Total monthly number of StackOverflow questions over time

Overall shape of the decline

• Query (questions only) shows:
  • Peak around 2014 (~207k questions/month).
  • Plateau / gentle decline from ~2014; clearer downtrend from ~2017.
  • COVID spike around April 2020, then resumed decline.
  • By 2025: entire year had fewer questions than a single month in 2021.
  • Late 2025: ~3.7k questions/month; early Jan 2026 extrapolates to similar levels.
• Some note the current rate is lower than in the site’s earliest days, though not literally zero.

LLMs vs pre‑AI decline

• Many see ChatGPT (late 2022) as the inflection point: a visible acceleration downward from an already-declining baseline.
• Others stress the decline began years before usable coding LLMs:
  • Question growth stalled ~2014; sustained decline from ~2017.
  • GitHub Copilot, GitHub Discussions, issues, Discord, and Reddit had already been siphoning questions.
• Broad consensus: LLMs “body‑slammed” SO, but onto a slope created by earlier issues.

Culture, moderation, and user experience

• Large number of anecdotes about:
  • Hostile tone, snark, “RTFM” attitudes, and pile‑on downvotes.
  • Legitimate questions closed as “duplicate,” “off‑topic” or “too broad,” with linked threads that didn’t actually solve the problem.
  • Difficulty asking as the corpus grew and standards tightened; some users banned or rate‑limited for deleting comments or asking about policy.
• Counter‑view from long‑time curators:
  • SO was never meant as a personal help forum but as a curated, duplicate‑reduced, matter‑of‑fact knowledge base.
  • Downvotes and closures are framed as quality control, not personal attacks.
  • Many complaints are seen as misunderstanding the site’s goals and mechanics (e.g., duplicates as “signposts” to canonical answers).
• General agreement that incentives (reputation, review queues) and unpaid moderation produced brittle, bureaucratic behavior.

Question saturation and ecosystem shifts

• One camp argues a natural saturation effect:
  • “All the basic questions” for mainstream languages were answered; new questions either duplicates or very niche.
  • Google got better at surfacing existing SO threads (and later rich snippets), so fewer people needed to ask.
• Others counter that new technologies and versions constantly create fresh question space; a healthy community should have maintained more growth.
• Many now prefer:
  • GitHub issues / Discussions and project‑specific forums.
  • Discord/Slack for “help desk” style support (though this hides knowledge from search and archives).
  • Reddit for more conversational, opinion, or “soft” questions.

Impact on LLMs and the knowledge commons

• Widespread worry that as public Q&A dries up:
  • Future models lose fresh, real‑world troubleshooting data.
  • Remaining public web is increasingly “AI slop” recycling old answers.
• Some argue LLMs can fall back to:
  • Official docs, code repositories, and live web tools/RAG.
  • Telemetry and agentic coding logs in closed ecosystems.
• Others note SO’s special value:
  • undocumented workarounds, bug lore, and “ship‑relevant” edge cases that don’t appear in manuals or code.
  • rich comment discussions and corrections that LLMs currently do not replicate.

What people valued – and what’s lost

• Many recall:
  • Canonical, high‑quality answers with deep explanations and tradeoffs.
  • Occasional gems: novel algorithms, elegant tricks, and insights from core library/language authors.
  • The ability to “pay back” by answering hard questions and building public artifacts.
• Others emphasize persistent flaws even for experts:
  • Advanced, niche questions often went unanswered or got low engagement.
  • Accepted answers frequently became stale, while better later answers stayed buried.
• Several express grief that:
  • Real human‑to‑human problem solving is being replaced by private, ephemeral LLM chats.
  • New engineers will never experience SO’s “golden age.”

Proposed futures / alternatives

• Ideas floated:
  • An AI‑first Q&A platform where:
    • LLMs draft initial answers; humans verify, correct, and add production context.
    • Reputation accrues for validating or fixing answers, not just posting first.
  • A Q&A‑plus‑wiki model, with strong incentives to update answers as tech evolves.
  • New, open, CC‑licensed communities (or federated systems) combining human curation and LLM assistance.
• Skeptics note:
  • Fixing AI‑generated content is tedious and unrewarding at scale.
  • Without strong product and moderation design, any replacement risks repeating SO’s trajectory.