Google DeepMind and Singapore’s Government Technology Agency (GovTech) have announced an on-device AI scam detection model for smartphones in Singapore, unveiled at Singapore International Cyber Week 2026 on 14 October 2026.DeepMind Blog
The collaboration matters for regulators and platforms because it offers a concrete template for fraud detection that runs locally on user devices rather than in the cloud, aiming to preserve end-to-end encryption while still flagging suspected scam conversations.DeepMind Blog
According to Google DeepMind, scams account for over half of reported crimes in Singapore and more than S$4 billion in losses since 2019, with complex social engineering scams now responsible for up to 80% of those losses.DeepMind Blog The company says the work with GovTech took place under a broader National AI Partnership with the Singapore Government.DeepMind Blog
Google DeepMind states that the new tool is built by fine-tuning its lightweight Gemma model on local scam conversations, including impersonations of public institutions, banks, local figures and regional schemes, to distinguish between ordinary social chat and manipulative attempts to obtain money or sensitive data.DeepMind Blog A separate post on Google DeepMind’s social media account similarly describes the work with @GovTechSG on “on-device fraud detection using our open model Gemma,” fine-tuned on local scam tactics to identify threats in real time.Signals — AI company announcements
The model runs entirely on the user’s device and is designed to operate on what Google DeepMind calls “typical smartphones in Singapore,” avoiding server-side scanning that is incompatible with end-to-end encrypted messaging platforms.DeepMind Blog When it detects what the company describes as manipulative conversational patterns, the system triggers a real-time warning on the device advising users to pause before sharing sensitive information or funds.DeepMind Blog
To manage battery and performance constraints, Google DeepMind says it uses a two-stage pipeline: a lightweight classifier first screens inbound messages for potential threats, and a second stage runs the fine-tuned Gemma model on a larger slice of the conversation to confirm or dismiss a suspected scam.DeepMind Blog
The company describes a governance framework in which fine-tuning took place in a GovTech-managed Google Cloud Platform environment described as a "Sovereign Sandbox," with training data kept isolated.DeepMind Blog Google DeepMind says the system did not use private or personally identifiable citizen data, instead relying on anonymised datasets derived from scam reports and intelligence feeds, and reports that the model was stress-tested on synthetic dialogues and local text corpora to cover different scam types and changing impersonation tactics.DeepMind Blog
According to Google DeepMind, the model is intended for integration by messaging platforms and operating system vendors to provide local scam protection without exposing user conversations, but the announcement does not specify a deployment timetable, participating apps, or any enforcement or liability framework for providers that choose not to adopt it.DeepMind Blog