Google DeepMind and the Government Technology Agency of Singapore (GovTech) announced on October 14, 2026, at Singapore International Cyber Week, an AI model that detects scam conversations directly on smartphones without sending messages to the cloud. The project targets social engineering scams, which account for most scam losses, while keeping users’ personal communications private.
Key Facts At A Glance
- Google DeepMind and GovTech announced the on-device scam detection model on October 14, 2026.
- The model is built on Google’s lightweight Gemma model and fine-tuned on local scam conversations.
- Detection runs in two stages: a fast screening classifier, followed by deeper analysis from the fine-tuned Gemma model.
- Conversations are processed on the device, and neither Google nor the Singapore Government can access them.
- Training used anonymised scam reports and intelligence feeds and excluded personally identifiable information.
- Scams account for more than half of Singapore’s reported crimes, with losses exceeding $4 billion since 2019.
- No rollout date or participating applications have been announced.
A Locally Trained Model Built For Privacy
Google DeepMind and GovTech have developed an AI model that identifies manipulative patterns in smartphone conversations and warns users before they share sensitive information or transfer money. The model was presented at Singapore International Cyber Week 2026 and forms part of Google’s National AI Partnership with the Singapore Government.
The system is built on Gemma, Google’s lightweight open model. It was fine-tuned on local scam conversations involving impersonation of public institutions, banks and local figures, as well as emerging regional schemes. The model is intended for Android devices and will be updated as scam methods change.
The project responds to the scale of fraud in Singapore. Scams make up more than half of the country’s reported crimes, and losses have exceeded S$4 billion since 2019. Google DeepMind said social engineering, where scammers use convincing conversations to manipulate victims, now accounts for up to 80 percent of scam losses.
How The Two-Stage Detection Works
To limit battery use, the system screens messages in two stages. A lightweight classifier first checks incoming messages for potential threats. When it flags suspicious content, the fine-tuned Gemma model reviews a larger portion of the conversation to confirm the threat or rule out a false alarm. If manipulative patterns are confirmed, the user receives a real-time warning advising caution.
Because analysis happens on the phone itself, messages are not sent to external servers. Mark Chen, Director of Trust and Safety at GovTech, said fighting scams “shouldn’t mean sacrificing personal privacy.” He added that the approach is designed to catch targeted local scams while keeping communications private.
Training Safeguards And Rollout Plans
The teams fine-tuned the model in a Google Cloud environment managed by GovTech, keeping training data isolated. They used anonymised datasets drawn from scam reports and intelligence feeds and excluded private and personally identifiable information. The model was tested against synthetic conversations and local datasets covering a range of scam types and impersonation tactics.
The two organisations plan to introduce the model gradually through applications developed with the Singapore Government. Once deployed, it will access conversations only with user consent and under each application’s terms of use. Google also released Be Scam Ready 2.0, a shorter version of its scam awareness game, as part of the same announcement.
Publicly available information remains limited. No rollout date, participating applications or performance figures have been disclosed. Google DeepMind said it is exploring whether similar technology could be deployed in other countries.