Google DeepMind and Singapore’s Government Technology Agency (GovTech Singapore) have developed a fine-tuned AI model intended to detect scams in conversations while keeping message data on the user’s device. The project is part of the National AI Partnership with the Singapore Government and was presented at Singapore International Cyber Week 2026.
The model uses a lightweight version of Google’s Gemma, adapted with local scam conversations. These included impersonations of public institutions, banks and local figures, as well as regional schemes. The aim is to distinguish ordinary chats from conversations that show signs of manipulation.
Unlike detection systems that scan messages on remote servers, the model is designed to analyse conversations on-device. If it identifies manipulative patterns, it can issue a real-time warning advising the user to pause before sharing money or sensitive information. Any deployment in applications will be phased and developed with the Singapore Government. The model would access conversations only with user consent under the relevant app’s terms, and its on-device operation means Google and the government cannot access those chats through the deployment.
To limit battery use, the system follows a two-stage process. A lightweight classifier first screens incoming messages for potential threats; the Gemma model then examines a broader portion of the conversation to confirm a threat or dismiss a false alarm.
Development took place in a GovTech-managed Google Cloud environment, where training data was kept isolated. The teams excluded private and personally identifiable information, using anonymised data derived from scam reports and intelligence feeds. They also tested the model against synthetic conversations and local datasets covering different scam types and impersonation tactics.
The announcement cites Singapore figures showing that scams—including investment fraud, fake job offers and impersonation of public officials—account for more than half of reported crimes and have caused over S$4 billion in losses since 2019. Social-engineering scams involving persuasive conversations are said to account for up to 80% of losses.
Sources: Google DeepMind Blog.