News · 2026-08-08

Stack Overflow took 1,490 questions in July

Stack Overflow took 1,490 new questions in July 2026. In July 2025 it took 6,414. In July 2014 it took 176,610. Those counts come from a direct query against the public Stack Exchange API, and they describe a 118-fold collapse in the volume of public programming questions on the site that taught a generation of developers -- and, through its archives, taught the models now answering those questions instead.

Key facts

• July 2026: 1,490 questions asked, about 48 per day.
• July 2025: 6,414 questions -- a 77% year-over-year decline.
• July 2014: 176,610 questions, near the site's peak.
• Method: counts retrieved from the Stack Exchange API v2.3 questions endpoint, filtered by creation date, site=stackoverflow.

These numbers are not an estimate or a scrape. The Stack Exchange API exposes a total count of questions created between two dates; anyone can reproduce the query in a browser. Community discussion on Stack Exchange's own meta sites has tracked the same decline from the inside, noting a 2014 peak above 6,700 questions per day and daily averages through the first half of 2026 of 96, 82, 62, 50, 42, and 49 -- a trajectory the July figure continues without deviation.

Some context on what a "question" is here. Stack Overflow's value was never the volume of text. It was that a working programmer with a specific broken thing wrote it down, strangers argued about it, and a correct answer got voted to the top and edited over years as the underlying technology changed. That is an unusually clean training signal: a real problem, a verified solution, and a public record of the wrong turns in between. It is why the site's archive shows up in essentially every code-model training corpus, including large public code datasets.

Attributing the decline entirely to AI would be too quick, and the record does not support it. The site has been changing its own product and moderation posture throughout the period: an AI-assisted answer feature launched in February 2025 with community verification built in, new spam tooling in January 2026 that the company said cut the time spam stays live by half, an experiment allowing open-ended and opinion-based questions expanded to all users in February 2026, a redesigned site in beta in March, and in June 2026 a product aimed at coding agents, pitched on the argument that isolated agents keep rediscovering the same fixes because they have no shared verified source of truth. Any of these changes what gets asked and what survives. Stack Overflow's own engineering blog has been arguing throughout that a trust and validation gap remains, and that developers still need human expertise.

But the direction is not seriously in dispute, and the mechanism is not mysterious. Most of what Stack Overflow received historically was routine: an error message, a misused API, a configuration that would not load. Those questions now go to a model, which answers instantly, without judgement, without a duplicate-close, and without publishing anything. The answer is correct roughly as often as the site's would have been, and it leaves no public trace.

Here is the loop that makes this more than a nostalgia story. The models are good at these questions because millions of humans answered them in public, at length, and corrected each other for fifteen years. The corpus of new human answers to new problems is now roughly one percent of what it was. When a framework ships a breaking change in 2027, the pattern that used to play out -- a hundred developers hit it, argue in public, and converge on the right fix within days -- has no venue. The model will still answer. It will answer from the last era it has data for, and the thing it lacks is precisely the thing it cannot know it lacks: this is hallucination with a schedule attached.

The industry's working answer is synthetic data -- models generating training material for models -- plus retrieval against documentation. Both help. Neither reproduces what the site actually provided, which was an adversarial process: a claim, a challenge, a vote, and an edit war that eventually settled on something true. Documentation states intent. Stack Overflow recorded what happened when intent met reality.

One honest caveat about the numbers. The API counts questions that currently exist, so questions deleted over the years are absent. Deletion has been far more aggressive on the 2014 cohort than on this year's, which means the historical figure is understated relative to the recent one -- the real ratio between peak and now is larger than 118-fold, not smaller. The comparison is conservative in the direction that matters.

There is a second-order effect worth watching too. If new public programming knowledge is thinning while agents increasingly write code by pattern-matching existing code, the field's shared record of how to fix things is being replaced by a private one, held per-session in tools that publish nothing. That may be fine. It is a change nobody voted on, and the meter reading is 1,490.

Primary source, verified: read the paper →

Ground Truth. (2026, August 8). Stack Overflow took 1,490 questions in July. Ground Truth. https://groundtruth.day/news/stack-overflow-took-1490-questions-in-july.html

@misc{groundtruth:stack-overflow-took-1490-questions-in-july,
  title  = {Stack Overflow took 1,490 questions in July},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/stack-overflow-took-1490-questions-in-july.html}
}

Topics: stack-overflow · training-data · developers · industry · coding