What do people use AI models for?
- Source
- youtube.com
- Author
- Anthropic
- Date
Context
Researchers on Anthropic's Societal Impacts team describe Clio, short for Claude insights and observations, in a roundtable. Earlier work started from harms chosen in advance, such as discrimination or persuasion, or paid testers to probe the . Clio adds a bottom-up view of what people actually ask Claude, so evaluations can be aimed at real uses.
Research engineer Miles explains the pipeline. A model writes a private high-level summary of each conversation, related summaries are grouped, and another model describes each group, repeated into a hierarchy of uses. A further model checks the groups for private or identifying details, down to anything that could identify roughly a thousand people, and minimum counts of distinct organizations and conversations are enforced. The team says no human needs to read the conversations in this , speaking of general traffic. That privacy design is their own description.
Checking sign-in…
Loading comments…






