What AI has done to Stack Overflow, in one chart

A contributor extracted Stack Overflow's public stats and produced a chart. The results speak for themselves: ChatGPT has taken over.

The concrete case

A user posted on data.stackexchange.com (the public SQL query interface for Stack Overflow data) a query that tracks the number of questions asked per month on the site from its creation to mid-2026. The graph is linked on Hacker News as a headline on July 18, 2026, and it does what few opinion pieces manage to do: show, without any detour, what AI has changed.

The curve gently rises from 2008 to 2013, plateaus around 2013-2020 with a slight decline already underway, and then a sharp drop starting from late 2022 - exactly when ChatGPT was released (November 30, 2022) - and the decline continues without recovery. In 2026, the monthly volume of questions has dropped to a level not seen since the site's early years.

The code behind the graph

For the curious, the published query is of the type:

SELECT
 DATEFROMPARTS(YEAR(CreationDate), MONTH(CreationDate), 1) AS Month,
 COUNT(*) AS Questions
FROM Posts
WHERE PostTypeId = 1
GROUP BY YEAR(CreationDate), MONTH(CreationDate)
ORDER BY Month;

PostTypeId = 1 filters for questions (answers have PostTypeId = 2). The data.stackexchange.com tool allows anyone to verify the result in thirty seconds, without installing anything - a model of data openness that few platforms still offer.

The why

For a senior developer, the question "why such a drop?" has an obvious answer: when you're stuck on a bug or syntax, you no longer search for "python typeerror int str" on Google and open the first Stack Overflow response from a 2014 thread marked as a duplicate. You open Claude, ChatGPT, or Copilot and ask the question directly, with your context, your piece of code, your stack trace. The answer arrives personalized, in French if you want, without the moderator wars about closing for being a duplicate.

It's faster, it's gentler (no one accuses you of RTFM), and it works often enough - except when the model hallucinates. Again, the senior user knows this and double-checks; the junior, often, does not. This is one of the current debates in teams.

What we lose along the way

The real cost of this shift is not immediately visible. Stack Overflow was not just an answer engine: it was a corpus that Google indexed, that recruiters consulted (to assess a profile), that newcomers read to learn how to properly formulate a technical question. This corpus is now shrinking, while the AI models that were trained on it continue to benefit from it without contributing back. Potentially problematic feedback loop: what will the models of 2028 be trained on?

There is also the aspect of traceability. A Stack Overflow answer, you know who wrote it, when, with what community score. An LLM's answer is an opaque aggregation. For production code, this traceability is valuable.

Key takeaways

The graph is just one data point among others: Stack Overflow has announced several waves of layoffs since 2023, motivated in part by the drop in traffic (see notably the blog post by its CEO Prashanth Chandrasekar in October 2023 announcing the first wave). But the image itself condenses three years of evolution in developers' habits. The question to ask in 2026, as a team, is no longer "should we use LLM". It's: how not to let die the common goods (documentation, public Q&A, tutorials) that made our profession possible?