Google DeepMind and GovTech Singapore announced a collaboration on October 14, 2026, to fine-tune a lightweight Gemma model for detecting scam conversations on-device. The design screens incoming messages first, then checks broader conversation context if the initial classifier flags a possible threat. Google DeepMind says an integrating app may access conversations only with the user’s consent and under that app’s terms of use.
How the two-stage detector is designed to work
A lightweight classifier screens incoming messages for signs of a possible scam. If it flags a message, the fine-tuned Gemma model examines a larger snapshot of the conversation. Google DeepMind says that second stage is intended to confirm a threat or dismiss a false alarm.
A detection can also prompt a real-time warning advising the user to pause before sharing sensitive information or money. The model’s role is to help identify manipulative conversations—not to take action on the user’s behalf.
What the announcement says about conversation access
The model is designed to analyze conversations on the device rather than send them to a server for analysis. Separately, Google DeepMind says an app integrating the system may access conversations only with user consent and in line with the app’s terms. The company also says neither Google nor the Singapore Government can access conversations through deployment.
The project’s training process is a separate part of the design: Google DeepMind says fine-tuning took place in a GovTech-managed Google Cloud environment. It says training excluded private or personally identifiable information and used anonymized data derived from scam reports and intelligence feeds. The partners also describe testing with synthetic dialogues and local corpora.
The integration path
Google DeepMind describes phased integration into applications developed with the Singapore Government. It also said the collaboration was presented at Singapore International Cyber Week 2026.
