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

Google DeepMind and Singapore’s Government Technology Agency announced a fine-tuned, on-device scam detection model built under Singapore’s National AI Partnership. Scams account for more than half of Singapore’s reported crimes and more than S$4 billion in losses since 2019, with social-engineering conversations now driving up to 80% of losses, according to the DeepMind post.

The model is a lightweight Gemma variant fine-tuned on local scam conversations—impersonations of banks, public institutions, and regional schemes—so it can flag manipulative chat patterns without sending message contents to Google or the government. A two-stage pipeline screens messages cheaply, then runs deeper in-context analysis only when needed.

Rollout into citizen-facing apps will be phased, with user consent required for on-device access. Training used anonymized scam-report datasets in a GovTech-managed GCP sandbox, not private citizen chats.

This is the kind of AI deployment we actually want to see more of. Not a chatbot that lectures you about phishing after the wire went out—an on-device classifier trained on local scam vernacular that can interrupt the conversation before money moves.

Singapore’s numbers make the urgency obvious. When social engineering drives most of the loss, keyword filters and SMS blocklists are fighting yesterday’s attack. Fine-tuning for local impersonation patterns is the difference between a generic model and one that recognizes how a fake bank officer talks in Singapore.

The privacy design is the other half of the story. Cloud scanning of encrypted chats is a political and technical dead end for many messaging apps. Running on-device, with consent, and with neither Google nor GovTech reading the conversation through the deployment, is a credible answer to the “safety vs encryption” trap. We still want to see the consent UX, false-positive rates, and what happens when the model is wrong about a family group chat.

Governments elsewhere will copy the press release. Fewer will copy the governance: sovereign sandbox, no citizen PII in training, phased app integration. We’re watching whether this stays a Singapore showcase or becomes a template messaging platforms can adopt without gutting end-to-end encryption.

Scammers increasingly use long social-engineering chats rather than one-shot phishing links. Messaging platforms and governments are searching for detection that works without centralizing private conversations.