AL+

RO EN

Email us

Is Stack Overflow dying? What the OpenAI deal showed developers

The 2024 Stack Overflow and OpenAI deal showed who controls public developer answers. Two years on, the knowledge a company can rely on is what it writes down.

. By . Careers and industry

Citiți articolul în română

On 6 May 2024, Stack Overflow and OpenAI announced an API partnership: OpenAI would use Stack Overflow's OverflowAPI to improve its models and surface Stack Overflow content in ChatGPT with attribution, and Stack Overflow would build its own AI features on OpenAI models. The press release counted "more than 59 million questions and answers" written by developers over 15 years.

My position, two and a half years on: the deal settled a question many developers had never asked. The answers they wrote in public were, in every practical sense, an asset the platform controlled and could license. Developers should assume that anything they write in public will train a model. Companies should stop treating public Q&A as their knowledge base and write down what their own systems do, because that is the only technical knowledge they actually control.

What the deal said, and what users found out

The announcement was framed as a benefit to developers: vetted data for AI tools, attribution inside ChatGPT, and revenue Stack Overflow would "reinvest in community-driven features". The reaction on Hacker News that day was less warm. One commenter asked where to send a bank account number for residuals. Another described a profession whose open habit of sharing would be used to make it obsolete.

Some users went further and overwrote their best answers with protest text. The Register reported on 9 May 2024 that one developer's edits were reverted within an hour and his account suspended for seven days, with a moderator email explaining that posts "become part of the collective efforts of others who have contributed to that content". The same article quoted the terms of service: content is "perpetually and irrevocably licensed to Stack Overflow on a worldwide, royalty-free, non-exclusive basis", and "you cannot revoke permission for Stack Overflow to publish, distribute, store and use such content."

The distinction is between licence survival and deletion rights. The current public terms, updated 13 November 2025, grant continuing, irrevocable rights to use and commercially exploit contributed content even after its removal. Public posts also carry CC BY-SA licences. That does not establish that every deletion, anonymisation or personal-data request is unavailable. The reverted edits above are a reported protest case, not a complete account of those rights. The platform controls the structured feed: in September 2024 it described data licensing as recurring subscription revenue.

What happened to the questions

The coverage at the time treated the deal as a licensing story or a protest story. The more important story was already underway: people were asking fewer questions.

By May 2025, Gergely Orosz, using Stack Exchange Data Explorer figures shared by Marc Gravell, showed that monthly question volume was back to the level of Stack Overflow's launch in 2009. By January 2026, DevClass reported 3,862 questions in December 2025, against more than 200,000 a month in early 2014. Much of that asking appears to have moved to AI assistants: in the 2025 Stack Overflow Developer Survey, 84% of respondents said they use or plan to use AI tools in their development process.

The business did not collapse with the traffic. Prosus, which bought Stack Overflow for $1.8 billion in 2021, reported revenue growth for the half-year to September 2025 for Stack Overflow and GoodHabitz together, driven primarily by Stack Overflow, according to the same DevClass piece. DevClass describes a business model based first on recruitment and later on paid private versions of the company's collaboration software; to that the company has added data licensing, which its own blog calls a recurring revenue model. None of it depends on developers asking public questions.

So the answer to "is Stack Overflow dying?" depends on which Stack Overflow you mean. The company appears to be doing fine. The public commons, the place where a new problem got a new answer from a person who had solved it, is thin. The archive that remains is being sold, and the flow of new questions that kept it current, and with it the flow of new answers, has shrunk to a few thousand a month.

What developers lost

The loss is easy to understate because the old answers are still there, and the models trained on them answer the old questions well. What is weakening is one mechanism that produces answers for new questions: a new framework version, a new cloud service, a new failure mode. Models can synthesize, infer and use tools, but claims about new APIs and failures still need reliable evidence. Fewer public answers remove one useful source for verification. When the people who used to write it down now ask a model instead, the next library release gets fewer public answers, and the next model has less to learn from.

