How People Use ChatGPT [pdf] (cdn.openai.com)

OpenAI researchers and academic co‑authors released a large-scale analysis of ChatGPT consumer use from launch through July 2025, finding rapid global adoption (≈700 million users, ~18 billion messages/week by mid‑2025) and a shift toward non‑work activity: non‑work messages rose from 53% to over 70% of consumer traffic. Using an automated, privacy‑preserving pipeline that classifies randomly sampled user messages with LLM‑based prompts (no human reading content) and links aggregated employment/education data in a secure clean room, they show that nearly 80% of conversations fall into three topics—Practical Guidance, Seeking Information, and Writing. Writing is the dominant work use (≈40% of work messages), and two‑thirds of writing requests are edits or rewrites rather than generation from scratch. They also report output types: Asking (49%), Doing (40%), Expressing (11%); as of July 2025, 56% of work messages are Doing, and ~75% of those Doing messages are writing tasks. Technically and economically important takeaways include that ChatGPT’s value is concentrated in decision support and producing digital outputs (especially writing), not primarily in coding (only ~4.2% of messages) or companionship (~1.9%). Demographics show the early male usage gap largely closed, heavy use by under‑26s, and faster growth in low‑/middle‑income countries; educated and high‑paid professionals are more likely to use it for paid work. The study’s automated taxonomies and secure linking approach provide a scalable, privacy‑minded model for measuring real‑world LLM behavior and its implications for labor, education, and consumer welfare.