Singapore has a new idea for stopping scammers: put the AI detective inside your phone and never let it phone home.
On October 14, 2026, Minister for Digital Development and Information Josephine Teo announced a strategic collaboration between the Government Technology Agency of Singapore (GovTech) and Google DeepMind. The goal is to build lightweight, on-device AI models, based on Google’s Gemma models, that can catch conversational scams in private messaging apps as they happen.
The pitch rests on one design choice. The analysis runs entirely on the user’s device, with zero cloud data egress. Your chats stay on your phone, and the scam-spotting happens there too.
How the scam interceptor is supposed to work
Conversational scams don’t always arrive with a dodgy link. They unfold slowly, through friendly messages, manufactured urgency, and emotional pressure designed to walk a victim toward a bad decision.
The system is meant to read the flow of a private conversation in real time and spot the manipulative tactics scammers rely on. The aim is early intervention, ideally before money moves.
Teo made the announcement at Singapore International Cyber Week (SICW). She framed it as part of a broader push to counter AI-powered scams, noting that the fight has grown tougher as scammers adopt more sophisticated AI tools.
The minister did not give a launch date for the Gemma-based tool. She also did not share performance figures, such as detection rates or false-positive rates.
Part of a bigger anti-scam machine
The new tool joins an existing stack of Singapore government defenses. The most prominent is the Scam Analytics and Tactical Intervention System, or SATIS.
SATIS works on behalf of the Singapore Police Force and disrupts approximately 20,000 malicious websites per month.
GovTech also runs an AI system called rMSE, which scans over 600,000 websites daily.
ScamShield rounds out the picture as another piece of the country’s wider anti-scam effort.
Why privacy is the whole ballgame
The no-egress design is what makes this project politically viable. A tool that sends chats to a government-linked server would face serious trust problems, even if the intent were purely protective.
Keeping everything on-device sidesteps much of that debate. The model can warn a user without anyone else seeing the conversation.
What this means for tech and public safety
There are open questions worth watching. The first is timing, since no rollout date has been set. The second is accuracy: a scam detector that cries wolf too often will get ignored, and one that misses too much offers little protection.
Distribution is another unknown. A detection model only helps if it’s built into the apps people actually use, and that typically requires cooperation from messaging platforms and device makers.
Scammers adapt quickly, and Teo’s own remarks acknowledged that AI is making their tactics harder to counter. Any model trained to recognize today’s scripts will need regular updates to keep up with tomorrow’s.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.