The survey shows developers know this. More of them distrust the accuracy of AI tools (46%) than trust it (33%), and the most common reason they would still ask a person, chosen by 75%, is "when I don't trust AI's answers". The fallback they name is a person who knows. The public place where you used to find that person is the thing that shrank.

Nothing was taken that was not already given

First, nothing was taken that had not already been given. The content was CC BY-SA from the start, so anyone could reuse it under that license. A paid deal that requires attribution is arguably better for contributors than the unlicensed scraping that was happening anyway; Stack Overflow said in September 2024 that all its OverflowAPI contracts require attribution.

Second, the decline did not start with OpenAI. Orosz dates the first drop to around 2014, when moderation became stricter and many questions were closed faster, and notes that the site could have faded even without LLMs. He also shows the steepest fall starting in November 2022, when ChatGPT arrived. On that reading, the deal did not cause the collapse in questions; AI assistants accelerated a decline in a community that had already become hard for newcomers.

Third, Stack Overflow disputes that the commons is dying. On 29 October 2026 it published The reports of Stack Overflow's death are greatly exaggerated, citing its 2026 Developer Survey: 64% of respondents said they visit the site less often for simple questions, but developers also report asking more advanced questions and visiting to verify AI-generated answers. The same post presents Stack Overflow for Agents, "an API-first knowledge exchange where agents can share knowledge and verify solutions", and describes Stack Internal as a "governed AI context layer for companies".

That is a fair correction to the idea that nobody uses the site. It does not change the measured fact: the cited January 2026 report counted 3,862 questions in December 2025, against more than 200,000 at the peak. "More advanced questions" is the company's description of the questions that remain, not a measure of how many there are. And using the site to check an AI answer means reading the archive, not adding to it. The 2026 survey discusses trust through situations in which respondents accept AI answers. That framing is not directly comparable to the 2025 trust/distrust split; the figures above describe 2025, not a measured trend into 2026.

I accept the first two points, and part of the third. They change the cause, and the picture of who still visits, not the conclusion. Whether the commons was sold, scraped or simply abandoned, a company that relies on it for answers about its own work is relying on something it does not control and whose supply of new public questions has declined.

Write down what your own systems do

Public Q&A never knew the things that cause most trouble inside a company: why the invoice export runs at 2 a.m., which customer depends on the old API field, why the team chose one queue over another. Those answers lived in people's heads and in chat threads. When those people left, the answers left with them.

The most telling evidence comes from Stack Overflow itself. Its internal product, Stack Internal (formerly Stack Overflow for Teams), opens its pitch with "The team who built the legacy system have left, with no handover" and "Crucial decisions are getting buried in chat threads." The company that built the public commons now sells private knowledge bases, because that is where the knowledge a business depends on has to live.

You do not need a product to start. Three kinds of document carry most of the value:

  • Decision records (often called ADRs). A short note for each significant choice: the context, the options, what was decided and why. Months later, this is the only thing that explains why the code looks the way it does.
  • An operating manual. How to deploy, roll back, restore a backup and rotate a credential, written so that someone new can do it without calling the previous developer.
  • Handover notes. What is unfinished, what is fragile and what is known to be wrong. We set out what a complete one contains in what a software handover should contain.

AI assistants make these documents more useful, not less. A model pointed at a repository with good decision records can answer questions about that system. Pointed at a repository with none, it guesses.

Where public documentation is enough

For an open-source project, public discussion is the point: issues, forum threads and documentation should be public, and the trade-off is that they will be used for training. For a small team on a common stack building a standard product, public documentation and an AI assistant will answer most daily questions, and an internal knowledge base would be overhead. And for questions about a language or framework itself, the vendor's documentation remains the primary source, better than a forum or a model.

The argument applies to everything specific to your business: your data, your integrations, your decisions and your failures. Nobody else will write that down, and a model cannot be assumed to know it.

Treat written knowledge as part of the software

The Stack Overflow deal did not end developer knowledge sharing, but it showed who controls public answers, and the following two years showed how quickly the flow of new ones could shrink. Treat anything you post in public as training data. Treat your company's own written knowledge as part of the software, and maintain it like code. If no one on your team owns that work today, it belongs in a maintenance plan.

Sources