Google reports moving client-gradient computation—the calculation of model updates—from devices to server-side trusted execution environments (TEEs) in its federated-learning system. In an October 2, 2026 announcement, Google Research says Gboard has adopted the system for English and Japanese next-word prediction. The announcement does not establish the adoption date.

Google says the change brings faster training and improved accuracy, alongside stronger privacy protections. It describes server workloads that external auditors can inspect through published access policies and verifiable code execution. For teams assessing privacy-sensitive machine learning, the relevant change is where computation happens and how server processing can be audited—not a demonstrated business return.

Those privacy claims have conditions. Google acknowledges current-generation TEE limitations. When proprietary information is loaded into a running training program, Google says its stated privacy assurances depend on all privacy-relevant logic remaining hardcoded in that program. The company describes ongoing research into side-channel observations and expects future hardware to provide deeper protections. Full formal correctness proofs remain a possibility for the future, not an achievement established by this announcement.

This review inspected Google's announcement, not the linked whitepaper or code, and did not independently verify Gboard adoption. It provides an attributed account of the architecture change; it does not establish an adoption recommendation.