How Gboard can learn without collecting your texts
A conventional keyboard-learning system might send examples of what people type to a central server, where they can be combined and used to improve predictions. Federated learning changes where the examples are processed: the training happens on users’ devices, while the shared model is updated from the results.
A model is a set of numerical parameters that help software predict what word or correction is likely. In a federated-learning cycle, a device receives a model, uses eligible local data to make a small training adjustment, and sends an update—not the underlying messages—to a server. The server combines updates from many devices to improve a shared model, which can then be sent back out.
What an update reveals—and what it does not
The update is mathematical information about how the model should change. It is not a copy of a text conversation. That separation matters for a keyboard: private messages can contain names, plans, and intimate details, yet a model can still learn broad patterns that help it predict words or recognize common typing errors.
Federated learning is not the same as proving that no information about a user can ever leak. Updates are derived from local data, so their handling and protection still matter. The key difference is that the training examples themselves do not have to be uploaded for the shared model to improve.
Why the approach suits a keyboard
A phone sees the kind of writing a general-purpose model may miss: abbreviations, unusual names, and personal shorthand. Learning on-device lets that data contribute without first pooling everyone’s raw text in one place. The resulting improvement is collective rather than a server reading and memorizing each person’s messages.
That tradeoff is the point: the keyboard can benefit from patterns across many users while keeping the original conversations on their devices.
