Despite the rapid adoption of LLM chatbots, little is known about how they are used. We approach this question theoretically and empirically, modeling a user who chooses whether to complete a task herself, ask the chatbot for information that reduces decision noise, or delegate execution to the chatbot.

The model—a rational inattention problem with a noisy delegation option—predicts that querying is favored for high-stakes, high-context decisions, while delegation is favored for routine, low-context tasks where the chatbot has a productivity advantage. Empirically, we study the growth of ChatGPT’s consumer product from its launch in November 2022 through July 2025, combining usage logs with a privacy-preserving pipeline that classifies a representative sample of conversations. We document rapid global diffusion—reaching around 10% of the world’s adult population—with work use concentrated among highly educated, highly paid professionals and an increasing share of non-work use, which has risen from 53% to more than 70% of all messages.

Conversations are dominated by “Practical Guidance,” “Seeking Information,” and “Writing,” and users with more complex, knowledge-intensive jobs—managers, highly educated, and higher-paid professionals—are more likely to use ChatGPT as a decision aide than as an agent, consistent with the model’s predictions.