Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection

Google DeepMind says it has worked with the Government Technology Agency of Singapore to fine-tune a version of its lightweight Gemma model that detects scams directly on users' phones. The project, part of a National AI Partnership with Singapore and presented at Singapore International Cyber Week 2026, targets a problem DeepMind describes as severe: scams make up more than half of the country's reported crimes and have cost over S$4 billion since 2019. Social engineering, in which attackers build trust through conversation before asking for money, now represents up to 80% of losses.

The model was trained on local scam conversations, including impersonations of banks, public institutions and local figures. It uses two stages to conserve battery: a small classifier screens incoming messages, then the Gemma model examines a larger portion of the conversation to confirm a threat or dismiss a false alarm. If manipulation is detected, the user gets a real-time warning.

DeepMind says the model only reads conversations with user consent and that neither Google nor the government can access them through its deployment. Training took place in a GovTech-managed cloud environment using anonymised data from scam reports and intelligence feeds, excluding personal information. Mark Chen, GovTech's director of trust and safety, said the approach catches targeted local scams while keeping communications private.

GovTech has launched an open beta of its Anti-Scam Pango app for Android users in Singapore, which applies the joint research, and the partners will keep testing and integrating the models during the beta. Wider rollout will be phased, and DeepMind says it may explore similar deployments elsewhere.