed one million in 2011 and two million in 2013. This is a textbook hockey stick phenomenon.

The second segment is the plateau. The single-month peak stalled at 207,000 in March 2014, followed by 201,000 in March 2017 and March 2016. The three highest months spread over four years indicate a plateau, not a sharp spike. The annual total remained stable above two million between 2013 and 2017, peaking at 2.18 million in 2016. It began to loosen afterward but didn't collapse: 1.88 million in 2018, 1.75 million in 2019, rebounding to 1.85 million in 2020, with April 2020 alone seeing 180,000 questions.

The third segment is the decline phase. 1.34 million for the whole year in 2022, 790,000 in 2023, 400,000 in 2024, and 110,000 in 2025. The entire 2025 question volume is less than 53% of the single month of March 2014.

Looking more granularly, the compression is even more dramatic. 17,935 in December 2024, 3,312 in December 2025 (an 81.5% year-on-year drop), 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 with its 4 questions, June's 1,072 is the lowest full month in the site's entire history. July 2026 saw a slight rebound but remained low at only 1,304 questions.

The first seven months of 2026 total about 12,000 questions. Based on the daily question rate of 2016, this is roughly the amount from about two days that year.

Stack Overflow's Decline Is Not Entirely Due to AI

From the graph, Stack Overflow's decline actually started eight years earlier than ChatGPT.

The inflection point appeared in 2014 when Stack Overflow began systematically improving moderation efficiency, more actively closing duplicate questions and queries that "did not meet standards." The site's intention was to preserve the signal-to-noise ratio, at the cost of blocking new users. A 2022 study showed that in a sample of 968 new user posts, 49% encountered at least one of: closure, no replies, or unexplained downvotes.

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

This culture has been criticized for a decade. Devclass, reporting on the January data wave, quoted a developer's reaction, roughly stating that AI indeed accelerated the decline, but the root lies in the community long punishing those trying to participate, and people finally have a tool that doesn't say "your question is stupid."

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

The problem is that a hostile culture alone isn't enough to drive annual question volume down to the tens of thousands. Before 2022, that number consistently stayed above 1.3 million. The real cliff is the result of two things combined: ChatGPT reduced the marginal cost of asking to zero, and then AI assistants in IDEs made the act of asking disappear altogether.

Developers no longer need to leave the editor, don't need to formulate the problem into a minimal reproducible example a stranger can understand, and don't need to wait hours.

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

The site hasn't been idle. In October 2023, it laid off 28% (following 58 layoffs in May), launched OverflowAI, pivoted to data licensing for companies like OpenAI, and erected anti-scraping walls to force platforms back to the negotiating table.

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

In 2025, it fully migrated to the cloud, decommissioning all ~50 servers from the New Jersey data center (its NYC location was actually NJ), retired the Colorado disaster recovery site in June, and ended the year with the blog post "The Great Unracking," giving this sixteen-year physical infrastructure history a subtitle: "So long, and thanks for all the bits."

In February 2026, the site underwent a redesign and got a new logo. CEO Prashanth Chandrasekar, in place since 2019, remains, and the company is still owned by Prosus. The $1.8 billion acquisition in June 2021 was completed exactly eighteen months before the steepest decline began.

Not Just Stack Overflow

Stack Overflow looks particularly conspicuous because it clearly publishes question statistics. But in reality, similar curves are appearing in many places with different slopes.

Chegg is widely acknowledged as the fastest to die. This online education company, which started with textbook rentals in 2005 and went public in 2013, relied on the same model: turning homework answers into searchable stock assets, using search engines for traffic, and charging a monthly fee.

After ChatGPT launched, students had no reason to pay anymore. Needham's research showed that in November 2024, 62% of students planned to use ChatGPT, up from 43% in spring 2023; Chegg dropped from 38% to 30% over the same period.

In 2023, it attempted a self-rescue, creating CheggMate using GPT-4 plus its over 100 million academic data points, trying to prove a specialized AI was worth paying for, but students didn't buy it.

