Google DeepMind and Singapore's Ministry of Digital Development and Information (MDDI) have jointly built a lightweight AI system designed to catch phone and chat scams in real time — without sending any user data to the cloud.
The system uses a fine-tuned version of Google's open Gemma model, optimized to run entirely on-device. According to the announcement, the model can intercept conversational scams in private chats with zero cloud data egress, meaning the user's conversations never leave their phone.
The collaboration comes as AI-powered voice cloning and social engineering scams have surged across Southeast Asia. Singapore's GovTech, the government's technology agency, helped design the deployment pipeline, ensuring the model works on mid-range devices commonly used in the region.
Unlike traditional cloud-based scam detection, which requires uploading audio or text for analysis, the on-device approach processes everything locally. DeepMind researchers say the model was trained on a curated dataset of known scam patterns and can generalize to new conversational tactics in real time.
The project is part of a broader push by DeepMind to demonstrate practical applications of small, efficient AI models. The Gemma family has become one of the most widely deployed open model families, and this scam-detection use case shows how fine-tuning can turn a general model into a specialized security tool.
MDDI officials said the system is being evaluated for integration with Singapore's national anti-scam infrastructure, with potential deployment in early 2027.