Google DeepMind and GovTech scam detector

Google DeepMind and Singapore's Government Technology Agency announced an on-device AI scam detector at Singapore International Cyber Week 2026, using a fine-tuned Gemma model to analyze conversations in real time without sending data to a server. The tool uses a two-stage pipeline: a lightweight classifier screens messages first, and the heavier Gemma model is invoked only when confirmation is needed, balancing detection accuracy with battery efficiency. The model was trained on Singapore-specific scam scenarios, including impersonations of public institutions and banks, using a sovereign sandbox environment that excluded personally identifiable information. Scams account for over half of Singapore's reported crimes and more than S$4 billion in losses since 2019, with social engineering schemes responsible for up to 80% of losses. Running analysis on-device also sidesteps the limitation of server-side scanning, which cannot inspect end-to-end encrypted messages. Separately, Google launched an opt-in Enhanced Protection mode in Android Messages that uses cloud-based Gemini models to analyze conversations from unknown senders for scam intent. A distinct Gemini Nano feature checks unfamiliar links using real-time web signals, but requires a Pixel 10 or newer, a Galaxy S26, or select other devices. The Android Messages feature targets reply-based scams, which a University of Florida study found adapt to each victim and grew faster than link-based scams, making keyword filters ineffective. The FTC reported $470 million in text scam losses in 2024, and total reported fraud losses reached $15.9 billion in 2025, with texting as the top contact method.