Google and Singapore unveil an on-device Gemma scam detector

Scam warnings without cloud access to chats

Google DeepMind and Singapore’s Government Technology Agency say they have developed a scam-detection model designed to run on ordinary smartphones. Their October 14 announcement describes a two-stage system: a lightweight classifier first screens incoming messages, then a fine-tuned Gemma model checks a wider slice of the conversation for manipulation. If the system flags a threat, it is designed to warn the user on the device before sensitive information or money is shared. The approach aims to work with private, end-to-end encrypted conversations because the message analysis stays on the phone.

What the announcement does and does not establish

The partners say they fine-tuned the model on local scam patterns, including impersonation of banks and public institutions. Google says the training excluded private citizen data and used anonymized material from scam reports and intelligence feeds. It also says the model was stress-tested against synthetic dialogues and local corpora. Those are the partners’ descriptions of their development and testing, not a published independent measure of how many real scams it catches or how often it alarms on harmless conversations. The announcement says messaging platforms and operating-system vendors can integrate the model; it does not name a public consumer rollout or claim that every phone is already protected.