Google DeepMind and the Singapore Government unveil an AI that detects scams on the phone itself
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
In Singapore, GovTech and Google DeepMind have built a scam detector that works inside the phone: a slimmed-down Gemma trained on local cons. When a chat smells of manipulation, an alert pops up to stop the victim before they hand over data or money. The analysis happens on the handset, even with encrypted chats, and DeepMind has not yet given a date or named any apps.
Google DeepMind and GovTech, the Singapore Government's technology agency, have developed a model that spots scams on the phone itself and is tuned for the smartphones commonly used in Singapore. They presented it at Singapore International Cyber Week 2026 (SICW). It comes out of the national AI partnership that DeepMind announced with the Singapore Government in May.
In Singapore, scams account for more than half of reported crimes and have caused losses of over S$4 billion since 2019. According to DeepMind, they have moved on from basic phishing to elaborate social engineering, where the attacker wins the victim over with persuasive conversations before asking for money. Scams of this kind now represent up to 80% of the losses.
Generic models often miss tricks with a local flavor. So DeepMind fine-tuned a lightweight version of Gemma, Google's family of open models, on Singapore scam conversations, from people posing as government bodies, banks and well-known local personalities to scams currently popular in the region, so that it can tell everyday chat apart from real manipulation.
When the model detects a manipulative conversation, it shows a warning on the phone urging the person to pause and take care before sharing sensitive data or money. To save battery, it works in two steps: a light classifier checks the messages coming in and, if something looks suspicious, the Gemma model looks at a wider stretch of the conversation to decide whether the threat is real or a false alarm.
The analysis runs on the phone rather than on cloud servers, which is how threat detection usually works, because many messaging platforms encrypt chats end to end and cannot look inside them. "Fighting scams shouldn't mean sacrificing personal privacy," said Mark Chen, director of trust and safety at GovTech. According to DeepMind, the fine-tuning was done in a Google Cloud environment run by GovTech, which kept the training material isolated. It used anonymized data from scam reports and intelligence sources, with no personal information on citizens, and the model was tested on synthetic dialogues and local texts.
DeepMind has not said when it will reach users or in which apps. According to the company, messaging platforms and operating system makers can build these lightweight models straight into their products to protect users on the device itself.
Why it matters · analysis and opinion
The approach resolves a clash that seemed to have no way out: end-to-end encryption stops the platform from watching chats, and that is exactly where the costliest cons arrive. If the analysis lives on the phone, nobody outside reads anything and the alarm can still go off in time. Having a government drive it alongside a tech company also sets a path for other countries facing the same problem. The decisive part is still missing: messaging apps or operating systems actually building it in, and knowing how many false alarms it raises in daily use. Meanwhile, the defense is the usual one: distrust anyone who pushes for money or passwords in a hurry, even if they claim to be the bank or a public body.
Official source: Google DeepMind · Written with the help of AI: how we make the news
