DeepMind and GovTech Singapore presented an on-device scam-detection model on October 14. A version of Gemma is adapted to local scam conversations to identify manipulation beyond phishing links alone. The announcement describes development and phased integration into applications, not protection already enabled on every Android phone.
The system has two stages. A lightweight classifier screens incoming messages; suspicious cases move to Gemma for multi-turn conversation analysis. This separation aims to limit energy use while retaining more context for ambiguous cases. A warning should encourage users to pause before transferring money or sensitive information.
The partners say analysis remains local, with access to conversations subject to user consent within the relevant application. Fine-tuning took place in GovTech’s managed GCP environment; training data is described as anonymised and free of personally identifiable information. These are partner statements, not the outcome of an independent audit performed here.
Belgian availability is not announced, and the post supplies no detailed performance or false-positive measurements. For developers, the cascade architecture is instructive, but effective protection also requires consent management, local evaluation and recourse when warnings are wrong. Running a model on a phone does not by itself guarantee the security of the entire application.