Google DeepMind and Singapore GovTech build on-device AI to flag scam chats

Google DeepMind and Singapore’s GovTech have fine-tuned an on-device scam-detection model built on Gemma, DeepMind wrote on 14 October 2026. The goal is to spot manipulative chat patterns — investment fraud, fake jobs, official impersonation — without sending message text to the cloud.

Why Singapore framed the problem this way

DeepMind cites Singapore figures: scams make up over half of reported crimes and more than S$4 billion in losses since 2019, with complex social engineering now behind up to 80% of losses. Generic models often miss local tricks, so the team fine-tuned on local scam conversation patterns (impersonations of banks, institutions and regional schemes), using anonymised datasets rather than private citizen messages, according to the post.

How the on-device pipeline works

DeepMind describes a two-stage design for phones:

1. A lightweight classifier screens inbound messages with low power use.
2. If something looks suspicious, a fine-tuned Gemma model reviews a larger conversation snapshot to confirm or dismiss the alert.

When the model flags manipulation, the device can show a real-time warning before someone shares money or credentials. DeepMind says the model accesses on-device conversations only with user consent via the host app’s terms, and that neither Google nor the Singapore government can read chats through this deployment because inference stays on the device.

Governance and rollout

Training ran in a GovTech-managed GCP sandbox. No private PII was used for training, DeepMind says; testing included synthetic dialogues and local corpora. App rollout will be phased with the Singapore government. No public accuracy numbers or consumer launch date appear in the blog post.