Google DeepMind and Singapore Develop On-Device AI Scam Detection Model
Google DeepMind and Singapore’s Government Technology Agency (GovTech) have developed an artificial intelligence model designed to detect scam conversations directly on users’ devices, aiming to identify sophisticated social-engineering attempts without sending private messages to external servers.
Announced on October 14, 2026, the collaboration uses a fine-tuned version of Google’s lightweight Gemma model to recognise local scam tactics, including impersonation of public institutions, banks and other figures. The initiative forms part of the National AI Partnership between Google and the Singapore government.
The model is designed to analyse conversation patterns and warn users when messages appear manipulative. Its deployment within applications will be phased, with access to conversations requiring user consent under the relevant application’s terms of use.
On-Device AI Targets Sophisticated Scam Tactics
Traditional scam detection systems may rely on server-side analysis of messages. The new approach processes conversation data directly on the device, allowing the model to examine potentially suspicious exchanges while keeping the underlying conversations local.
The model has been fine-tuned using anonymised datasets derived from scam reports and intelligence feeds. Its training process excluded private citizen data and personally identifiable information, according to the project announcement.
The focus is on social engineering, in which scammers build convincing conversations before asking victims to transfer money or disclose sensitive information. The model is intended to distinguish ordinary conversations from exchanges that show signs of manipulation.
Two-Stage Screening System Aims to Reduce False Alerts
The scam detection system uses a two-stage process designed to balance detection accuracy with smartphone efficiency.
A lightweight classifier first screens incoming messages for potential threats. Messages flagged as suspicious proceed to a fine-tuned Gemma model, which analyses a broader portion of the conversation to determine whether it contains manipulative patterns.
If the system identifies a likely scam, it can trigger a warning on the device, advising users to pause before sharing sensitive information or transferring funds. The initial screening stage is also intended to limit unnecessary processing and battery consumption.
Privacy and Deployment Remain Central
The model was developed in GovTech Singapore’s managed Google Cloud environment, where the training data was kept isolated. Testing included synthetic dialogues and local datasets designed to cover different scam categories and impersonation tactics.
Google DeepMind and GovTech presented the collaboration during Singapore International Cyber Week 2026. The project addresses a key challenge for digital security: identifying suspicious conversations without requiring broad access to private communications on centralised servers.
The partners plan to introduce the technology gradually through applications in collaboration with the Singapore government. They have also indicated that similar approaches could be explored in other countries. However, the announcement does not specify a public release date or confirm widespread availability.
The Google DeepMind and GovTech collaboration demonstrates how on-device AI could strengthen scam detection while limiting exposure of personal messages. Its effectiveness in real-world use will depend on deployment, detection performance and its ability to adapt as scam techniques evolve.
