Advances in private training for production on-device language models
Not a pilot. Google deployed differential privacy across 30+ production language models in Gboard, protecting user data for 15+ countries—the largest known DP deployment anywhere.
Why it matters
Google achieved a major privacy milestone in on-device AI by deploying formal differential privacy guarantees at scale across Gboard's language models, including the first production models with strong DP (ε ≤ 1) trained directly on user data. This signals a critical shift in how AI companies must balance utility with privacy in consumer-facing applications.
The key facts
8 to know30+ Gboard on-device language models deployed with differential privacy guarantees
7+ languages and 15+ countries covered
First production announcement of strong DP guarantee (ε ≤ 1) for models trained directly on user data (Portuguese Brazil and Spanish Latin America)
ε values range from 0.994 to 13.69 with δ fixed at 10^-10
12,000+ devices participating in training rounds for ε ≤ 1 models
Largest known deployment of user-level differential privacy in production
All future Gboard LM launches trained on user data now require formal DP guarantees
Spanish model in Spain improved from ε=8.9 to ε=5.37 through optimization
Go to the source
Google Research Blogblog.research.google
Publisher excerpt: Posted by Zheng Xu, Research Scientist, and Yanxiang Zhang, Software Engineer, Google Language models (LMs) trained to predict the next word given input text are the key technology for many applications [1, 2]. In Gboard, LMs are used to improve users’ typing experience by supporting features like…