Google DeepMind and Singapore's GovTech fine-tune a Gemma model to warn people about scams on their phones, with conversations staying on the device

Google DeepMind said on 14 October that it has worked with Singapore's Government Technology Agency (GovTech) to fine-tune its lightweight Gemma model to detect scams on the user's own device. The work is part of DeepMind's national AI partnership with Singapore and was presented at Singapore International Cyber Week.

The system has two stages. A light classifier screens incoming messages. The fine-tuned Gemma model then reviews a larger part of the conversation, confirms a threat or dismisses a false alarm, and warns the user in real time before they share information or money. The model was trained in GovTech's managed Google Cloud environment, which the post calls a sovereign sandbox, on anonymised scam reports and local scam patterns such as impersonation of banks and public institutions. It reads conversations only with the user's consent.

The post says that as the model runs entirely on the device, “neither the Singapore Government nor Google, will be able to access users' conversations through its deployment.” It gives no rollout date, no named app, no user numbers and no detection or false-alarm rates, and says the rollout will be “carefully phased”.

A government-backed detector reading conversations on citizens' phones raises privacy questions, which keeping the model on the device and asking for consent are meant to answer. Without accuracy figures, it is not yet possible to judge how well it works.

• Confirmed A Gemma model fine-tuned with GovTech Singapore runs on the device in a two-stage pipeline and warns users in real time. Google DeepMind
• Claimed Neither the Singapore Government nor Google will be able to access users' conversations. Google DeepMind
• Confirmed No rollout date, accuracy figure or user count is given. Google DeepMind