DeepMind and GovTech Singapore deploy on-device Gemma scam detector for private chats

Google DeepMind and Singapore's GovTech, working with the Ministry of Digital Development and Information, have deployed a scam-detection system built on small Gemma models that run entirely on the user's device. The tool watches conversational patterns typical of scams, such as impersonation, urgency or payment pressure, and flags them inside private messaging. Messages are never sent to a server.

The design matters because scam detection has a privacy problem. The most useful signal lives in end-to-end encrypted chats that cloud classifiers cannot and should not read. Running a compact Gemma model locally sidesteps that conflict and keeps data residency trivially satisfied. That is a selling point for a government partner.

The launch fits a broader edge-AI push this week. Liquid AI released a 2B-active on-device agent model, EmbeddingGemma 2 was repurposed for speech-to-text on HN, and llama.cpp shipped a Windows app. Apple's Siri AI also leans on on-device processing plus Private Cloud Compute. Sovereign and government deployments are becoming a meaningful channel for small open models.

Key open questions are false-positive rates and how the system handles fast-evolving scam scripts. Small models are quick but brittle, and scammers adapt. It is also unclear whether Google will offer the detector beyond Singapore, for example inside Android Messages globally.