Google DeepMind and Singapore build on-device AI scam detector

Google DeepMind and Singapore’s Government Technology Agency say they developed an on-device AI system designed to flag manipulative scam conversations without sending those conversations to either partner. The design pairs a lightweight initial classifier with a locally fine-tuned Gemma model, but the partners have not named an integrating app or a public availability date.

Google DeepMind said any rollout inside applications will be phased with the Singapore Government and will require users’ consent under the integrating app’s terms. In other words, the model-development work is complete, but the detector is not yet publicly available.

According to Google DeepMind, the first stage screens incoming messages while using minimal power. Messages it marks as suspicious move to the second stage, where the fine-tuned Gemma model examines a larger snapshot of the conversation. The model then either confirms a suspected threat or dismisses the initial alert as a false alarm. The project extends Google’s on-device model work beyond uses such as Google’s EmbeddingGemma 2 release for on-device multimodal search.

When the system identifies a manipulative pattern, it produces an on-device warning advising the user to pause before sharing sensitive information or sending money. Google DeepMind said the model runs entirely on the device and that neither it nor the Singapore Government will be able to access users’ conversations through the deployment.

Google DeepMind said GovTech Singapore fine-tuned the lightweight Gemma model on local scam conversations, including impersonations of public institutions, banks and local figures, as well as regional schemes. The work took place in GovTech’s managed Google Cloud environment. The partners said training excluded private or personally identifiable citizen data and used anonymized datasets derived from scam reports and intelligence feeds.

The system was tested with synthetic dialogues and local corpora, according to Google DeepMind, but the announcement did not provide precision, recall, false-positive rates or an independent evaluation. It also did not identify the Gemma variant, the first-stage classifier architecture, supported handsets or languages, or expected effects on memory, storage, latency and battery life.

The project targets a substantial volume of fraud. Singapore Police recorded 37,308 scam cases and about S$913.1 million in losses in 2025. Its annual brief said victims were manipulated into transferring funds themselves in 81.8% of reported scams, while online platforms were used to contact victims in 84.1% of cases.