Singapore’s Government Technology Agency (GovTech) and Google DeepMind have outlined an AI system designed to spot scam conversations on smartphones while keeping message analysis on the device. The announcement describes a working model, but does not say when ordinary users will be able to install or use it.
A small classifier screens messages before Gemma reviews flagged conversations
DeepMind says it adapted a lightweight Gemma model by training it further on anonymized conversation data derived from local scam reports and intelligence feeds. The company says training excluded private messages and personally identifying information. The examples included impersonation of public agencies, banks and local figures, as well as regionally trending schemes. The goal is to tell manipulative exchanges from ordinary conversation, a harder task than flagging a suspicious phrase in isolation.
The system uses two stages. A small classifier first screens incoming messages with minimal power, according to DeepMind. If it marks a conversation as suspicious, a fine-tuned Gemma model reviews a wider portion of the conversation. The second step can either raise a scam warning or decide the conversation is not a scam. Keeping the more demanding review for flagged cases is intended to make contextual analysis practical on a typical phone.
DeepMind says the model analyzes conversations locally and can warn people to pause before sharing money or credentials. That design matters for messaging services with end-to-end encryption, where the service provider cannot read private chats to scan them centrally. The announcement describes the intended on-device design; it does not report an independent privacy audit.
Public release and performance results are unreported
The public status is not yet clear. DeepMind says the collaboration has delivered a model, while Singapore’s CNA reported the same day that GovTech and DeepMind were developing real-time on-device detection. Neither account gave a public release date or a list of supported messaging apps. DeepMind says it presented the work at Singapore International Cyber Week on October 14, 2026.
The announcement says the model was tested on synthetic conversations and local scam datasets, but it publishes no accuracy, false-alarm or battery-use figures. Those results would help show how often the system catches real scams, how often it interrupts legitimate chats, and how it performs on a phone outside a test set.