In May 2025, it laid off 248 employees (22%) and closed US and Canada offices; in October, it laid off another 388 (45% of remaining staff). By late April 2026, its stock hovered around $1.07, with a market cap of about $125 million in May, roughly 99% evaporated from its peak. From peak to near zero, about 39 months.

Freelance platforms are the second sample, and they show clear stratification. Fiverr's Q2 2026 earnings released on July 29: revenue of $97.8 million, down 10% year-over-year; marketplace revenue from matching transactions fell 15.5% to $63.1 million. The real eye-catcher is buyer count: 2.7 million annual active buyers as of June 30, down 21.9% from 3.4 million a year earlier. But average spend per buyer rose 15.6% to $368 over the same period, take rate increased from 27.6% to 28.0%, and the number of clients completing projects over $1000 grew 13% year-over-year. Management gave the most straightforward annotation on the earnings call: on a TTM basis, the writing and translation categories saw the largest declines, over 24%. Full-year guidance was lowered to $356-$372 million, down 14% to 17% year-over-year, and they stated this transformation might take at least six quarters to show results.

The same logic repeats on Upwork and Freelancer.com.

Upwork's active client count at the end of 2025 was about 785,000, down from about 832,000 in 2024—a drop of about 47,000, its largest contraction since going public.

Freelancer's parent company's FY25 group GMV was 881.5 million AUD, down 7.1% year-over-year.

The raison d'être of these platforms was to commoditize human skills to lower costs; now there's a technology that further commoditizes "output." Research from Manav Raj et al. at the Wharton School of the University of Pennsylvania in 2024 already quantified this displacement: measurable substitution occurred in freelance demand for writing, translation, and basic coding after ChatGPT.

Wikipedia's situation is subtler because it's losing readers, not contributors. In May 2025, the Wikimedia Foundation updated its bot identification system, reclassifying a batch of human-impersonating crawlers, and discovered real human traffic had fallen by about 8% year-over-year.

Senior Director of Product Marshall Miller indicated: fewer visits mean fewer volunteers growing into editors and fewer individual donors. Wikipedia's contributions were already extremely concentrated: research shows 77% of articles were written by 1% of editors. The Foundation's 2026–2027 fiscal year annual plan draft states the judgment: this is not temporary, it's a structural shift.

Quora took another path: instead of guarding Q&A, build an AI gateway itself. It poured resources into Poe, securing $75 million in funding from a16z in January 2024 to expand this multi-model aggregation platform. A Q&A community ultimately bet its future on something that replaces Q&A...

One step further out is the entire informational content business. A randomized controlled experiment covering 1065 desktop Chrome users found that when AI Overviews appear, off-site organic clicks drop 39.8%, zero-click searches rise 34.5%, and users show no measurable improvement in experience ratings. Ahrefs' February 2026 research calculated a click-through rate drop of up to 58% for top-ranking pages. Consequences have already manifested in layoffs and site closures: Bauer Media Group announced in April 2026 the closure of its German digital subsidiary, cutting 160 positions; the 21-year-old public reference site Overfishing.org shut down.

To summarize the pattern: the faster the collapse, the more the category is one where "machine-generated answers are functionally equivalent to the original." Reference materials, technical documentation, definitional content, how-to tutorials—all are in this set. Conversely, work requiring long-term relationships, context, and accountability shows little change. Fiverr's earnings also illustrate this: low-value transactions are disappearing, while complex projects over $1000 are growing.

More Troublesome: Experts Are Retreating

If users were merely asking questions elsewhere, the story would be simpler. But Kenny Ching from the University of Auckland Business School, in a working paper released in July 2026, pointed out a more problematic mechanism.

He tracked the behavior of 24,304 contributors on Stack Overflow over 17 months, finding that since 2022, the departure rate of high-reputation users began accelerating, gradually catching up with the faster-departing low-reputation users. In other words, it's not just question-askers leaving, but also answerers, and the more people who spent years accumulating expertise and earning recognition in the community, the more resolutely they are leaving.

