Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams
Google DeepMind and Singapore rolled out an on-device scam detector for phones. It spots shady chats without sending private messages to the cloud.
Based on reporting by Google — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind and Singapore’s Government Technology Agency have teamed up on a scam-detection model that runs directly on a phone. The pitch is simple: catch fraud early, without forcing people to give up private conversations to a server somewhere else.
That matters in Singapore, where scams account for more than half of reported crimes and have caused more than S$4 billion in losses since 2019. The source of the damage has also changed. This is no longer just crude phishing. Attackers now lean on social engineering, dragging victims into convincing back-and-forth conversations before asking for money, and those scams are said to make up as much as 80% of all losses.
The model is built from Google DeepMind’s lightweight Gemma system, then fine-tuned on local scam conversations. Those examples include impersonations of public bodies, banks, local figures, and newer regional schemes, so the system can tell ordinary chat apart from manipulation that is trying to steer someone toward sharing money or credentials.
The interesting part is where the analysis happens. Instead of scanning messages in the cloud, the model works on the device itself, which matters for end-to-end encrypted messaging apps that cannot inspect private chats on their own. When it spots manipulative patterns, it gives a warning on the user’s phone and tells them to slow down before handing over sensitive information or funds.
There is also a battery problem to solve, so the system uses two stages. A light first pass screens incoming messages with minimal power use, and then a larger slice of the conversation is checked by the tuned Gemma model to confirm the threat or toss out a false alarm. Google DeepMind says the training stayed inside GovTech’s managed GCP environment, used no citizen data, and relied on anonymised scam reports, intelligence feeds, synthetic dialogues, and local corpora. The collaboration was presented at Singapore International Cyber Week 2026.
My take — AI-written commentary, not fact-checked reporting
This is the kind of AI project that actually deserves a press release: narrow, useful, and not allergic to privacy. The industry loves talking about safety while quietly funneling everything to the cloud; on-device filtering is the more honest move. More governments should push for that, instead of asking citizens to trade away private chats for the privilege of not getting scammed.
Read more about this at: Google