What One Million Prompts Tells Us About AI Usage, Topics, and Preferences

With extensive research and significant progress made in AI in recent years, generative conversational AI products (such as ChatGPT, Claude, etc.) have been increasingly integrated into daily life, across a variety of tasks. While much research has focused on improving the capability of conversational AI, such as through developing novel model architectures and richer training corpora, comparatively less attention has been given to understanding how people use conversational AI. In this study, We analyzed a large, naturalistic dataset of 1 million human-AI prompt-response logs to explore and quantify common types of use cases. Focusing primarily on human prompts in a conversational dialogue system, we apply topic modeling techniques, and found that usage preferences were primarily for assisting users with coding and programming (15.54%), business management (10.33%), and visual design (8.76%). We also examined how usage dynamics of multi-turn conversation with a dialogue system-like conversational AI (GPT-series models). The results indicate that the majority of users' multi-turn conversations stay on one topic, and for the few multi-turn conversations identified, most of conversations focused on one topic instead of transitioning between multiple topics. Our research sheds light into usage dynamics for conversational AI, and also highlights the kinds of conversational topics that users use these AI for.