Introduction of Team and Their Interests:
Deep Ganguli, a research scientist on the Societal Impacts team at Anthropic:
Driven by understanding how people are using and being affected by AI systems.
Focuses on using this understanding to make AI systems safer.
Works on anticipating the societal impacts of these systems.
Emphasizes the general-purpose nature of the systems and their myriad potential uses and effects.
Esin Durmus, a research scientist on the Societal Impacts team:
Interested in understanding the societal impact of AI systems.
Researches the values AI systems should have, how to incorporate those values, and how to evaluate whether systems embody those values.
Mentions prior involvement in "clear work."
Miles, a research engineer on the Societal Impacts team:
Focused on understanding how AI systems are used in the real world and their impact on people.
Particularly interested in building systems that enable empirical understanding of how AI systems are used.
Involved in the development of Clio.
Alex, a researcher on the Societal Impacts team:
Interested in understanding how general-purpose AI systems are used today as a way to predict their future use.
Aims to build societal resilience in the face of new technology.
Considers the perspective of those outside the AI lab and what information they would need.
Involved in the Clio project.
Clio (Claude Insights and Observations) is a tool that provides a high-level overview of how people are using the AI system Claude.
Clio reveals:
High-level aggregate clusters of usage patterns.
Potential risks and benefits associated with Claude's use.
Insights into future directions of AI technology.
Methods used before Clio:
Top-down approaches: Identifying a potential harm (e.g., discrimination) and then measuring its presence in the system. Example given: Discrimination in high-stakes decision making scenarios.
Red teaming: Employing contract workers to adversarially probe the systems for potential harms and assess their success.
The gap Clio fills:
Previous methods lacked insight into real-world usage patterns.
Clio provides a way to observe actual user interactions and understand where evaluation efforts are most relevant (e.g., identifying contexts where discrimination or persuasion are more likely to occur).
Clio bridges the gap between hypothetical scenarios in laboratory settings and empirical evidence gathered from real-world interactions.
Data Collection: Clio analyzes a large dataset of real-world conversations people have with Claude.
Summarization: A language model processes each conversation and extracts a private, high-level summary focusing on the user's overall request or intent.
Clustering: Related conversation summaries are grouped together, forming clusters that represent common user intents or topics.
Cluster Analysis: Another language model analyzes these clusters and generates descriptions that explain what is happening in each group of conversations. This process is repeated to create a hierarchy of use cases.
Privacy Filtering: A separate model scans all clusters to ensure they do not contain any private or identifying information. Any information that could be linked back to approximately 1,000 individuals is removed.
Aggregation Minimums: Quantitative thresholds are applied to ensure that each cluster contains a minimum number of unique organizations and conversations, further protecting user privacy.
Internal Insights: The results are then made available internally to inform the design of better evaluations, guide product development, and deepen the understanding of how AI systems are being used across various contexts.
Key Points:The entire process is designed to maintain a high standard of user privacy.
No human reads the raw conversations.
Claude, the AI system, analyzes conversations people are having with it.Discussion of Clio:Prior Methods and the Gap Clio Fills:How Clio Works:
