Google and GovTech Put Scam Detection On Your Phone

A phone that can sniff out a scam conversation on its own, with no chat logs shipped to a remote server, is what Google DeepMind and Singapore's GovTech have built together. The pair unveiled the system at Singapore International Cyber Week 2026, describing it as one strand of Google's National AI Partnership with the government here.

Locally, fraud is the crime story that dwarfs the rest: reported cases tied to scams exceed half of the total, and cumulative losses since 2019 have climbed past S$4 billion. The money increasingly leaves victims' accounts through social engineering, long persuasive exchanges that wear people down until they transfer funds. Google DeepMind estimates that such manipulation drives as much as 80% of the losses.

Under the hood sits Gemma, Google's lightweight model line, refined on genuine local scam chatter: crooks posing as public agencies, banks and well-known figures, along with newer cons spreading across the region. Screening runs in two stages. A small classifier watches arriving messages first; only when it raises a flag does the heavier Gemma model read a wider slice of the exchange to confirm a threat or clear a false alarm. If manipulative patterns surface, the handset shows a live warning, telling users to think twice before revealing details or moving money.

Local processing is the point. Cloud-based scanning would mean conversations leaving the device, and Mark Chen, who leads Trust and Safety at GovTech Singapore, argues that fighting fraud should not cost anyone their privacy. Rollout will be gradual, through applications built with the Singapore Government; the model reads conversations only where users consent under those apps' terms, and tuning takes place inside a GovTech-managed Google Cloud environment that keeps training data walled off. The datasets were anonymised, drawn from scam reports and intelligence feeds, with personal identifiers removed. Researchers also pitted the system against synthetic exchanges and local collections covering a range of scam categories and evolving impersonation tactics.

Why it matters for Singapore: Scam defence that stays on the handset sidesteps the usual trade-off between privacy and protection, and it sets a template other markets could copy, since DeepMind says it is weighing similar deployments abroad. What is still missing is a timeline: no launch date or list of participating apps has been announced.