news

Google says Gboard training now uses protected servers

Google says Gboard uses a new way to train next-word suggestions on protected servers. Its privacy claims have conditions and remain unverified by this review.

MasterNodeAI EditorialBy MasterNodeAI Editorial TeamOctober 4, 20262 min read
news

Google says Gboard training now uses protected servers

Google says Gboard, its keyboard app, is using a new system to train English and Japanese next-word suggestions. Its October 2 announcement describes a change in where the work of learning from device data happens.

In Google's earlier system, devices calculated updates to the prediction model. Google says those calculations now run on servers. It calls this federated learning: devices contribute to learning coordinated by a service provider.

How is information protected when that work moves? Google says devices encrypt training examples before uploading them and specify which computations may use them. Processing takes place in protected server environments, called trusted execution environments, or TEEs. Google says these protect the running program's internal information and let third parties check which code runs, subject to limitations in today's TEE technology. Google says the permitted computations release only anonymized results.

The protections are conditional. When extra proprietary information is loaded while a training program runs, Google says all rules affecting privacy must remain fixed in the program's code. Google describes deeper side-channel protections for these dynamically loaded workloads and full mathematical proofs of software correctness as future possibilities.

For a business reader assessing AI privacy claims, the useful distinction is between where data is processed and what restricts its use there. Google's account explains a way to make server processing checkable; it does not establish a business return or a reason to adopt the system.

This review read Google's announcement, not its linked whitepaper or code, and did not independently verify Gboard's adoption or the privacy protections. The announcement does not establish when Gboard adopted the system.

Updated (2026-10-05): We have rewritten the explanation of Google's reported Gboard training change and clarified its privacy conditions and the limits of this review. The original publication date is unchanged.