Ching named this mechanism "signal compression." His explanation: these people are leaving because the expertise they painstakingly earned is no longer distinguishable from a chatbot's answer.

When everyone can use AI to produce something that looks decent, the act of putting in genuine effort ceases to be a recognizable signal. In an interview, he extended this logic beyond platforms: the same thing is happening in classrooms, companies, and scientific communities. The long-term risk is that "by destroying the incentive to demonstrate genuine effort, AI may cut off the formation of future human expertise."

This aligns with Stack Overflow's own survey data. The 2025 survey covering about 49,000 developers from 177 countries showed 84% of respondents using or planning to use AI tools (76% in 2024), but only 29% trusted the accuracy of AI output, 46% explicitly distrusted it, and only 3% "highly trusted" it—2.6% among senior developers.

https://survey.stackoverflow.co/2025

The biggest complaint wasn't that AI can't code, but that its output is "almost right, but not quite," with 66% troubled by this and 45% saying they spend significant time debugging AI code. And when they don't trust an AI answer, 75.3% said they would ask a person.

The problem is, that person is leaving.

Conclusion

Now let's synthesize a few things.

Stack Overflow's stock corpus is openly licensed under CC BY-SA, making it one of the most thoroughly scraped technical corpora. The models that ingested it are now the reason for the halt in new contributions.

Stock corpus has timestamps. Answers about jQuery, Python 2, and Angular.js on Stack Overflow will gradually become outdated. Knowledge about new frameworks, breaking changes in new versions, compilation errors of a library used by only three thousand people under a specific CUDA version... This type of knowledge was previously written on a public webpage by the first person to step on the landmine. Now it happens in a private conversation between someone and a model, disappearing after resolution.

The site's answer is the "trusted human intelligence layer." Chandrasekar repeatedly uses this term: AI has risks of misleading, may lack complexity and relevance, so a path relying on curated knowledge bases and responsible data use is critical. This is logically sound, but it requires people willing to continuously supply that layer; but Ching's data says those people are diminishing.

There's also a reverse possibility. Models may not need forum-style Q&A to learn new knowledge: they can learn from code repositories, official documentation, interaction logs with users.

Google opened a documentation API in 2026 covering about 40 million documents with 24-hour re-indexing, clearly meant for machines to read. Framework maintainers are also catching up: Next.js documentation features "Common Errors" sections, written almost like Stack Overflow answers; the Svelte team simply trained its own chatbot based on docs and source code. The site of knowledge production might just be migrating, not disappearing.

But there's an unanswered difference between these two prospects. Forum-style Q&A has a byproduct: it's public, searchable by third parties, and doesn't belong to any single company. Documentation is written by maintainers, interaction logs belong to model vendors. When a programmer's first reaction to a problem shifts from "search if someone asked" to "ask the AI," what's lost isn't just a website's traffic, but a public record anyone can consult and correct.

In the discussion under Lockyer's tweet, developer Mayberry offered a less pessimistic prediction: blogs, niche communities, and personal accounts might regain importance because "people with reputations and sites with reputations for original content will become valuable again."

This article is from WeChat public account "Almost Human" (ID: almosthuman2014), author: Panda

Trending Cryptos

Related Questions

QAccording to the article, what is the key data point highlighting Stack Overflow's decline in user activity?

AIn July 2026, users posted only 1,304 new questions on Stack Overflow, which is lower than the number of questions posted during its private beta month in July 2008 (except for the very first month with 4 questions). This is a drastic drop from a peak of 207,000 monthly questions in March 2014.

QWhat are the two main phases of decline discussed, and what primary factor is blamed for the earlier phase starting around 2014?

AThe article discusses two phases of decline. The first phase, beginning around 2014, is primarily attributed to Stack Overflow's increasingly strict moderation culture and policies that aggressively closed 'duplicate' or 'low-quality' questions, which created a hostile environment for new users and stifled participation.

QWhat is the 'signal compression' effect described in the article, and why is it problematic?

