For years, platforms like Stack Overflow were where people went when they got stuck. Now large language models are taking over more and more of that problem solving. Julian Just's study looks at 212,891 problem communities from 2020 to 2025 across four major model releases: ChatGPT, GPT-4, GPT-4o and o3. Using a difference-in-differences design, it asks which problems AI takes over and which ones still need the crowd.

Sharing some selected key insights:
🔹 Knowledge availability drives substitution. The more problem-specific knowledge already exists, the more sharply crowd activity drops after each model release.
🔹 Complexity buys time, not immunity. Complex problems that are interdependent and hard to structure resist substitution at first. With each new model generation, that protection weakens.
🔹 A moving knowledge frontier. Crowds stay relevant mainly where knowledge is scarce, and that boundary keeps shifting. This raises a key question for IS research: what happens to collective knowledge creation when fewer problems are solved in the open?

Thanks again, Julian Just, for the inspiring talk and the lively discussion!