Google DeepMind trained an AI on Singapore’s scams to catch fraudsters inside your chats

Scammers in Singapore no longer just send dodgy links. They chat, build trust and only then ask for money, and Google DeepMind thinks a small AI living on your phone can spot the moment a friendly conversation turns. The lab and the Government Technology Agency of Singapore (GovTech) announced on October 14 that they have fine-tuned Google’s open Gemma model on local scam conversations to build a scam detector that runs entirely on the device.

The model was presented at Singapore International Cyber Week (SICW 2026). It has not shipped yet: Google DeepMind says it will roll out inside apps in carefully phased stages, developed with the Singapore government, and it hasn’t named which apps will carry it or when.

Why Singapore wants it

The numbers explain the urgency. According to Google DeepMind’s post, scams such as investment fraud, fake job offers and impersonation of public officials make up over half of Singapore’s reported crimes and have cost victims more than S$4 billion (about US$3 billion) since 2019.

The tactics have also changed. Simple phishing has given way to social engineering, where the scammer manipulates the victim through a long, convincing conversation before asking for money. Those scams now account for up to 80% of all losses, the post says, and they are hard for keyword filters to catch because no single message looks suspicious.

How the on-device scam detector works

Google DeepMind says generic models struggle with local tricks, so the team fine-tuned Gemma on conversations covering impersonations of Singapore’s public institutions, banks and local figures, plus trending regional schemes. The aim is to tell normal social chatter from genuine manipulation.

Running a model like that on a phone all day would drain the battery, so the work is split into two stages:

• Stage 1: a lightweight, low-power classifier screens incoming messages for possible threats.
• Stage 2: only when something looks off, the fine-tuned Gemma model reads a larger snapshot of the conversation to confirm the threat or dismiss a false alarm.

If it decides a chat is manipulative, the phone shows a real-time warning telling the user to pause before sharing sensitive information or sending money. It is a similar approach to the one Google took with its offline Foresight meeting-notes app: a small Gemma model doing the work locally instead of in the cloud.

‘Fighting scams shouldn’t mean sacrificing personal privacy’

The obvious worry with any tool that reads private messages is who else gets to see them. Google DeepMind says the model only accesses conversations with the user’s consent, under the terms of whichever app it is built into, and because it runs entirely on the device, neither Google nor the Singapore government can see users’ conversations through it. Those are the companies’ own claims and haven’t been independently tested.

Mark Chen, GovTech’s Director of Trust and Safety, put it this way:

Fighting scams shouldn’t mean sacrificing personal privacy. Working with Google DeepMind, we’re bringing advanced AI protection directly onto citizens’ devices. This allows us to catch highly targeted, local scam attempts in real time while keeping personal communications private.

Mark Chen, Director, Trust and Safety, GovTech Singapore

Google DeepMind also set out how the model was trained. Fine-tuning took place in a GovTech-managed Google Cloud environment that kept the training data isolated; training used anonymised data drawn from scam reports and intelligence feeds, with no private or personally identifiable citizen data; and the model was stress-tested against synthetic dialogues and local text collections covering different scam types. The post gives no accuracy or false-alarm figures.

Part of a bigger Singapore AI push

The project comes under Google’s National AI Partnership with the Singapore government, announced in May. It lands on the same day Anthropic opened a Singapore office and signed an AI safety pact with the country’s Ministry of Digital Development and Information, as Singapore lines up deals with the big AI labs. Google DeepMind says it hopes to share more on the collaboration soon and is exploring whether similar technology could be deployed in other countries.

Why it matters

Conversational scams are a global problem, and the usual fix, scanning messages on a server, clashes with end-to-end encryption. A small model that flags manipulation on the phone itself, without the messages leaving it, could offer messaging apps a way round that, if the accuracy and the privacy promises hold up once it reaches real users.

Sources: Google DeepMind.