A'Signal compression' refers to the phenomenon where the effort and expertise of high-reputation contributors on platforms like Stack Overflow become indistinguishable from AI-generated answers. This devalues their hard-earned professional standing, removing the incentive to contribute and leading experts to leave the platform, which risks cutting off the formation of future human expertise.

QBesides Stack Overflow, what other types of online platforms or businesses does the article mention as being significantly impacted by generative AI?

AThe article mentions several other impacted platforms: educational Q&A services like Chegg, freelance marketplaces for writing/translation/coding (e.g., Fiverr, Upwork), Wikipedia (seeing a decline in human readers), Q&A sites like Quora, and general informational websites/SEO-based content businesses.

QWhat potential future for knowledge sharing does the article suggest, as an alternative to centralized Q&A platforms like Stack Overflow?

AThe article suggests knowledge production might migrate to other venues. This includes official documentation enhanced with troubleshooting guides (like Next.js), framework-specific AI chatbots (like Svelte's), and a potential resurgence of blogs, niche communities, and personal accounts, where reputation and original content from trusted individuals become valuable again.

Related Reads

The Tokenization of Gold: A Financial Revolution Driven by Regulation

The Tokenization of Gold: A Regulation-Driven Financial Revolution In early 2025, anticipation of US tariffs triggered a rush to withdraw physical gold from the Bank of England, exposing the friction in London's $1.2 trillion gold market. While daily trading of $737 billion occurs via ownership certificates without moving bullion, demand for physical metal revealed a pricing split based on deliverability. This article explores why the UK's Financial Conduct Authority (FCA) is developing rules for tokenized gold, focusing regulation squarely on the ledger system. London's gold market, cleared by four major banks, performs four key functions: physical storage, quality verification, ownership registration, and credit provision. Tokenization excels at the third—offering cheap, precise ownership records—but fundamentally challenges the fourth by potentially rendering banks' credit intermediation obsolete. Currently, most gold is held in "unallocated" accounts, where investors are unsecured creditors of banks, trading speed for credit risk. Tokenized gold combines the instant transfer of unallocated accounts with the direct ownership claim of allocated bullion, eliminating bank risk. The FCA is prioritizing its use as collateral, where speed is critical, unlocking this vast pool of currently illiquid assets for margin calls. This shift is partly driven by post-2008 banking regulations like Basel III's Net Stable Funding Ratio, which penalized banks for holding unallocated gold. The push now comes from the清算 banks themselves, like HSBC, which have launched proprietary tokenization systems. The crucial regulatory decision will be whether these tokens are allowed to circulate outside their creators' closed ecosystems. True adoption will be visible in LBMA data: high overall trading volume paired with a decline in settlements through the traditional central ledger, as transactions move to instant blockchain transfers. Ultimately, tokenization doesn't replace the physical vaults, security, and assayers but could elevate their importance, shifting their revenue to standardized service fees. The revolution lies not in moving the gold, but in reinventing how its ownership is recorded and transferred.

marsbit8m ago

The Tokenization of Gold: A Financial Revolution Driven by Regulation

marsbit8m ago

Yield Leverage and Liquidity Leverage: The STONKBROKER Mechanism is Very Suitable for the RWA Scenario

The article argues that the STONKBROKER mechanism is highly suitable for Real-World Asset (RWA) tokenization scenarios due to its two key leverages: **Profit Leverage** and **Liquidity Leverage**. **Profit Leverage:** Traditional RWA models simply tie tokens to real-world asset yields. STONKBROKER's model is more engaging by allowing NFT holders to earn not only the underlying real-world cash flows but also additional yields generated from the on-chain system's own operations and transaction activity (like fees and slippage). This creates a dual income stream. **Liquidity Leverage:** Standard RWA asset bundles often suffer from poor liquidity. STONKBROKER's mechanism, which allows fixed-ratio swapping between NFTs and a platform's fungible token (and vice versa), solves this. It provides real-time market pricing and flexible exit options for each "RWA asset" NFT by unifying NFT liquidity with ERC-20 token liquidity via an AMM pool. This enhances capital efficiency and attracts more trading activity. The article further speculates on applying this model to broader RWAs (e.g., real estate, bonds). The core idea is to create a more dynamic, liquid, and profitable ecosystem where ordinary users can participate, combining real-world yields with the speculative and transactional energy of a crypto-native system. The author mentions being inspired by the launch of a project called @TheCardWall, which tokenizes high-grade physical trading cards.

