Google DeepMind and Singapore’s Government Technology Agency have built an artificial intelligence model that can detect scam conversations directly on smartphones while keeping the analysis on the device. The collaboration was presented October 14 at Singapore International Cyber Week 2026.
Singapore’s scam losses
Google DeepMind said scams account for more than half of Singapore’s reported crimes and have caused more than S$4 billion in losses since 2019. Social-engineering scams, including fake investment offers, job schemes and impersonation attempts, can involve long conversations designed to build trust before a victim is asked for money or credentials.
The partnership fine-tunes Google’s lightweight Gemma model for local scam patterns. Training examples covered impersonations of public institutions, banks and local figures, along with regional schemes that generic models can struggle to identify.
Gemma runs on phones
Unlike systems that scan messages on remote servers, the new model processes conversation data on the device. Google said it will be integrated into applications through a phased rollout and will only access conversations with the user’s consent. Neither Google nor the Singapore government will be able to access users’ conversations through the deployment, according to the announcement.
Google has already described on-device scam detection for calls and messages on Android. Its February 2026 announcement said the processing remains on the phone and that the feature is being expanded to additional devices and regions.
Two-stage warning system
The proposed system uses two layers to balance detection quality and battery use. A lightweight classifier first screens incoming messages for possible threats. Suspicious conversations then receive more detailed analysis from the fine-tuned Gemma model, which examines a larger portion of the exchange to confirm a scam or dismiss a false alarm.
When manipulative patterns are detected, the system can issue a warning on the device advising the user to pause before sharing sensitive information or sending funds. Google said the model was tested with synthetic dialogues and local data sets covering multiple scam categories and changing impersonation tactics.
Privacy rules for rollout
GovTech said fine-tuning took place in its managed Google Cloud environment, keeping training data isolated. Private and personally identifiable citizen data was excluded from training; the project used anonymized data derived from scam reports and intelligence feeds.
The model is not yet described as a general release for all smartphone users. Its deployment will be phased through applications developed with the Singapore government. Singapore’s AntiScam Pango privacy and security guidance covers permissions, on-device processing and data handling.
Google DeepMind said the project demonstrates how governments and model providers can build real-time safety tools without sending private conversations to the cloud, and that it is exploring similar deployments in other countries.
References
Geewax, J. J., & de Andrés, M. (2026, October 14). Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection. Google DeepMind. https://deepmind.google/blog/on-device-ai-scam-detection-mddi/
Google. (2026, February 25). Staying one step ahead: Strengthening Android’s lead in scam protection. Google Security Blog. https://blog.google/security/strengthening-android-lead-in-scam-protection/
Chua, N. (2026, August 26). Over $410m lost to scams in first half of 2026; 8 in 10 victims voluntarily handed money to scammers. The Straits Times. https://www.straitstimes.com/singapore/courts-crime/over-410m-lost-to-scams-in-first-half-of-2026-8-in-10-victims-voluntarily-handed-money-to-scammers
Open Government Products. (n.d.). AntiScam Pango: Privacy and security. AskGov Singapore. https://ask.gov.sg/pango?topic=Privacy+and+Security