Google DeepMind and GovTech Singapore have fine-tuned a version of Google’s lightweight Gemma model to spot scam conversations on a user’s own device rather than scanning messages on a server. The work, shown during Singapore International Cyber Week, falls under a National AI Partnership with the Singapore government.
The design is a two-stage pipeline built to spare the battery. A small classifier screens incoming messages, and only when it flags something does the tuned Gemma model read a larger slice of the conversation to confirm or dismiss the threat. If it sees a manipulative pattern it warns the user in real time, before money or credentials move. Fine-tuning ran inside GovTech’s managed Google Cloud environment on anonymised scam reports and intelligence feeds, with personal data excluded, and the model was stress-tested against synthetic dialogues and local corpora.
Why it matters: the local tuning is the point. Generic models miss scams because the lures are local, impersonating named banks, agencies and public figures. Google says scams make up over half of Singapore’s reported crimes and have cost more than S$4bn since 2019. There is no launch date, the rollout inside applications will be phased, and Google says it is looking at other countries.