marsbit29m ago

Yield Leverage and Liquidity Leverage: The STONKBROKER Mechanism is Very Suitable for the RWA Scenario

marsbit29m ago

Deciphering TapeOut: A Child Built a CPU On-Chain

**Title: Understanding TapeOut: A Kid Built a CPU On-Chain** **Summary:** The article explores a remarkable experiment where a young developer named Blonskr has used blockchain technology to build a functional, on-chain CPU. Named "Behemoth," this CPU is based on the 1971 Intel 4004 processor specification. It runs on the BNB Chain, uses block production as its clock (running at a glacial 2.22 Hz), and is perhaps the slowest CPU in existence, yet it represents a novel and intriguing concept in the on-chain space. The core concept is the TapeOut protocol, named after the final "tape out" stage in semiconductor manufacturing. It treats the blockchain as a silicon wafer, with tokens representing fundamental logic components. The protocol's atomic unit is an ERC-1155 token representing a single NAND gate, the fundamental building block of digital logic. Users can visually design circuits on a "canvas," connecting these logic tokens. Once a design is finalized, the underlying tokens are burned, and an ERC-721 NFT representing the permanent, immutable circuit is minted. Blonskr used 2,300 such tokenized NAND gates to construct the Behemoth CPU. Key features of this and any TapeOut circuit include **permanent operation** (running as long as the chain exists), **free, permissionless evaluation** (anyone can query it without gas), and **safe composability** (circuits cannot call external contracts or modify state, making them inherently safe to use). This enables a new form of open, collaborative hardware design, where anyone can build upon and integrate components created by others. While highly niche, TapeOut reimagines the foundations of computation using blockchain's properties of permanence and openness. It pushes the boundaries of what can be built on-chain, potentially enabling future applications like custom mining circuits or decentralized, specialized processors. The experiment demonstrates an exciting combination of extreme technical vision and creative curiosity.

marsbit1h ago

Deciphering TapeOut: A Child Built a CPU On-Chain

marsbit1h ago

Novig Sues Wisconsin as Sports Prediction Market Dispute Heats Up

Novig, a sports event prediction market operator, has filed a lawsuit against Wisconsin's attorney general. The legal dispute centers on whether prediction markets fall under federal derivatives regulations or state gambling laws. The lawsuit, filed in a U.S. district court, seeks a declaratory judgment after Novig began offering event contracts to Wisconsin residents. This case is part of a broader conflict. In April, Wisconsin sued other prediction market platforms like Kalshi and Polymarket, alleging their sports-related contracts violate state gambling laws. Novig argues its contracts are "swaps" regulated exclusively by the federal Commodity Futures Trading Commission (CFTC), under the Commodity Exchange Act. Wisconsin contends they are sports betting. Federal courts have issued conflicting rulings on similar cases, creating legal uncertainty. The outcome is significant as prediction markets are a fast-growing sector, with trading volume reaching $9.5 billion on a single day in August. These markets are also increasingly intertwined with cryptocurrency infrastructure. Simultaneously, Novig is pursuing a nationwide legal strategy, having filed similar lawsuits in four other states. The company also recently secured a marketing partnership with the New York Mets, marking the first such deal with an MLB team. The Wisconsin case's resolution will influence whether prediction markets can operate as national financial products or remain constrained by individual state laws.

cryptonews.ru1h ago

Novig Sues Wisconsin as Sports Prediction Market Dispute Heats Up

cryptonews.ru1h ago

Trading

Spot

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of S (S) are presented below.

活动图片