Toward provably private insights into AI use

Google Research · 9 months ago · 48

Google announced a system called provably private insights (PPI) that uses large language models, differential privacy, and trusted execution environments to analyze how people use generative AI tools while protecting individual user data. The system uses an open-source Gemma 3 model running inside a secure processor to classify unstructured data like transcripts, then aggregates results with differential privacy guarantees before sharing insights with developers. Google deployed this system in the Recorder app for Pixel phones and open-sourced the underlying confidential federated analytics stack to allow external verification of privacy claims.