Michael JohnsonOn October 14, Google DeepMind and Singapore’s government technology agency GovTech introduced a model they developed together to detect scam conversations directly on smartphones. The system analyzes message context locally and is designed to be integrated into apps; deployment will be gradual and require user consent.
The solution is based on a lightweight Gemma model tuned using local examples of scams. According to Google DeepMind, training used anonymized datasets derived from scam reports and intelligence. The company says the model was tuned in a GovTech-managed cloud environment, and citizens’ personal data was not used for training.
First a filter, then conversation analysis
The system works in two stages. A lightweight classifier quickly checks incoming messages and passes along those that appear suspicious. The tuned Gemma model then analyzes a broader segment of the conversation to identify a manipulative scheme or rule out a false alarm. This approach reserves complex analysis for selected cases, while taking the importance of conserving phone battery into account.
If the model detects manipulative patterns, it is intended to display an on-device warning and prompt the user to pause before sending money or disclosing sensitive information. Analyzing the full conversation is necessary because scammers may first build trust and only later ask for a transfer or a code.
What on-device processing means
The model is designed to process conversations on the device, rather than send their contents to a server for screening. Google DeepMind says that after integration into an app, neither the company nor the Singapore government will be able to access messages through the model. Access to conversations will be granted with the user’s consent and under the terms of the specific app.
On-device analysis could be useful for end-to-end encrypted apps: processing takes place on the user’s device, while conversation contents are not sent to a server for screening. Google DeepMind describes integration as a future possibility for apps; it has not named specific messaging services or partners.
What is known about readiness
In its October 14 post, Google DeepMind reports testing the system on synthetic conversations and local datasets covering different categories of scams. The company describes future integration as a gradual effort with the Singapore government. The announcement does not name specific apps, a date when the feature will become available to users, or figures for accuracy and false-alert rates.
For now, the confirmed details concern the architecture and planned direction of deployment, not a publicly available feature that users can install. Once integrated, the quality of warnings on real conversations, language support, and the effect of analysis on battery life will help show its practical value. The proposed approach demonstrates how an on-device model can check conversation context while keeping the messages themselves on the device.