Google DeepMind and GovTech Singapore build on-device AI to catch scams

Google DeepMind and GovTech Singapore, the agency that runs the city-state’s government technology, have built a conversational scam detection system using fine-tuned versions of Gemma, Google’s family of open AI models. Google DeepMind announced the collaboration as part of what it calls a national partnership to strengthen Singapore’s AI capabilities.

The system is designed to spot scam conversations, exchanges where a fraudster tries to manipulate someone into sending money or sharing personal information, as they happen. What sets it apart is where the analysis takes place. Rather than sending a conversation to a remote server for the AI to read, the fine-tuned Gemma model runs directly on the user’s device. Google DeepMind describes this as zero cloud data egress, meaning the conversation text being screened never leaves the phone to reach outside servers.

That architecture matters for a scam-detection tool specifically because the content it needs to analyze, private messages, is also the content most people want kept off company servers. Processing on-device lets the system flag a suspicious message without transmitting the conversation itself anywhere.

GovTech Singapore builds and runs digital services for the Singapore government, which makes it the natural partner for deploying a consumer-facing protection tool at national scale. Google DeepMind is Google’s AI research arm and the developer of Gemma, which it has positioned as an openly available alternative to its closed Gemini models for developers who want to run AI locally rather than through an API.

What hasn’t been detailed

The available material does not specify which size of Gemma model was fine-tuned, which app or service will carry the feature, or when it will reach the public. There is also no published data yet on how accurately the system distinguishes real scams from ordinary conversations, or how many people it is expected to protect. Google DeepMind and GovTech Singapore have not said whether the tool will be open-sourced, limited to Singapore, or offered as a template for other governments.

What to watch next is whether either organization follows up with a concrete release, such as a named app update or a technical report detailing the model’s size and accuracy. Confirmation of a public rollout date, and any independent testing of how well the on-device model catches real scam attempts, would mark the next real checkpoint for this partnership.