Anthropic has developed a tool called Clio, designed to identify new threats and unknown harms in using its AI chatbot, Claude. Recently, Clio detected a coordinated spam network attempting to manipulate search engine optimization by generating text through Claude. This internal tool analyzes a million conversations to cluster similar topics and highlight suspicious activities. Anthropic found that the top uses for Claude include coding and software development, educational purposes, and business strategy, with the latter three categories accounting for only about twenty-three percent of interactions. The company aims to share Clio’s methodology to encourage other AI labs to adopt similar strategies for monitoring and improving the safety of AI technologies. In the future, Anthropic envisions using Clio to understand the evolving job landscape and enhance safety evaluations based on real-world usage. The cost of operating Clio is roughly forty-eight dollars per one hundred thousand conversations.
Why do we care?
It’s encouraging to see a model provider invest in an observability and risk detection tool, which must be integrated into AI deployments. Service providers offering AI solutions to customers must build similar governance frameworks or partner with vendors who can provide monitoring capabilities.
Service providers could leverage Clio’s model to develop new AI safety and monitoring services, offering clients real-time observability into their AI tools to identify risks, optimize performance, and ensure compliance. This would be a notable competitive difference in a landscape where model providers are rapidly moving to commoditization.