What AI did to stackoverflow in a graph

Original Article ↗ Hacker News Discussion ↗

Stack Overflow’s question volume has collapsed from a 2014 peak to a tiny fraction of its former activity, prompting debate over whether generative AI or longstanding structural problems are more to blame. Many argue that hostile moderation, aggressive duplicate-closing, gamification, and outdated answers had already driven users away years before ChatGPT, while others see AI as the decisive accelerant that made waiting for human replies obsolete. The thread also raises worries about where future AI systems will get fresh training data as human-driven Q&A sites, internal help channels, and technical blogs all see declining engagement.

Shape of the decline (per the graph)

  • Activity peaks around 2014, then trends downward from ~2016 onward.
  • COVID produces a temporary spike in 2020.
  • Around late 2022–2023 (ChatGPT era), questions drop sharply, roughly halving within a year and falling ~99% from peak by 2026.
  • Some argue the curve looks like a “bell” and was already well-explained by long-term dynamics; others say multiple different forces shape different parts of it.

AI vs. pre-existing problems

  • One camp: AI is the main cause; LLMs provide instant, conversational answers, so users stop asking on SO and on internal “SO-like” channels (Slack, etc.).
  • Another camp: SO was already in serious decline due to culture and design; AI simply delivered the “death blow” or sped up an inevitable outcome.
  • Some stress that hiring downturns, higher interest rates, and other macro factors also reduced programming Q&A demand.

Moderation, culture, and gamification

  • Many describe SO as hostile: aggressive closing (especially as “duplicate”), snarky or contemptuous comments, and a Meta culture seen as self-righteous.
  • Reputation thresholds, “this should be a comment” deletions, and nitpicky edits made newcomers feel unwelcome; even high-rep users report frustration.
  • Others defend strict moderation as essential for quality and argue critics misunderstood the site’s goal (a curated reference, not a chat forum).
  • Gamification incentivized busywork, gatekeeping, and early adopters’ dominance instead of long-term knowledge stewardship.

Design flaws and aging content

  • The “canonical Q&A” model clashes with a fast-changing field: old accepted answers remain, while new questions get closed as duplicates.
  • Closing by duplicate to unanswered or only-vaguely-related threads was common, wasting askers’ effort.
  • Lack of conversation and community focus meant little reason to stay once a better answer channel (LLMs, Discord, GitHub) appeared.

Shifts to alternatives and data reuse

  • Many now prefer:
    • LLMs (for immediacy, politeness, and “rubber-duck” debugging).
    • GitHub issues/discussions, project Discords, and Reddit.
  • SO’s content is widely believed to be a major training source for coding LLMs; some worry about future training data as Q&A sites and blogs also decline.

Comparisons and governance issues

  • Users link the downturn to private-equity ownership and controversial incidents (e.g., moderator conflicts) as signs of mismanagement.
  • Parallels drawn to Reddit and Wikipedia: similar tensions around heavy-handed moderation, community burnout, and AI cannibalizing traffic.