Google DeepMind said it worked with the Government Technology Agency of Singapore (GovTech Singapore), under a National AI Partnership, to fine-tune a privacy-preserving, on-device scam-detection model. A lightweight Gemma model was fine-tuned on local scam conversations covering impersonations of public institutions, banks and local figures. The system uses a two-stage pipeline: a lightweight classifier screens incoming messages, then the fine-tuned Gemma model analyses a larger snapshot of the conversation to confirm a threat or discard a false alarm, triggering a real-time warning on the device. DeepMind said the model runs entirely on-device, so neither the Singapore government nor Google can access users' conversations, and that fine-tuning took place in GovTech's managed cloud environment using anonymised datasets with no citizen data. The work was presented at Singapore International Cyber Week.
Our takeThis is one of the more useful AI-safety stories of the week because it solves a real constraint instead of advertising a capability. Scam detection has always meant either cloud scanning (a privacy problem) or brittle keyword filters; moving a fine-tuned Gemma onto the phone, behind a cheap first-pass classifier, threads both needles. DeepMind's framing — a template "for other countries" — is also a quiet export pitch. The honest caveats: rollout depends on app makers and government phasing, and efficacy rests on benchmarks we have not seen.