Google DeepMind and Singapore's GovTech develop on-device scam-detecting AI

The system uses a two‑stage approach: a power‑efficient lightweight classifier screens incoming speech, and a second stage runs a fine‑tuned Gemma model trained on local scam tactics to analyse context and issue real‑time warnings. It operates only within user‑consented apps, keeping conversations encrypted and never transmitting them to Google or the government. The technology is already in practical use, with social implementation having begun.

Authorities say the rollout will be phased, and plans are already announced to expand the solution to other countries. The privacy‑first design aims to protect users while preserving the confidentiality of their personal communications.