
Your phone may soon read your messages to keep you safe — and, in a twist, the whole point is that nobody else gets to read them. On October 14, 2026, Google DeepMind and Singapore's Government Technology Agency (GovTech) announced a first-of-its-kind on-device AI model that scans your chats for scam tactics and warns you before you send money or share sensitive details. The analysis happens entirely on your phone, so neither Google nor the Singapore government can see your conversations. As reported by The Straits Times and Channel NewsAsia, the project was unveiled on the sidelines of the Singapore International Cyber Week.
Built on Gemma, Trained on Local Scam Tricks
The system is built on Google's lightweight Gemma model, fine-tuned with local scam conversation data so it understands how fraudsters actually talk in Singapore. According to Google DeepMind's announcement, the training covered tactics ranging from impersonation of public agencies, banks, and local celebrities to scam methods currently trending in the region. That local flavor matters: generic fraud filters often miss culturally specific manipulation, such as fake calls "from the police" or messages impersonating government officials — the exact schemes the Straits Times says are fine-tuned into this model.
Privacy was engineered into the training itself. The fine-tuning ran inside a Google Cloud environment managed by GovTech Singapore, and the training data excluded personally identifiable information entirely — only anonymized data extracted from scam reports and intelligence feeds was used. The finished model was also stress-tested against synthetic conversations designed to reflect local speech patterns.

How It Works: A Two-Stage Check on Your Phone
To keep battery drain low, the system uses a two-stage pipeline:
- Stage 1 — Lightweight classifier: every incoming message passes through a small, power-efficient filter that screens for suspicious signals.
- Stage 2 — Deep analysis: if the message looks risky, the fine-tuned Gemma model examines a larger slice of the conversation to confirm the threat or rule out a false alarm.
When it detects manipulation patterns — say, an urgent plea to transfer money to a "bank officer" — the system displays a real-time warning urging the user to pause before sharing information or sending money. Because the model lives on the device, it works across messaging apps like WhatsApp and Telegram in real time, and it analyzes conversations only with the user's consent, under the app's terms of use.
Why Singapore? The Numbers Are Staggering
The stakes explain the urgency. Scams account for more than half of Singapore's reported crimes, with losses exceeding S$4 billion since 2019, according to reports from the October 14 announcements. Social-engineering cons — scammers using convincing conversation to manipulate victims — now account for up to 80 percent of scam losses, Google DeepMind said. When fraud is that conversational, the defense has to be conversational too.
"Fighting scams shouldn't mean sacrificing personal privacy," said Mark Chen, GovTech's director of Trust and Safety. "Working with Google DeepMind, we're bringing advanced AI protection directly onto citizens' devices." Singapore's Minister for Digital Development and Information, Josephine Teo, framed the collaboration as part of a broader strategy: scams now move across social media, messaging apps, phone calls, and banks, so defenses must be connected across the whole ecosystem.

Part of a National AI Partnership
The scam detector is one piece of Google DeepMind's National AI Partnership with the Singapore government. The model will be rolled out into apps in phases after further consultation with the government, and Google says it will keep retraining the model as scammers change tactics — an arms race where the detector must learn as fast as the fraudsters. A release date has not been announced, and it is still unclear whether the feature will become a default part of Android.
It is worth keeping expectations honest. Independent checks of how often the detector catches scams — or how often it cries wolf on harmless chats — have not been published yet, and real-world accuracy will only be proven after the phased rollout. Still, the design choice is notable: an AI that reads your most private messages, whose core promise is that only your phone reads them.
The Bottom Line
DeepMind and GovTech are betting that the best anti-scam tool is one that lives where the scams live — inside your chats — without sending your words to the cloud. If the phased rollout delivers accurate, low-noise warnings, it could become a template for other countries battling the same social-engineering epidemic. Watch for the first apps carrying it in Singapore in the months ahead.

Disclaimer: Rollout timing, availability, and detection capabilities are based on October 14 announcements and are subject to change as the project develops.