Community When AI Answers the Questions

Stack Overflow is back to 2009 question volume, but Q&A was always a function, not the community itself. What remains when AI answers the questions (identity, trust networks, showcase, and taste) is exactly what DevRel should build for.

In December 2025, Stack Overflow received 3,862 new questions. Not per day. For the month. That's down 78% from the year before, and roughly 95% below the 2014 peak, when the site fielded around 200,000 questions a month. Question volume is back to where it was in May 2009, ten months after launch.

The easy read is that developer community is dying. I think the easy read confuses a function with the thing itself, and DevRel teams that accept it will dismantle the part of their program that matters most right when it matters most.

Two honest caveats before the argument. Stack Overflow's decline started around 2018, well before ChatGPT, and had as much to do with community culture as with technology. AI accelerated the collapse; it didn't originate it. And the business didn't collapse with the traffic. Prosus reported the segment containing Stack Overflow growing revenue 12% in its most recent half-year. The community function and the company decoupled, a reminder that community health and platform revenue were never the same metric, which is exactly the confusion this post is about.

Q&A was a function, not the community

Here's the reframe. Q&A was deficit-driven interaction: you showed up because you were stuck. It was the highest-volume thing communities did, so it became the thing we measured, and then quietly the thing we meant when we said "community."

AI now serves that function inside the editor, instantly, privately, without the social cost of asking in public and getting your question closed as a duplicate. That interaction volume is evaporating no matter what any DevRel team does. Watching your forum thread count drop and concluding your community is failing is like watching phone calls decline after texting arrived and concluding people stopped talking.

The useful question is what all that Q&A volume was proxying for. Because people didn't stay in communities for answers alone. They stayed for things that don't show up in a thread count.

Identity. Being a "Rails person" or a "Postgres person" is part of a developer's professional self-image, and communities are where that identity gets performed and confirmed. An AI can answer your question. It can't confer belonging.

Trust networks. Communities are where you learn whose judgment to weight. This one is getting more valuable, not less. When an agent generates ten plausible solutions, "which one is right?" becomes "who do I trust to tell me which one is right?" Sunil Pai's point about judgment, that AI is a "cheat code machine" producing output faster than developers build the judgment to vet it, applies directly here: judgment develops through exposure, comparison, critique, and repeated contact with consequences. Every one of those is a social process. Communities are where it happens.

Showcase. People want to show what they built to people equipped to appreciate it. Jono Bacon points to the Tailwind Discord as the model of self-sustaining help culture. Watch it closely and it's as much an identity and showcase economy as a support channel.

None of those functions are automatable. Austin Kodra put it plainly in his community-building series: "You can't automate away trust-building, no matter how accurate a given chatbot or recommendation system might be."

What the metrics will do to you

This shift breaks community measurement in a specific, predictable way. Bacon's best community health metric is self-sufficiency: members answering members without maintainer intervention. Keep that metric, but expect its denominator to fall. AI is absorbing the easy questions, so total thread volume drops. What's left should be higher signal: the edge cases and judgment calls AI handles badly, the "it works but is this the right shape?" questions.

Falling volume with rising depth is a healthy community in 2026. Falling volume with falling depth is an actual decline. If your dashboard only counts threads, both look identical, and the healthy case will get its budget cut. I covered the broader measurement problem in the metrics post in this series; the community-specific version is to measure identity and trust behaviors rather than interaction volume: showcase posts, member-to-member help on hard questions, contributor retention, discussion depth, and whether the same names keep coming back when they don't need anything.

What DevRel builds now

If Q&A migrates to AI and docs, community energy goes to the layer AI can't serve. Concretely:

Showcase infrastructure. Channels, programs, and rituals for members to show work to peers. This is the identity economy, and it's the stickiest thing you can build.

Taste debates. Two working solutions, which has the better shape and why? Sessions and threads structured around judgment rather than correctness. This is where the community becomes the place developers get better, not unstuck. Live witnessed practice sessions, the format from earlier in this series, fit here exactly.

Contribution ladders. Bacon's time-to-second-PR logic, where the interval before a first-time contributor comes back for round two is the real signal of a working ladder, applies to community roles too: clear paths from member to helper to trusted voice. Trust networks need rungs.

The hard-question layer. Route the questions AI answers badly, your product's genuinely sharp edges, into the community deliberately, and celebrate the people who handle them. That's where member expertise becomes visible and trust gets built.

The Kodra series has a warning worth carrying into this: communities run on value creation, not value capture, and they take longer to show results than any other GTM motion. The AI era doesn't change that. It removes the vanity volume that let teams pretend otherwise, which might be the most useful thing about it.

Your forum was never the community. The people who kept showing up after they stopped being stuck were. AI made that distinction impossible to ignore.

Part 8 of 9 in the Developer Relations in the Age of AI series.

← Previous: Evaluate Your DevRel Program Like an Agentic Workflow · Next: Value Creation Just Became the Distribution Strategy → · ← All articles