Clio: Privacy-Preserving Insights into Real-World AI Use

AI assistants are being used for diverse, practical tasks in the real world: Analysis of one million Claude.ai conversations revealed that the most common use cases are coding, writing, and research tasks, with users seeking everything from hairstyle advice to technical guidance on Git operations, demonstrating that AI assistants have become embedded in everyday workflows across multiple domains.

Privacy-preserving analysis is technically feasible: Clio successfully analyzed millions of conversations without human reviewers reading raw data by using AI assistants themselves to identify patterns, achieving "a high degree of accuracy and privacy" according to the paper's extensive evaluations—suggesting that companies can study real-world AI use while protecting user privacy.

Usage patterns vary significantly across languages and regions: The research found meaningful differences in how AI is used globally, such as conversations in Japanese discussing elder care and aging populations at higher-than-typical rates, indicating that AI assistants serve different cultural and demographic needs that may require localized approaches.

Real-world usage data helps identify safety risks: Beyond understanding how people use AI, Clio enabled the researchers to detect coordinated abuse attempts, monitor for unexpected harms during product launches and major world events, and improve existing safety monitoring systems—demonstrating that analyzing actual usage patterns is critical for responsible AI deployment.