Google DeepMind and Singapore's GovTech unveil on-device AI tool to flag chat scams

Google DeepMind and Singapore's Government Technology Agency (GovTech) said on 14 October 2026 that they have built an on-device AI model to warn smartphone users in real time when a chat conversation looks like a scam, without sending private messages to the cloud. The partners said scams make up over half of Singapore's reported crimes and have accounted for more than S$4 billion in losses since 2019, with social-engineering schemes — where attackers build a convincing conversation before asking for money — now behind up to 80% of all losses, according to the DeepMind announcement.

What the tool does

• The model is a lightweight version of DeepMind's Gemma, fine-tuned on local scam conversations covering impersonations of public institutions, banks and local figures, so it can tell "benign social chatter from genuine manipulation", DeepMind said.
• It runs in two stages on the phone itself: a low-power classifier screens incoming messages, then a contextual pass by the fine-tuned Gemma model confirms a threat or discards a false alarm, according to the announcement.
• Because analysis happens on the device, the system can flag manipulative patterns on end-to-end encrypted messaging platforms that cloud-based scanning cannot inspect, DeepMind said.
• The work was carried out inside a "Sovereign Sandbox" — a GovTech-managed cloud environment — using anonymised scam-report data rather than citizens' own messages, and was presented at Singapore International Cyber Week 2026.
• "Fighting scams shouldn't mean sacrificing personal privacy," said Mark Chen, GovTech's Director of Trust and Safety, in the DeepMind announcement.

Why it matters

The project is one of the first public examples of a frontier AI lab fine-tuning a model for a government partner specifically to run offline on ordinary phones, rather than through a cloud API — a design choice aimed at reconciling fraud detection with encrypted messaging rather than pushing platforms to break it. For regulators and banks outside Singapore, including in the Caucasus and Central Asia where mobile banking is expanding fast and social-engineering fraud is a growing concern, it offers a concrete template for on-device, privacy-preserving scam alerts as an alternative to demands for message-scanning access.