The Commons Was Already Dying

Stack Overflow is one of the most consequential institutions programming ever produced: a public archive where, for a decade and a half, practitioners wrote each other's documentation. At its peak it absorbed a new question every fourteen seconds. The standard story is that ChatGPT killed it. The data says something more uncomfortable: by the time the AI arrived, the commons had already lost a third of its volume, most of its answering capacity, and nearly all of its warmth toward strangers.

1  The arc

Question volume grew explosively to 2.2 million in 2016 — and then declined, year over year, to 1.63 million by 2021. No AI existed to blame. The decline coincided with the maturing of the archive itself: by the mid-2010s, most beginner questions had already been asked and answered, and new questions were increasingly either duplicates (closed) or hard specialty questions (unanswered).

2  The commons hardened

Volume is the least of it. In 2009, 99.7% of questions received at least one answer, typically within 12 minutes. By 2021 the answer rate had fallen to 76% and the median wait had quintupled to about an hour. The accepted-answer rate — the asker confirming their problem was actually solved — fell from 58% to 40%. The machine that turned questions into documentation was seizing up, one percentage point a year, for a decade.

3  What it did to newcomers

Track each user cohort by the quarter they created their account, and look at what happened to the first question they ever asked. For the 2009 cohorts, first questions were answered 99.7% of the time. For 2020 cohorts: 78%. The share of first questions downvoted into negative territory — the community's way of saying "you asked wrong" — quadrupled from 4% to a peak of 16% around 2018. A newcomer's first contact with the commons went from a welcome to a toll gate.

4  Empires rise and fall inside the archive

The tag time-series double as a stratigraphy of software itself. jQuery — the most-asked-about technology on earth in 2013, 2.4% of all questions — decays to a rounding error by 2022. React crosses it in 2018 and keeps climbing. Python grows from 1.3% to 5.3%, quietly becoming the language of the archive's second half. Each panel below is a technology's life story in one shape.

5  The weekend index

One more register, a happier one: divide each tag's weekend share of questions by the site-wide weekend share, and you get a clean instrument for what programmers do for love. Weekend tags: Discord bots, pygame, Haskell, Rust, x86 assembly — and linked lists and time-complexity, because interview prep is a weekend sport. Weekday tags: SSIS, SharePoint, Azure DevOps, reporting services — nobody builds an enterprise report pipeline for fun on a Sunday.

Epilogue: then the machine came

This archive stops 66 days before ChatGPT launched. What happened next is outside our data and deserves its own careful treatment with a different source — Stack Overflow's own traffic disclosures and the Internet Archive's later dumps. But the numbers above reframe that story before it starts. Generative AI did not kill a healthy commons; it arrived at one that had spent six years losing volume and a decade losing the ability to answer its own questions, and it offered the people the commons had been turning away — the newcomers of Figure 3 — an oracle that never downvotes.