Google DeepMind and GovTech Singapore are proposing a phased rollout, not releasing a consumer app today.
Google DeepMind and Singapore’s Government Technology Agency, or GovTech Singapore, are developing an on-device scam detector for integration into applications. The project matters to AI builders because it applies a language model under real phone limits: limited battery, private data and fast decisions.
The system targets social engineering, where scammers build trust through conversation before requesting money or sensitive information. DeepMind says its lightweight Gemma model was fine-tuned on local scam conversations involving public institutions, banks, local figures and regional schemes.
Its proposed design uses two stages. A smaller classifier first screens incoming messages with low power use. If it spots a possible threat, Gemma examines a larger slice of the conversation to confirm the danger or discard a false alarm.
A confirmed threat could trigger a warning on the phone, telling the user to pause before sharing funds or credentials. The analysis would run on the device rather than through cloud-based message scanning, according to DeepMind. An application could access conversations only with the user’s consent and under its terms of use; DeepMind says neither Google nor Singapore’s government could access those conversations through deployment.
DeepMind says training used anonymized data from scam reports and intelligence feeds, excluding private or personally identifiable information. Testing included synthetic dialogues and local corpora.
The announcement describes development and a planned phased rollout. It names no participating applications or release date, and gives no accuracy, false-alarm, response-time or battery results. For engineers and product leaders, the next useful milestone is evidence from integration testing: whether the two-stage design catches new local scams without draining phones or interrupting ordinary conversations.